Proceedings of International Joint Conference on Computational Intelligence: IJCCI 2018 [1st ed.] 978-981-13-7563-7;978-981-13-7564-4

Table of contents :
Front Matter ....Pages i-xiii
Factorial Analysis of Biological Datasets (H. M. Shahriar Parvez, Saqib Hakak, Gulshan Amin Gilkar, Mahmud Abdur Rahman)....Pages 1-9
Classification of Motor Imagery Events from Prefrontal Hemodynamics for BCI Application (Md. Asadur Rahman, Md. Mahmudul Haque, Anika Anjum, Md. Nurunnabi Mollah, Mohiuddin Ahmad)....Pages 11-23
Diabetic Retinopathy Detection Using PCA-SIFT and Weighted Decision Tree (Fatema T. Johora, Md. Mahbub -Or-Rashid, Mohammad A. Yousuf, Tumpa Rani Saha, Bulbul Ahmed)....Pages 25-37
GIS-Based Surface Water Changing Analysis in Rajshahi City Corporation Area Using Ensemble Classifier (Mahbina Akter Mim, K. M. Shawkat Zamil)....Pages 39-47
Leveraging Machine Learning Approach to Setup Software-Defined Network(SDN) Controller Rules During DDoS Attack (Sajib Sen, Kishor Datta Gupta, Md. Manjurul Ahsan)....Pages 49-60
A Fuzzy-Based Study for Biomedical Imaging Applications (Fahmida Ahmed, Tausif Uddin Ahmed Chowdhury, Md. Hasan Furhad)....Pages 61-71
Meta Classifier-Based Ensemble Learning For Sentiment Classification (Naznin Sultana, Mohammad Mohaiminul Islam)....Pages 73-84
Mining Periodic Patterns and Accuracy Calculation for Activity Monitoring Using RF Tag Arrays (Md. Amirul Islam, Uzzal Kumar Acharjee)....Pages 85-95
Can the Expansion of Prediction Errors be Counterbalanced in Reversible Data Hiding? (Hussain Nyeem, Sultan Abdul Hasib)....Pages 97-109
Drowsiness Detection Using Eye-Blink Pattern and Mean Eye Landmarks’ Distance (Abdullah Arafat Miah, Mohiuddin Ahmad, Khatuna Zannat Mim)....Pages 111-121
Routing Protocol Selection for Intelligent Transport System (ITS) of VANET in High Mobility Areas of Bangladesh (Md. Kamrul Hasan, Orvila Sarker)....Pages 123-135
An Intelligent Children Healthcare System by Using Ensemble Technique (Nishargo Nigar, Linkon Chowdhury)....Pages 137-150
Microprocessor-Based Smart Blind Glass System for Visually Impaired People (Md. Tobibul Islam, Mohiuddin Ahmad, Akash Shingha Bappy)....Pages 151-161
A Study on Monitoring Coastal Areas for Having a Better Underwater Surveillance Perspective (Md. Hasan Furhad, Mohiuddin Ahmed, Abu S. S. M. Barkat Ullah)....Pages 163-174
Ethanol Detection Through Photonic Crystal Fiber (Etu Podder, Md. Bellal Hossain, Abdullah Al-Mamun Bulbul, Himadri Shekhar Mondal)....Pages 175-182
Computer-Aided Speckle Noise Analysis in Ultrasound Images Through Fusion of Convolutional Neural Network and Wavelet Transform with Linear Discriminate Analysis (Rafid Mostafiz, Md. Mezbahul Islam, Md. Motiur Rahman)....Pages 183-195
A Dynamic Bandwidth Allocation Algorithm for Gigabit Passive Optical Network for Reducing Packet Delay and Bit Error Rate (Md. Hayder Ali, Mohammad Hanif Ali)....Pages 197-206
Feature Selection and Biomedical Signal Classification Using Minimum Redundancy Maximum Relevance and Artificial Neural Network (Md. Masud Rana, Kawsar Ahmed)....Pages 207-214
An Identity-Based Encryption Scheme for Data Security in Fog Computing (Nishat Farjana, Shanto Roy, Md. Julkar Nayeen Mahi, Md Whaiduzzaman)....Pages 215-226
Modeling Photon Propagation Through Human Breast with Tumor in Diffuse Optical Tomography (Shisir Mia, Md. Mijanur Rahman, Mohammad Motiur Rahman)....Pages 227-233
A Network-Based Approach to Identify Molecular Signatures and Comorbidities of Thyroid Cancer (Md. Ali Hossain, Tania Akter Asa, Fazlul Huq, Julian M. W. Quinn, Mohammad Ali Moni)....Pages 235-246
Alcoholic Brain State Identification from Brain Signals Using Support Vector Machine-Based Algorithm (Siuly Siuly, Enamul Kabir, Hua Wang, Frank Whittaker, Hongbo Kuang)....Pages 247-253
A Machine Learning Approach to Detect Diabetic Retinopathy Using Convolutional Neural Network (Muhammad Mahir Hasan Chowdhury, Nishat Tasnim Ahmed Meem, Marium-E-Jannat)....Pages 255-264
A Comparative Overview of Classification Algorithm for Bangla Handwritten Digit Recognition (Md. Nazmul Hoq, Mohammad Mohaiminul Islam, Nadira Anjum Nipa, Md. Mostofa Akbar)....Pages 265-277
Fraud Detection of Facebook Business Page Based on Sentiment Analysis (Samia Nasrin, Priyanka Ghosh, S. M. Mazharul Hoque Chowdhury, Sheikh Abujar, Syed Akhter Hossain)....Pages 279-287
A Framework for Detecting Driver Drowsiness Based on Eye Blinking Rate and Hand Gripping Pressure (Md. Ashfakur Rahman Arju, Naib Hossain Khan, Kazi Ekramul Hoque, Arif Rizvi Jisan, Saifuddin M. Tareque, Md. Zahid Hasan)....Pages 289-304
A Day-Ahead Power Demand Prediction for Distribution-Side Peak Load Management (Khizir Mahmud, Weilun Peng, Sayidul Morsalin, Jayashri Ravishankar)....Pages 305-315
Simulation and Comparison of RPL, 6LoWPAN, and CoAP Protocols Using Cooja Simulator (Arif Mahmud, Faria Hossain, Tasnim Ara Choity, Faija Juhin)....Pages 317-326
Algorithms for String Comparison in DNA Sequences (Dhiman Goswami, Nishat Sultana, Warda Ruheen Bristi)....Pages 327-343
A New Approach for Efficient Face Detection Using BPV Algorithm Based on Mathematical Modeling (Tangina Sultana, M. Delowar Hossain, Niamul Hasan Zead, Nur Alam Sarker, Jannatul Fardoush)....Pages 345-358
A Computational Approach to Author Identification from Bengali Song Lyrics (Nazmun Nisat Ontika, Md. Fasihul Kabir, Ashraful Islam, Eshtiak Ahmed, Mohammad Nurul Huda)....Pages 359-369
Automatic Skin Lesion Segmentation and Melanoma Detection: Transfer Learning Approach with U-Net and DCNN-SVM (Zabir Al Nazi, Tasnim Azad Abir)....Pages 371-381
A Non-invasive Heart Rate Estimation Approach from Photoplethysmography (Monira Islam, Trisa Biswas, Abdul Munem Saad, Chowdhury Azimul Haque, Md. Salah Uddin Yusuf)....Pages 383-394
Issues of Internet of Things (IoT) and an Intrusion Detection System for IoT Using Machine Learning Paradigm (M. F. Mridha, Md. Abdul Hamid, Md. Asaduzzaman)....Pages 395-406
A Collaborative Platform to Collect Data for Developing Machine Translation Systems (Md. Arid Hasan, Firoj Alam, Sheak Rashed Haider Noori)....Pages 407-416
A Comparative Study of Classifiers in the Context of Papaya Disease Recognition (Md. Tarek Habib, Anup Majumder, Rabindra Nath Nandi, Farruk Ahmed, Mohammad Shorif Uddin)....Pages 417-429
A Hierarchical Learning Model for Claim Validation (Amar Debnath, Redoan Rahman, Md. Mofijul Islam, Md. Abdur Razzaque)....Pages 431-441
D-CARE: A Non-invasive Glucose Measuring Technique for Monitoring Diabetes Patients (Md. Mahbub Alam, Swapnil Saha, Proshib Saha, Fernaz Narin Nur, Nazmun Nessa Moon, Asif Karim et al.)....Pages 443-453
Enhancing the Classification Performance of Lower Back Pain Symptoms Using Genetic Algorithm-Based Feature Selection (Abdullah Al Imran, Md. Rifatul Islam Rifat, Rafeed Mohammad)....Pages 455-469
A CNN-Based Classification Model for Recognizing Visual Bengali Font (Md. Zahid Hasan, Kh. Tanzila Rahman, Rokeya Islam Riya, K. M. Zubair Hasan, Nusrat Zahan)....Pages 471-482
Performance Analysis of SDN-Based Intrusion Detection Model with Feature Selection Approach (Samrat Kumar Dey, Md. Raihan Uddin, Md. Mahbubur Rahman)....Pages 483-494
Query-Oriented Active Community Search (Badhan Chandra Das, Md. Shoaib Ahmed, Md Musfique Anwar)....Pages 495-505
Olympic Sports Events Classification Using Convolutional Neural Networks (Shahana Shultana, Md. Shakil Moharram, Nafis Neehal)....Pages 507-518
Type 2 Diabetics Treatment and Medication Detection with Machine Learning Classifier Algorithm (Md. Kowsher, Farhana Sharmin Tithi, Tapasy Rabeya, Fahmida Afrin, Mohammad Nurul Huda)....Pages 519-531
Initial Point Prediction Based Parametric Active Contour Model for Left Ventricle Segmentation of CMRI Images (Md. Al Noman, A. B. M. Aowlad Hossain, Md. Asadur Rahman)....Pages 533-546
Bangla Handwritten Digit Recognition and Generation (Md. Fahim Sikder)....Pages 547-556
Portable Mini-Weather Station for Agricultural Sector of Rural Area in Bangladesh (Nazib Ahmad, Thajid Ibna Rouf Uday, Md. Toriqul Islam, Rayhan Patoary, Md. Mostasim Billah, Nuhash Ahmed et al.)....Pages 557-569
Appliance of Agile Methodology at Software Industry in Developing Countries: Perspective in Bangladesh (Abdus Sattar, Arif Mahmud, Sheak Rashed Haider Noori)....Pages 571-581
A Novel Approach for Tomato Diseases Classification Based on Deep Convolutional Neural Networks (Md. Ferdouse Ahmed Foysal, Mohammad Shakirul Islam, Sheikh Abujar, Syed Akhter Hossain)....Pages 583-591
Classification by Clustering (CbC): An Approach of Classifying Big Data Based on Similarities (Sakib Shahriar Khan, Shakim Ahamed, Miftahul Jannat, Swakkhar Shatabda, Dewan Md. Farid)....Pages 593-605
Brain–Machine Interface for Developing Virtual-Ball Movement Controlling Game (Md. Ochiuddin Miah, Al Maruf Hassan, Khondaker Abdullah Al Mamun, Dewan Md. Farid)....Pages 607-616
Vehicle Tracking and Monitoring System for Security Purpose Based on Thermoelectric Generator (TEG) (Md. Fahim Newaz, Abu Tayab Noman, Humayun Rashid, Nawsher Ahmed, Mohammad Emdadul Islam, S. M. Taslim Reza)....Pages 617-629
Improved Subspace Detection Based on Minimum Noise Fraction and Mutual Information for Hyperspectral Image Classification (Md. Rashedul Islam, Md. Ali Hossain, Boshir Ahmed)....Pages 631-641
A RSA-Based Efficient Dynamic Secure Algorithm for Ensuring Data Security (Himadri Shekhar Mondal, Md. Tariq Hasan, Md. Mahbub Hossain, Md. Mashrur Arifin, Rekha Saha)....Pages 643-653
Improved Time Complexity and Load Balance for DFS in Multiple NameNode (Mohammad Nurul Islam, Md. Nasim Akhtar)....Pages 655-664
Real-Time Crowd Detection to Prevent Stampede (Sabrina Haque, Muhammad Sheikh Sadi, Md. Erfanul Haque Rafi, Md. Milon Islam, Md. Kamrul Hasan)....Pages 665-678
Development of an Expert System-Oriented Service Support Help Desk Management System (Abrar Hasin Kamal, Mohammad Obaidullah Tusher, Shadman Fahim Ahmad, Nusrat Jahan Farin, Nafees Mansoor)....Pages 679-692
Range-Based Location Estimation of Machines in M2M Communications Over Cellular Networks (Sree Krishna Das, Ratna Mudi)....Pages 693-703
Developing a Technique for Overcoming the Searching Limitations of Documents (Md. Muntasir Shahriar, Mohammad Shamsul Arefin, M. Ali Akber Dewan)....Pages 705-719
An Intelligent Technique for Stock Market Prediction (Mohammad Mekayel Anik, Mohammad Shamsul Arefin, M. Ali Akber Dewan)....Pages 721-733
An Approach to Aggregate Intuitionistic Fuzzy Information with the Help of Linear Operator (Meenakshi Kaushal, Mohd Shoaib Khan, Q. M. Danish Lohani)....Pages 735-746
Back Matter ....Pages 747-749

Citation preview

Algorithms for Intelligent Systems Series Editors: Jagdish Chand Bansal · Kusum Deep · Atulya K. Nagar

Mohammad Shorif Uddin Jagdish Chand Bansal Editors

Proceedings of International Joint Conference on Computational Intelligence IJCCI 2018

Algorithms for Intelligent Systems Series Editors Jagdish Chand Bansal, Department of Mathematics, South Asian University, New Delhi, Delhi, India Kusum Deep, Department of Mathematics, Indian Institute of Technology Roorkee, Roorkee, Uttarakhand, India Atulya K. Nagar, Department of Mathematics and Computer Science, Liverpool Hope University, Liverpool, UK

This book series publishes research on the analysis and development of algorithms for intelligent systems with their applications to various real world problems. It covers research related to autonomous agents, multi-agent systems, behavioral modeling, reinforcement learning, game theory, mechanism design, machine learning, meta-heuristic search, optimization, planning and scheduling, artificial neural networks, evolutionary computation, swarm intelligence and other algorithms for intelligent systems. The book series includes recent advancements, modification and applications of the artificial neural networks, evolutionary computation, swarm intelligence, artificial immune systems, fuzzy system, autonomous and multi agent systems, machine learning and other intelligent systems related areas. The material will be beneficial for the graduate students, post-graduate students as well as the researchers who want a broader view of advances in algorithms for intelligent systems. The contents will also be useful to the researchers from other fields who have no knowledge of the power of intelligent systems, e.g. the researchers in the field of bioinformatics, biochemists, mechanical and chemical engineers, economists, musicians and medical practitioners. The series publishes monographs, edited volumes, advanced textbooks and selected proceedings.

More information about this series at http://www.springer.com/series/16171

Mohammad Shorif Uddin Jagdish Chand Bansal

•

Editors

Proceedings of International Joint Conference on Computational Intelligence IJCCI 2018

123

Editors Mohammad Shorif Uddin Department of Computer Science and Engineering Jahangirnagar University Dhaka, Bangladesh

Jagdish Chand Bansal Department of Mathematics South Asian University New Delhi, Delhi, India

ISSN 2524-7565 ISSN 2524-7573 (electronic) Algorithms for Intelligent Systems ISBN 978-981-13-7563-7 ISBN 978-981-13-7564-4 (eBook) https://doi.org/10.1007/978-981-13-7564-4 © Springer Nature Singapore Pte Ltd. 2020 This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd. The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721, Singapore

Preface

This book contains high-quality research papers as the proceedings of the International Joint Conference on Computational Intelligence (IJCCI 2018). IJCCI 2018 has been jointly organized by Daffodil International University (DIU), Bangladesh; Jahangirnagar University (JU), Bangladesh; and South Asian University (SAU), India. It was held on 14–15 December 2018 at DIU, Dhaka, Bangladesh. The conference was conceived as a platform for disseminating and exchanging ideas, concepts and results of the researchers from academia and industry to develop a comprehensive understanding of the challenges of the advancements of intelligence in computational viewpoints. This book will help in strengthening a congenial and nice networking between academia and industry. The conference focused on collective intelligence, soft computing, optimization, cloud computing, machine learning, intelligent software, robotics, data science, data security, big data analytics, signal and natural language processing. This conference is a biennial update of the first conference named International Workshop on Computational Intelligence (IWCI 2016) that was held on 12–13 December 2016 at Jahangirnagar University, Dhaka, Bangladesh, in collaboration with South Asian University (SAU), India, under the technical co-sponsorship of IEEE Bangladesh Section. All accepted and presented papers of IWCI 2016 are in IEEE Xplore Digital Library. We have tried our best to enrich the quality of IJCCI 2018 through a stringent and careful peer review process. IJCCI 2018 received 182 papers from 496 authors (459 are local authors and 37 are foreign authors from 9 countries), and 63 papers were finally accepted for presentation. However, the proceedings contain 61 papers. In fact, this book presents the novel contributions in areas of computational intelligence and it serves as a reference material for advanced research. Dhaka, Bangladesh New Delhi, India

Mohammad Shorif Uddin Jagdish Chand Bansal

v

Contents

1

Factorial Analysis of Biological Datasets . . . . . . . . . . . . . . . . . . . . . H. M. Shahriar Parvez, Saqib Hakak, Gulshan Amin Gilkar and Mahmud Abdur Rahman

2

Classification of Motor Imagery Events from Prefrontal Hemodynamics for BCI Application . . . . . . . . . . . . . . . . . . . . . . . . Md. Asadur Rahman, Md. Mahmudul Haque, Anika Anjum, Md. Nurunnabi Mollah and Mohiuddin Ahmad

3

4

5

Diabetic Retinopathy Detection Using PCA-SIFT and Weighted Decision Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Fatema T. Johora, Md. Mahbub -Or-Rashid, Mohammad A. Yousuf, Tumpa Rani Saha and Bulbul Ahmed

1

11

25

GIS-Based Surface Water Changing Analysis in Rajshahi City Corporation Area Using Ensemble Classifier . . . . . . . . . . . . . . . . . Mahbina Akter Mim and K. M. Shawkat Zamil

39

Leveraging Machine Learning Approach to Setup SoftwareDefined Network(SDN) Controller Rules During DDoS Attack . . . . Sajib Sen, Kishor Datta Gupta and Md. Manjurul Ahsan

49 61

6

A Fuzzy-Based Study for Biomedical Imaging Applications . . . . . . Fahmida Ahmed, Tausif Uddin Ahmed Chowdhury and Md. Hasan Furhad

7

Meta Classifier-Based Ensemble Learning For Sentiment Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Naznin Sultana and Mohammad Mohaiminul Islam

73

Mining Periodic Patterns and Accuracy Calculation for Activity Monitoring Using RF Tag Arrays . . . . . . . . . . . . . . . . . . . . . . . . . . Md. Amirul Islam and Uzzal Kumar Acharjee

85

8

vii

viii

9

Contents

Can the Expansion of Prediction Errors be Counterbalanced in Reversible Data Hiding? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Hussain Nyeem and Sultan Abdul Hasib

97

10 Drowsiness Detection Using Eye-Blink Pattern and Mean Eye Landmarks’ Distance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 Abdullah Arafat Miah, Mohiuddin Ahmad and Khatuna Zannat Mim 11 Routing Protocol Selection for Intelligent Transport System (ITS) of VANET in High Mobility Areas of Bangladesh . . . . . . . . . . . . . 123 Md. Kamrul Hasan and Orvila Sarker 12 An Intelligent Children Healthcare System by Using Ensemble Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 Nishargo Nigar and Linkon Chowdhury 13 Microprocessor-Based Smart Blind Glass System for Visually Impaired People . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151 Md. Tobibul Islam, Mohiuddin Ahmad and Akash Shingha Bappy 14 A Study on Monitoring Coastal Areas for Having a Better Underwater Surveillance Perspective . . . . . . . . . . . . . . . . . . . . . . . 163 Md. Hasan Furhad, Mohiuddin Ahmed and Abu S. S. M. Barkat Ullah 15 Ethanol Detection Through Photonic Crystal Fiber . . . . . . . . . . . . 175 Etu Podder, Md. Bellal Hossain, Abdullah Al-Mamun Bulbul and Himadri Shekhar Mondal 16 Computer-Aided Speckle Noise Analysis in Ultrasound Images Through Fusion of Convolutional Neural Network and Wavelet Transform with Linear Discriminate Analysis . . . . . . . . . . . . . . . . 183 Rafid Mostafiz, Md. Mezbahul Islam and Md. Motiur Rahman 17 A Dynamic Bandwidth Allocation Algorithm for Gigabit Passive Optical Network for Reducing Packet Delay and Bit Error Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197 Md. Hayder Ali and Mohammad Hanif Ali 18 Feature Selection and Biomedical Signal Classification Using Minimum Redundancy Maximum Relevance and Artificial Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207 Md. Masud Rana and Kawsar Ahmed 19 An Identity-Based Encryption Scheme for Data Security in Fog Computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215 Nishat Farjana, Shanto Roy, Md. Julkar Nayeen Mahi and Md Whaiduzzaman

Contents

ix

20 Modeling Photon Propagation Through Human Breast with Tumor in Diffuse Optical Tomography . . . . . . . . . . . . . . . . . . 227 Shisir Mia, Md. Mijanur Rahman and Mohammad Motiur Rahman 21 A Network-Based Approach to Identify Molecular Signatures and Comorbidities of Thyroid Cancer . . . . . . . . . . . . . . . . . . . . . . 235 Md. Ali Hossain, Tania Akter Asa, Fazlul Huq, Julian M. W. Quinn and Mohammad Ali Moni 22 Alcoholic Brain State Identification from Brain Signals Using Support Vector Machine-Based Algorithm . . . . . . . . . . . . . . 247 Siuly Siuly, Enamul Kabir, Hua Wang, Frank Whittaker and Hongbo Kuang 23 A Machine Learning Approach to Detect Diabetic Retinopathy Using Convolutional Neural Network . . . . . . . . . . . . . . . . . . . . . . . 255 Muhammad Mahir Hasan Chowdhury, Nishat Tasnim Ahmed Meem and Marium-E-Jannat 24 A Comparative Overview of Classification Algorithm for Bangla Handwritten Digit Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265 Md. Nazmul Hoq, Mohammad Mohaiminul Islam, Nadira Anjum Nipa and Md. Mostofa Akbar 25 Fraud Detection of Facebook Business Page Based on Sentiment Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279 Samia Nasrin, Priyanka Ghosh, S. M. Mazharul Hoque Chowdhury, Sheikh Abujar and Syed Akhter Hossain 26 A Framework for Detecting Driver Drowsiness Based on Eye Blinking Rate and Hand Gripping Pressure . . . . . . . . . . . . . . . . . . 289 Md. Ashfakur Rahman Arju, Naib Hossain Khan, Kazi Ekramul Hoque, Arif Rizvi Jisan, Saifuddin M. Tareque and Md. Zahid Hasan 27 A Day-Ahead Power Demand Prediction for Distribution-Side Peak Load Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305 Khizir Mahmud, Weilun Peng, Sayidul Morsalin and Jayashri Ravishankar 28 Simulation and Comparison of RPL, 6LoWPAN, and CoAP Protocols Using Cooja Simulator . . . . . . . . . . . . . . . . . . . . . . . . . . 317 Arif Mahmud, Faria Hossain, Tasnim Ara Choity and Faija Juhin 29 Algorithms for String Comparison in DNA Sequences . . . . . . . . . . 327 Dhiman Goswami, Nishat Sultana and Warda Ruheen Bristi

x

Contents

30 A New Approach for Efficient Face Detection Using BPV Algorithm Based on Mathematical Modeling . . . . . . . . . . . . . . . . . 345 Tangina Sultana, M. Delowar Hossain, Niamul Hasan Zead, Nur Alam Sarker and Jannatul Fardoush 31 A Computational Approach to Author Identification from Bengali Song Lyrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359 Nazmun Nisat Ontika, Md. Fasihul Kabir, Ashraful Islam, Eshtiak Ahmed and Mohammad Nurul Huda 32 Automatic Skin Lesion Segmentation and Melanoma Detection: Transfer Learning Approach with U-Net and DCNN-SVM . . . . . . 371 Zabir Al Nazi and Tasnim Azad Abir 33 A Non-invasive Heart Rate Estimation Approach from Photoplethysmography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383 Monira Islam, Trisa Biswas, Abdul Munem Saad, Chowdhury Azimul Haque and Md. Salah Uddin Yusuf 34 Issues of Internet of Things (IoT) and an Intrusion Detection System for IoT Using Machine Learning Paradigm . . . . . . . . . . . . 395 M. F. Mridha, Md. Abdul Hamid and Md. Asaduzzaman 35 A Collaborative Platform to Collect Data for Developing Machine Translation Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407 Md. Arid Hasan, Firoj Alam and Sheak Rashed Haider Noori 36 A Comparative Study of Classifiers in the Context of Papaya Disease Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 417 Md. Tarek Habib, Anup Majumder, Rabindra Nath Nandi, Farruk Ahmed and Mohammad Shorif Uddin 37 A Hierarchical Learning Model for Claim Validation . . . . . . . . . . 431 Amar Debnath, Redoan Rahman, Md. Mofijul Islam and Md. Abdur Razzaque 38 D-CARE: A Non-invasive Glucose Measuring Technique for Monitoring Diabetes Patients . . . . . . . . . . . . . . . . . . . . . . . . . . 443 Md. Mahbub Alam, Swapnil Saha, Proshib Saha, Fernaz Narin Nur, Nazmun Nessa Moon, Asif Karim and Sami Azam 39 Enhancing the Classification Performance of Lower Back Pain Symptoms Using Genetic Algorithm-Based Feature Selection . . . . . 455 Abdullah Al Imran, Md. Rifatul Islam Rifat and Rafeed Mohammad 40 A CNN-Based Classification Model for Recognizing Visual Bengali Font . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 471 Md. Zahid Hasan, Kh. Tanzila Rahman, Rokeya Islam Riya, K. M. Zubair Hasan and Nusrat Zahan

Contents

xi

41 Performance Analysis of SDN-Based Intrusion Detection Model with Feature Selection Approach . . . . . . . . . . . . . . . . . . . . . . . . . . 483 Samrat Kumar Dey, Md. Raihan Uddin and Md. Mahbubur Rahman 42 Query-Oriented Active Community Search . . . . . . . . . . . . . . . . . . . 495 Badhan Chandra Das, Md. Shoaib Ahmed and Md Musfique Anwar 43 Olympic Sports Events Classification Using Convolutional Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 507 Shahana Shultana, Md. Shakil Moharram and Nafis Neehal 44 Type 2 Diabetics Treatment and Medication Detection with Machine Learning Classifier Algorithm . . . . . . . . . . . . . . . . . 519 Md. Kowsher, Farhana Sharmin Tithi, Tapasy Rabeya, Fahmida Afrin and Mohammad Nurul Huda 45 Initial Point Prediction Based Parametric Active Contour Model for Left Ventricle Segmentation of CMRI Images . . . . . . . . . . . . . . 533 Md. Al Noman, A. B. M. Aowlad Hossain and Md. Asadur Rahman 46 Bangla Handwritten Digit Recognition and Generation . . . . . . . . . 547 Md. Fahim Sikder 47 Portable Mini-Weather Station for Agricultural Sector of Rural Area in Bangladesh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 557 Nazib Ahmad, Thajid Ibna Rouf Uday, Md. Toriqul Islam, Rayhan Patoary, Md. Mostasim Billah, Nuhash Ahmed and Farhana Sharmin Tithi 48 Appliance of Agile Methodology at Software Industry in Developing Countries: Perspective in Bangladesh . . . . . . . . . . . . 571 Abdus Sattar, Arif Mahmud and Sheak Rashed Haider Noori 49 A Novel Approach for Tomato Diseases Classification Based on Deep Convolutional Neural Networks . . . . . . . . . . . . . . . 583 Md. Ferdouse Ahmed Foysal, Mohammad Shakirul Islam, Sheikh Abujar and Syed Akhter Hossain 50 Classification by Clustering (CbC): An Approach of Classifying Big Data Based on Similarities . . . . . . . . . . . . . . . . . . . . . . . . . . . . 593 Sakib Shahriar Khan, Shakim Ahamed, Miftahul Jannat, Swakkhar Shatabda and Dewan Md. Farid 51 Brain–Machine Interface for Developing Virtual-Ball Movement Controlling Game . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 607 Md. Ochiuddin Miah, Al Maruf Hassan, Khondaker Abdullah Al Mamun and Dewan Md. Farid

xii

Contents

52 Vehicle Tracking and Monitoring System for Security Purpose Based on Thermoelectric Generator (TEG) . . . . . . . . . . . . . . . . . . 617 Md. Fahim Newaz, Abu Tayab Noman, Humayun Rashid, Nawsher Ahmed, Mohammad Emdadul Islam and S. M. Taslim Reza 53 Improved Subspace Detection Based on Minimum Noise Fraction and Mutual Information for Hyperspectral Image Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 631 Md. Rashedul Islam, Md. Ali Hossain and Boshir Ahmed 54 A RSA-Based Efficient Dynamic Secure Algorithm for Ensuring Data Security . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 643 Himadri Shekhar Mondal, Md. Tariq Hasan, Md. Mahbub Hossain, Md. Mashrur Arifin and Rekha Saha 55 Improved Time Complexity and Load Balance for DFS in Multiple NameNode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 655 Mohammad Nurul Islam and Md. Nasim Akhtar 56 Real-Time Crowd Detection to Prevent Stampede . . . . . . . . . . . . . 665 Sabrina Haque, Muhammad Sheikh Sadi, Md. Erfanul Haque Rafi, Md. Milon Islam and Md. Kamrul Hasan 57 Development of an Expert System-Oriented Service Support Help Desk Management System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 679 Abrar Hasin Kamal, Mohammad Obaidullah Tusher, Shadman Fahim Ahmad, Nusrat Jahan Farin and Nafees Mansoor 58 Range-Based Location Estimation of Machines in M2M Communications Over Cellular Networks . . . . . . . . . . . . . . . . . . . . 693 Sree Krishna Das and Ratna Mudi 59 Developing a Technique for Overcoming the Searching Limitations of Documents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 705 Md. Muntasir Shahriar, Mohammad Shamsul Arefin and M. Ali Akber Dewan 60 An Intelligent Technique for Stock Market Prediction . . . . . . . . . . 721 Mohammad Mekayel Anik, Mohammad Shamsul Arefin and M. Ali Akber Dewan 61 An Approach to Aggregate Intuitionistic Fuzzy Information with the Help of Linear Operator . . . . . . . . . . . . . . . . . . . . . . . . . . 735 Meenakshi Kaushal, Mohd Shoaib Khan and Q. M. Danish Lohani Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 747

About the Editors

Dr. Mohammad Shorif Uddin is a Professor at Jahangirnagar University, Bangladesh. He completed his Master of Technology Education at Shiga University, Japan in 1999, his Doctor of Engineering in Information Science at Kyoto Institute of Technology, Japan, in 2002, and an MBA at Jahangirnagar University in 2013. He is the Editor-in-Chief of ULAB Journal of Science and Engineering, an Associate Editor of IEEE Access, and has served as General Chair or Co-Chair of various conferences, including the IJCCI 2018, EICT 2017 and IWCI 2016. He holds two patents for his scientific inventions, is a senior member of several academic associations, and has published extensively in international journals and conference proceedings. Dr. Jagdish Chand Bansal is an Assistant Professor at South Asian University, New Delhi, India and a Visiting Research Fellow at Liverpool Hope University, UK. A leading researcher in the field of swarm intelligence, he has published numerous research papers in respected national and international journals.

xiii

Chapter 1

Factorial Analysis of Biological Datasets H. M. Shahriar Parvez, Saqib Hakak, Gulshan Amin Gilkar and Mahmud Abdur Rahman

1 Introduction There are a lot of studies going in the area of bioinformatics applications like DNA analysis, protein analysis, DNA sequencing, and others. This has resulted in large biological data generation [1]. Some of the standard biological databases include the international cancer genome project—Encode project [2], 1,000 Genomes project [3], and other standard databases. For analyzing biological datasets, the first step involves retrieval and efforts are being put forward to retrieve a particular pattern from a given data with much higher efficiency and accuracy [4–6]. This problem of pattern matching is of utmost significance particularly for biological applications and text retrieval applications involving Arabic, French, and other similar big data related applications [7–10]. The problem of pattern matching is to find a particular pattern p from a given data denoted by d [11]. Pattern matching is employed in many applications like in search engines, text processing, DNA analysis, network intrusion detection, web retrieval, and other such applications [11]. To solve this problem of pattern matching, numerous algorithms were developed. A good review regarding the performance of H. M. S. Parvez · S. Hakak (B) University of Malaya, Kuala Lumpur 50603, Malaysia e-mail: [email protected] H. M. S. Parvez e-mail: [email protected] G. A. Gilkar Shaqra University, Riyadh 11961, Saudi Arabia e-mail: [email protected] M. Abdur Rahman University of Liberal Arts Bangladesh, Dhaka 1209, Bangladesh e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_1

1

2

H. M. S. Parvez et al.

some pattern-based algorithms is given by Simone et al. [5]. Numerous studies have focused on improving the efficiency of the algorithm by implementing different data structures, optimizing loops and other methods. Less attention has been paid regarding the factors which can directly affect the performance of an algorithm [12]. Although, theoretically, one can assume the factors like input size can directly affect the performance of an algorithm or pattern length can affect the performance of an algorithm. Practically, to the best of our knowledge, no study has focused on any mathematical technique which can show exactly using empirical values, how much a particular factor is having an effect on the performance of an algorithm while retrieving data from biological datasets. Thus, in this article, different performance analysis techniques were investigated and 2k factorial design was designed to suit requirements for the completion of this research. Retrieval of two biological datasets, i.e., Genome sequence and Protein Sequence was done using backward oracle matching (BOM) [13] and extended backward oracle matching (EBOM) [14] algorithms with varying pattern lengths. It is assumed that the findings of this paper will open new insight to future researchers in the area of biological computation where this model can be used to know the effect of different factors before designing an algorithm. The organization of this article is as follows: Sect. 2 presents a brief overview of the background of the research and factorial design techniques. In Sect. 3, the methodology of the research is mentioned. Results are shown in Sect. 4 followed by a conclusion in Sect. 5.

2 Background Due to a lot of applications related to biological data analytics, research has been carried out in different dimensions. Some studies have focused on the analysis of DNA, some have focused on sequencing alignments, and some have focused on detecting mutations and some on the retrieval and so on. In addition to past algorithms, many new algorithms were proposed to locate a particular pattern from a given dataset [11, 15]. There is such a large list of algorithms that it is difficult to mention each algorithm. One of the well-known algorithms is basic local alignment search tool (BLAST) [16]. This algorithm is used to analyze the sequence alignment and considered a standard benchmark. Similarly, other algorithms which are used for analysis related to biological datasets include Berry-Ravindran algorithm (BR) [17], two sliding window algorithm (TSW) [18], enhanced two sliding windows algorithm (ETSW) [19], and others. In a few related papers related to factorial design, the authors [12, 20–23] have used factorial design technique to determine the effect of packet size and direct-sequence spread spectrum (DSSS) rate on several network performance indicators like throughput, jitter average delay in mobile ad hoc networks. This work provided insight to implement and redesign the factorial model for biological applications.

1 Factorial Analysis of Biological Datasets

3

2.1 Method to Analyze Performance of an Algorithm An algorithm is said to be efficient if its computational utilization (or computational expense) is at or beneath some satisfactory level. There is no specific tool available that can measure the efficiency of an algorithm. The two most widely used measures for measuring the performance of an algorithm include time complexity and space complexity. These measures may rely on upon a number of factors like the number of loops within an algorithm, type of data structures, size of the input, encoding technique used, the arrangement of data, and other such factors. The two methods that are used to analyze the performance of algorithms are the following: (1) Theoretical Analysis. In this type of analysis, the typical practice is to evaluate the unpredictability of the algorithm in the asymptotic sense, i.e., utilize Big O notation. These notations are known as asymptotic notations and analyzes the behavior of an algorithm in terms of its running time as input for the algorithm increases. These notations are used to predict the time complexity of the algorithm. There are two ways by which time complexity of an algorithm can be known, i.e., to evaluate the algorithm using best case also known as Big Omega and worst case known as Big O notation. (2) Benchmark Analysis. Nowadays, benchmark work is commonly used to evaluate the efficiency of work. In order to verify whether a new algorithm is efficient or not, it is compared to the previous works. Based on the comparison, it is concluded whether the new algorithm is a better approach or not. These two approaches are commonly used for analyzing the performance of algorithms. However, these two approaches are not able to determine the importance of any particular factor. For example, the importance of pattern length for an algorithm, the importance of the type of database used and other related factors. Mathematical techniques can be used for the same. We find that the factorial design method can be used to achieve the objectives of this research.

2.2 Mathematical Performance Techniques There are many kinds of mathematical performance models available for evaluating the effect of a particular factor on a response variable. However, to identify a particular model to suit the needs of particular research is a major and key issue. In this research, the focus has been only on experimental design techniques due to the fact that it is easy to separate factors and study the influence of those factors individually as well as together along with some other factors. A brief description of some major experimental design performance techniques is discussed below: (1) Basic Terminology of Experimental Designs. In this section, some basic terms which are used in factorial design techniques [12, 24] are mentioned:

4

H. M. S. Parvez et al.

(a) Response Variable: By response variable, we mean the outcome of an experiment. In other words, the measured performance of the system is known as the response variable. For example, in this research, there is one respective response variable which are execution time. (b) Factors: The variable which has some effect on a response variable like the type of data, pattern length, algorithm, input size. For example, in this research, there are three factors, i.e., algorithm, type of datasets, and pattern length. (c) Levels: Levels are the values which a factor can have. For example, if one of the factors is a user. The user can be put in three levels like high school, graduate school, or postgraduate school and thus this factor user has three levels, respectively. Similarly, in this research also, two different levels have been taken for each factor like two algorithms have been selected; BOM and EBOM in which former algorithm is a little bit faster as compared to another one. Similarly, two pattern lengths have been taken. (d) Design: It consists of a complete set of experiments done for all possible factors at all possible levels. In other words, for example, if there are 5 factors with following levels, i.e., 3, 3, 4, 3, 3 then it will require 3 * 3 * 4 * 3 * 3 or 324 experiments and this experimental design can be repeated five times. This constitutes one possible experimental design. (e) Interaction: By interaction we mean if the effect of one factor suppose A depends on the level of another factor suppose B. The fundamental and primary purpose of factorial design is to get and provide the maximum amount of information through a minimum number of experiments. The most widely used experimental designs are simple designs, full factorial designs, and fractional designs. Below is the brief description of these three major experimental designs along with their advantages and disadvantages. (2) Factorial Designs. (a) Simple Designs: As the name itself implies simple, it means it is a very straightforward technique in which only one factor is varied and at a time effect of only one factor can be evaluated against a response variable. Although this design is simple statistically, it is not efficient due to the fact that if some factors have interaction with each other, using this model may lead us toward wrong conclusions [24]. (b) Full Factorial Designs: This model is an extension to the above model and utilizes all possible combinations at all levels of all factors [16]. It is given by the equation:

n=

k  i=1

ni

(1)

1 Factorial Analysis of Biological Datasets

5

where k denotes the number of factors with ith factor having ni levels. One of the major advantages of a full factorial design method is that all possible combinations of configuration are examined under the same workload. The effect of every factor can be found along with interactions among the factors. For this research, a full factorial design technique known as the 2k experimental design is used in which two levels are varied for a k single factor. It is one of the most popular and widely used techniques involving many studies like in chemical industry, management industry, material engineering, and so on [24]. (3) Fractional Factorial Designs. In most complex research studies which involve rigorous and too many experiments, factorial design using 2k technique becomes non-feasible due to the time limitations or costs in experiments involved. Other reasons may be too many factors with too many levels as well. In such cases, another method can be used known as full fractional factorial design. The advantage of the fractional factorial design method is that it saves time and cost as compared to a full factorial design method. However, there is one flaw and that is, this technique does not give much detailed information which a full factorial design technique gives. This technique cannot give all possible interactions among the factors [24]. These were the three major experimental designs and out of these 2k full factorial design model was selected as this model suits our requirements and can be designed according to the scenario needed in case of this research. From the above discussion, it can be concluded, a lot of research is going on in the area of analyzing biological datasets using pattern-based algorithms. As far as the effect of different factors on the performance of pattern-based algorithms is concerned, a lot of research needs to be done. Past studies have not considered knowing the effect of certain factors mathematically due to the lack of proper approach or methodology.

3 Methodology From the abovementioned factorial designs, the 2k design has been used in this research. The 2k experimental design model is used to find the effect of k factors with each k factor having two possible levels or values. This model is most widely used due to its easiness to analyze things and sorting out factors in a particular order. By 2k, it means there are k factors with two levels and total tasks are performed 2k times. After performing the experiment 2k times, we get 2k effects which include k main effects, two-factor interactions, three-factor interactions, and so on. In this model, the desired numbers of factors are taken. Suppose, in this research three factors were taken, each factor is assigned two levels denoted by 1 and −1. Once each factor is assigned two levels under 1 and −1, then the following formula is used to find the effect of factors:

6

H. M. S. Parvez et al.

Table 1 List of factors used in factorial design

Throughput

Levels −1

1

A

Pattern matching algorithm

BOM

EBOM

B

Type of database

Genome

Protein

C

Input size

2 MB

1024 MB

  SST = 23 q2A + q2B + qC2 + q2AB + q2AC + q2BC + q2ABC

(2)

Here A, B, C denotes different factors; AB, BC, and AC denote interaction between two factors and ABC denotes interaction among all three factors. And SST denotes Sum of Square Total [18]. Following steps have been taken to obtain the results: literature phase, data collection phase, design phase, and factorial design phase which are explained below: (1) Literature Phase: In this phase, investigation regarding performance techniques was done along with methodology followed by each approach. (2) Data Collection Phase: In this phase, data of protein sequence and genome sequence datasets were collected from previous work. The purpose was to save time in order to conduct experiments again. After reviewing lots of research articles, data was collected from the work of Simone [5]. All algorithms have been implemented using C programming language on Intel Core 2 Duo processor of 2.79 GHz with 2 GB of RAM and 2 MB of L2-cache. For compiling process, GNU C compiler has been used. The author has done appreciable work in the field of pattern-based algorithms. In the case of algorithm factor, backward oracle matching (BOM) and extended backward oracle matching (EBOM) algorithms are taken and put under two levels, i.e., 1 and −1. Pattern length was another factor chosen with a size of 2 and 1024 MB. The choice of BOM and EBOM was due to the fact both algorithms have the same structure with EBOM having a faster loop. However, the choice of algorithms is arbitrary. One can choose any other algorithm for this purpose as the main aim is to check how much impact does other factors like pattern length, dataset have on execution time. (3) Design Phase: In this phase, all abovementioned factors with their influence on execution time and the factorial model was mapped together to get the complete picture of the whole process. The list of factors used for the purpose of the research is shown in Table 1. (4) Factorial Design Phase: In this phase, the 2k factorial design model was designed. To evaluate the results based on above data obtained in data collection phase, the factorial design is designed as shown below to suit our scenario of knowing the effect of different factors on the performance of an algorithm which is usually based on execution time. All 3 factors with levels and results obtained from the work of Simone [5] is shown in Table 2.

1 Factorial Analysis of Biological Datasets

7

Table 2 Factorial design model B (−1)

B (1)

Database type (genome sequence database)

Database type (protein sequence database)

A

C

C

C

C

Pattern matching algorithm

Input size

Input size

Input size

Input size

−1

1

−1

1

2 MB

1024 MB

2 MB

1024 MB

BOM (1)

136.3

2.05

26.49

0.73

EBOM (−1)

49.77

3.42

12.13

2.57

4 Results From the calculations involved in Table 3 and results obtained in Table 4, it can be concluded the most significant factor for obtaining optimal performance of an algorithm is pattern length. Pattern length is having the effect of 38.5% on the execution time of an algorithm. The second most important factor is the type of database used. The database is having the effect of 18.50% on the execution time of an algorithm. Together pattern length and database are having the influence of 17.40% on the execution time. Displayed equations are centered and set on a separate line. Using Eq. 2, 

SST = 2

3

(12.21)2 + (−18.705) + (−26.99) + (−9.08) + (−13.0125) + (18.16) + (8.9625)



Therefore, SST = 15113.69015.

Table 3 Calculations involved using factorial design I

A

B

C

Y

AB

AC

BC

ABC

1

−1

−1

−1

49.77

1

1

1

−1

1

1

−1

−1

136.3

−1

−1

1

1

1

−1

1

−1

12.13

−1

1

−1

1

1

1

1

−1

26.49

1

−1

−1

−1

1

−1

−1

1

3.42

1

−1

−1

1

1

1

−1

1

2.05

−1

1

−1

−1

1

−1

1

1

2.57

−1

−1

1

−1

1

1

1

1

0.73

1

1

1

1

Total

97.68

−149.62

−215.92

0

−72.64

−104.1

145.28

71.7

Total

12.21

−18.7025

−26.99

−9.08

−13.0125

18.16

8.9625

8 Table 4 Effect of different factors on execution time

H. M. S. Parvez et al.

Effect of algorithm selection

7.89%

Effect of database

18.50%

Effect of pattern length

38.50%

Effect of algorithm and database

4.30%

Interaction b/w algorithm and pattern length

8.96%

Interaction b/w pattern length and database

17.40%

Interaction b/w algorithm, database, and input

4.25%

5 Conclusion In this study, the problem of designing algorithms without knowing the effect of other factors on the execution time was highlighted. Only biological datasets were considered due to vast scope and research in the area of algorithms. A brief description of experimental design techniques particularly simple design, full factorial design, and fractional design technique was given. Finally, the reason for selecting 2k factorial was also mentioned along with the methodology used in the research. The 2k factorial model was designed to suit the requirements of the present research and three parameters were considered which includes pattern length, dataset type, and patternbased algorithms. Results show that pattern length and type of datasets are important factors, which need to be taken into consideration while designing or proposing an algorithm. Pattern length alone has 38.5% effect on execution time with the type of dataset used 18.5%. These empirical values are bound to change while operating on different platforms due to processing-related issues. But overall, the conclusion will be the same in any case. This study further gives a view that algorithm which is to be proposed or designed should be target based like the same algorithm cannot be used for analyzing DNA and protein bases. The algorithm needs to be specific with a clear idea in mind which pattern length to use and on what dataset. In the future, it will be an interesting research, to keep the pattern length the same and vary different algorithms to check the performance of different algorithms using biological datasets. This way, the optimal length of a pattern can be obtained which may be used by other algorithms for optimal performance. Acknowledgements This research is supported by FRGS FP003A-2017, Faculty of Computer Science and Information Technology, University of Malaya.

References 1. Egholm M, Margulies M, Altman W, Attiya S, Bader J, Bemben L et al (2005) Genome sequencing in open microfabricated high density picoliter reactors. Nature 437:376–380 2. E.P. Consortium (2004) The ENCODE (ENCyclopedia of DNA elements) project. Science 306:636–640

1 Factorial Analysis of Biological Datasets

9

3. G.P. Consortium (2010) A map of human genome variation from population-scale sequencing. Nature 467:1061 4. Yang Z, Yu J, Kitsuregawa M (2010) Fast algorithms for top-k approximate string matching. AAAI 5. Faro S, Lecroq T (2013) The exact online string matching problem: a review of the most recent results. ACM Comput Surv (CSUR) 45:13 6. Hakak S, Kamsin A, Shivakumara P, Idris MYI (2018) Partition-based pattern matching approach for efficient retrieval of Arabic text. Malays J Comput Sci 31:200–209 7. Hakak S, Kamsin A, Tayan O, Idris MYI, Gani A, Zerdoumi S (2017) Preserving content integrity of digital holy Quran: survey and open challenges. IEEE Access 5:7305–7325 8. Hakak S, Kamsin A, Palaiahnakote S, Tayan O, Idris MYI, Abukhir KZ (2018) Residualbased approach for authenticating pattern of multi-style diacritical Arabic texts. PLoS ONE 13:e0198284 9. Hakak S, Kamsin A, Tayan O, Idris MYI, Gilkar GA (2017) Approaches for preserving content integrity of sensitive online Arabic content: a survey and research challenges. Inf Process Manag 10. Zerdoumi S, Sabri AQM, Kamsin A, Hashem IAT, Gani A, Hakak S et al (2017) Image pattern recognition in big data: taxonomy and open challenges: survey. Multimed Tools Appl 1–31 11. Hakak S, Kamsin A, Shivakumara P, Idris MYI, Gilkar GA (2018) A new split based searching for exact pattern matching for natural texts. PLoS ONE 13:e0200912 12. Hakak SI (2015) Evaluating the effect of routing protocol, packet size and DSSS rate on network performance indicators in MANET’s. Kulliyyah of Engineering, International Islamic University Malaysia 13. Allauzen C, Crochemore M, Raffinot M (1999) Factor oracle: a new structure for pattern matching. In: International conference on current trends in theory and practice of computer science. Springer, pp 295–310 14. Faro S, Lecroq T (2009) Efficient variants of the backward-oracle-matching algorithm. Int J Found Comput Sci 20:967–984 15. Khan ZA, Pateriya R (2012) Multiple pattern string matching methodologies: a comparative analysis. Int J Sci Res Publ 2:1–7 16. Altschul SF, Gish W, Miller W, Myers EW, Lipman DJ (1990) Basic local alignment search tool. J Mol Biol 215:403–410 17. Berry T, Ravindran S (1999) A fast string matching algorithm and experimental results. Stringology 16–28 18. Hudaib A, Al-Khalid R, Suleiman D, Itriq M, Al-Anani A (2008) A fast pattern matching algorithm with two sliding windows (TSW). J Comput Sci 4:393 19. Itriq M, Hudaib A, Al-Anani A, Al-Khalid R, Suleiman D (2012) Enhanced two sliding windows algorithm for pattern matching (ETSW). J Am Sci 8:607–616 20. Hakak S, Anwar F, Latif SA, Gilkar G, Alam M (2014) Impact of packet size and node mobility pause time on average end to end delay and jitter in MANET’s. In: 2014 International conference on computer and communication engineering (ICCCE). IEEE, pp 56–59 21. Hakak S, Latif SA, Anwar F, Alam MK (2014) Impact of key factors on average jitter in MANET. In: First international conference on systems informatics, modeling and simulation computer society. IEEE, pp 179–183 22. Hakak S, Latif SA, Anwar F, Alam M, Gilkar G (2014) Effect of mobility model and packet size on throughput in MANET’s. In: 2014 International conference on computer and communication engineering (ICCCE). IEEE, pp 150–153 23. Hakak S, Latif SA, Anwar F, Alam M, Gilkar G (2014) Effect of 3 key factors on average end to end delay and jitter in MANET. J ICT Res Appl 8:113–125 24. Jain R (1990) The art of computer systems performance analysis: techniques for experimental design, measurement, simulation, and modeling. Wiley

Chapter 2

Classification of Motor Imagery Events from Prefrontal Hemodynamics for BCI Application Md. Asadur Rahman, Md. Mahmudul Haque, Anika Anjum, Md. Nurunnabi Mollah and Mohiuddin Ahmad

1 Introduction Brain–computer interface (BCI) is a remarkable notion of current neurocomputational research that offers the scope to control the computer with the brain command. One of the most important applications of the BCI system is neurorehabilitation that implicates brain commanding devices for the paralyzed or physically challenged persons [1]. A successful BCI implementation demands precise noninvasive functional neuroimaging for construing the functional activities of the brain. The brain functionalities can be assessed noninvasively by either the electric impulsive signals or the hemodynamics of the brain tissue. For the functional neuroimaging, electroencephalography (EEG) and magneto-encephalography (MEG) are the well-known noninvasive methods based on electric impulsive signals of the brain. Due to some questionable features like noise sensitivity, poor spatial resoluMd. Asadur Rahman (B) · Md. Mahmudul Haque · A. Anjum Department of Biomedical Engineering, Khulna University of Engineering & Technology (KUET), 9203 Khulna, Bangladesh e-mail: [email protected] Md. Mahmudul Haque e-mail: [email protected] A. Anjum e-mail: [email protected] Md. Nurunnabi Mollah · M. Ahmad Department of Electrical and Electronic Engineering, Khulna University of Engineering & Technology (KUET), 9203 Khulna, Bangladesh e-mail: [email protected] M. Ahmad e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_2

11

12

Md. Asadur Rahman et al.

tion, motion sensitivity, these modalities should be replaced by some newer one [2, 3]. The hemodynamic method can be a suitable alternative instead of the electric impulsive signal. Functional near-infrared spectroscopy (fNIRS) is such a noninvasive optical method that is able to measure the functional neuro-activations based on hemodynamics of the brain tissue [4]. The planning for limb movement or imagery movement is a basic task for BCI and that is why considering motor imagery events classification through fNIRS, a number of research works have been accomplished, particularly, imagery hand movement classification. Both voluntary and imagery movements of the right and left hand were considered for classification from fNIRS data in [5]. Imagery movements related events of hemodynamic response are shown in these research works [6]. Multiple motor imagery events like left hand, right hand, left foot, and right foot movements are classified in [7]. The imagery movements of the left and right arm are classified by support vector machine algorithm in [8] with the fNIRS data for BCI application. All of the research works acquired the hemodynamic information from the motor cortex because this area of the human brain controls the voluntary movements. Nonetheless, it is often found that this motor area of paralyzed patients is either injured or fully damaged. Therefore, practical BCI related to the paralyzed persons cannot be achieved by the previous proposed works. This limitation creates a scope of research to find a possible solution. This research work proposes that the hemodynamics of the prefrontal cortex can be the possible solution to the aforesaid limitation. However, it is reported in [9, 10] that the motor action planning is introduced in the prefrontal lobe of the brain. In addition, the research work demonstrated that prefrontal hemodynamics has a strong correlation with voluntary hand movements [11] and these events can be classified with acceptable accuracy [12]. So, motor action planning or imagery movement may be decoded from the hemodynamic information of the prefrontal cortex. From this point of view, the paradigm of this proposed research work is developed to classify the imagery movements. Acquisition of motor imagery signal and classifying it would accelerate the suitable implementation of BCI for paralyzed people. In this research work, two classes (left hand and right hand) of motor imagery movements and their corresponding blood oxygen level-dependent (BOLD) signals have been recorded from prefrontal cortex using 16 channel fNIR devices. With the help of prominent preprocessing of the signals, features have been extracted according to the imagery events. The extracted features of some selected signals were used for the training of the network. In this work, we have used two classifiers: k-nearest neighbor (kNN) algorithm and artificial neural network (ANN). For the training and testing, features of the different dataset were used. The results have been presented with both subject-dependent and subject-independent approach. Our results convince that motor planning events can be significantly classified from the prefrontal BOLD signals.

2 Classification of Motor Imagery Events from Prefrontal …

13

2 Experimental Materials 2.1 Data Acquisition Protocol In this experiment, two types of events were performed sequentially (imagery left hand (LH) and right hand (RH) movements). Each adult participant was seated comfortably on a chair facing a computer monitor at a distance of 68–73 cm. At first, the subject was asked to relax at least 5 min so that his heart rate, blood pressure, and other involuntary issues remain in normal condition. The data acquisition trial for each task was taken 5 times in a session and 6 sessions were performed by each participant. The task scheduling pattern is given in Fig. 1. Therefore, 30 trials for each task (LH and RH) were performed by each participant.

2.2 Participants Five healthy adult male subjects participated (mean age 22.5 ± 1.5 years) in this experiment. All of the participants were right handed based on the recommendation of Edinburg Handedness Inventory [13] to minimize the variation of hemodynamic responses due to hemispheric dominance difference. Participants did not have any psychiatric, neurological, and visual disorder. All of them were provided with verbal consent and informed about the experimental procedure. The experiment was conducted as the Declaration of Helsinki [14].

2.3 Hemodynamic Data Acquisition by fNIRS In this work, the fNIR device we used has a headband of 16 channels for prefrontal hemodynamic data acquisition. The optode configuration and channel distribution of the fNIR devices (model 1200) headband are given in Fig. 2. The headband is attached to the skin of the prefrontal cortex (PFC). The positions of the light sources and detectors in the PFC of the human brain are presented in Fig. 3a. The headband is placed on the forehead in such a way that the bottom row of detectors is located just above the eyebrows as given in Fig. 3b. In addition, Fig. 3b presents the setup of data acquiring hardware with participant, imager, and computer.

Rest 15 sec

Left hand 3 sec

Rest 15 sec

Right hand 3 sec

Repeat…

Fig. 1 The schedule of the imagery task and rest regarding the experimental paradigm was conducted as a figure. This procedure was repeated 5 times in every session

14

Md. Asadur Rahman et al.

Fig. 2 The yellow-colored circles are indicating the detectors of the headband and the red circles are indicating 4 optical sources in this figure. The numbers given inside the headband represent the channels of this device

(a)

(b) fNIR Imager 1200

Fig. 3 a NIR emitter and detector placement upon the prefrontal cortex, and b data acquisition procedure utilized the 16 channel fNIR devices (model 1200)

Due to the neurovascular coupling, the functional activity of a human can be assessed from the concentration change of oxidized hemoglobin (HbO2 ) and deoxidized hemoglobin (Hb) in the brain. The fNIR devices can measure the concentration change of HbO2 and Hb. This device uses NIR (650–1000 nm) light and the absorption coefficient of HbO2 and Hb are different in this range. Suppose, the intensity of the injected NIR light by the fNIR device is i0 and the intensity of the scattered backlight after absorption by the chromophores (HbO2 and Hb) of the brain tissue is i. The relation between i and i0 is expressed by modified Beer–Lambert law (mBLL) as [15] i =i0 10−ζλ or, ζλ = − log

i i0

(1)

where ζλ is the optical density which depends on the wavelength, λ. In the case of continuous-wave fNIRS, the change in concentrations of the two chromophores are calculated as [16] 

  −1   1 ζλ 1 CHbO2 (t) αHbO2 (λ1 ) αHb (λ1 ) = CHb (t) αHbO2 (λ2 ) αHb (λ2 ) ζλ2 dPL × DEM

(2)

2 Classification of Motor Imagery Events from Prefrontal …

15

In (2) ζλ1 and ζλ2 are optical density for two corresponding NIR wavelengths, λ1 and λ2 . Here, α is the excitation coefficient of HbO2 and Hb in μM−1 mm−1 , d PL is the unit-less differential pathlength factor, and DEM is the distance between emitter and detector. The unit of DEM is in mm. Besides, C is the change in concentrations of HbO2 and Hb in μMole.

3 Methodology From data acquisition to classify for the performance test of the proposed work, a number of processing steps are needed. In this section, all the processing steps concerning this working methodology have been explained. The offline data processing methodology of this work has been presented with all the steps in Fig. 4 which summarizes the entire work.

3.1 Compression Since the data was acquired with 16 channels, the processing could have faced the curse of high dimensionality. To check the actual dimensions, the data were transformed by principal component analysis (PCA) and we found that the actual dimensions of the signals are 4 instead of 16. The result of PCA has been shown in Fig. 5 with the similar channel numbers of ilk dimension. The resulting 4 dimensions are termed as Left Lateral (LL: channel 1, 2, 3, 4), Left Medial (LM: channel 5, 6, 7, 8), Right Medial (RM: channels 9, 10, 11, 12), Right Lateral (RL: channels 13, 14, 15, 16). The signals having the same dimensions were averaged. Therefore, the 16 channel information become compressed from i × 16 to i × 4. This compression helps to reduce the feature dimension which is very important for achieving high classification accuracy in the machine learning approach.

3.2 Filtering and Baseline Correction The raw fNIRS signals were filtered by Savitzky–Golay filter of 3rd order with 21 frame size (equivalent to 0.1 Hz cutoff FIR lowpass filter in 2 Hz fNIR signal [17]) to remove physiological noises like heart rate, respiration, etc. The noisy raw and filtered fNIR signal by the method is given in Fig. 6. After filtering the data, the tasks were separated. Each trial of the fNIRS data was corrected by subtracting the baseline from the original signal. Baseline was calculated from the average of the first three seconds of the task.

16

Md. Asadur Rahman et al.

Start Load 16 channel data

Database

Data compression Filtering Baseline correction Separate the signals for two different events Left hand

Right hand

Feature extraction

Feature extraction Train kNN/ANN to create a predictive model

kNN/ANN based predictive model

Data for testing

Classifier

Accuracy (%)

Fig. 4 Block diagram of the total offline signal processing steps to create the machine learning based predictive model and finding the accuracy of the testing data

3.3 Feature Extraction Since the fNIRS signal does not exhibit very complexity like EEG, MEG, and fMRI, simple time domain features are enough to represent the signal characteristics [18]. In this experiment, we choose two important features—mean and slope to extract from the signal. Before training the machine learning-based network, all the values of training features were normalized between 0 and 1 by (3). π =

π − πmin πmax − πmin

(3)

2 Classification of Motor Imagery Events from Prefrontal …

Component 2

0.4

17

8 7 6 5

12 10

0.2

11

9

14 15 16 13

0

2

-0.2

1

4 3 -0.4 -0.1

-0.05

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

Component 1

Signal Amplitude in micromole/liter

Fig. 5 Principal component 1 versus component 2 presents the actual signal dimensions in Eigenspace. The figure represents four linearly uncorrelated spatial positions in solution space 8

Raw fNIR Signal Filtered fNIR Signal

6 4 2 0 -2 -4 -6

5

10

15

20

25

30

35

40

45

50

Number of Samples

Fig. 6 The raw fNIRS signal with the physiological artifacts and the filtered signal after SG filtering technique

Here, π  are the feature values that are rescaled between the range 0 and 1. On the other hand, π  ∈Rn are the actual values of the features. The maximum and minimum value of the features are presented as πmax and πmin , respectively.

3.4 kNN Classifier To classify the feature set of imagery hand movement fNIRS signals k-nearest neighbor method has been used in this work. The kNN algorithm is a nonparametric estimation method that implements a refinement, where the feature environment is high resolution in regions with dense training and low resolution in variance. The method works based on the following steps. Suppose, Z(z) ⊂ RN is a hypersphere of volume, ν and center, z. Let us consider that the training set S k consists of N k number of samples. Therefore, according to the binomial distribution, the probability of including

18

Md. Asadur Rahman et al.

precisely n samples within the hypersphere, Z(z) can be equated by the following expectation [19]:  p(y|wk ) ≈ Nk ν p(z|wk ) E[n] = Nk (4) y∈Z(z)

Here, wk represents the number of class. The radius, z of the sphere Z is selected in a way that Z contains exactly κ samples. In this sense, the volume becomes the function of z and it can be presented as ν(z). So, the estimation of density based on the number of κ can be calculated as 

p(z|wk ) =

κ Nk ν(z)

(5)

The parameter κcan control the stability between the variance and bias. To classify any vector zˆ , the radius of the sphere is selected such a way that the sphere contains exactly κsamples occupied from S k . These samples are termed as the κ-nearest neighbor.

3.5 ANN-Based Classifier The ANN is a complex and very efficient classifier. This algorithm was also used in this research work to find the highest classification accuracy of the feature set. ANN has the quality to mimic the comportment of the human brain. For the feedforward networks, commonly multilayer perceptors consist of three type layers: input, output, and hidden layers. The objective of the input layer is to buffer the distribution of the input signals x n (n = 1, 2, 3, …) toward the hidden layer neurons. Each hidden layer neuron adds the input signals (x n ) after weighing the input signals by the strengths W nj from the input layer and calculate the output, Y where Y is a function of their summations [12].  n   Yj = f Wjn xn (6) n=1

Here j is neuron numbers, W jn is the weight of a connection between n and j according to their relation, Wjn = ηδj xn . Here η is the rate of learning parameter and δ j depends on whether j is an input or hidden neuron. The adjustments of the weights are generally estimated by the backpropagation algorithm. Let V i be the prediction for jth observation in an ANN system and is a function of the network weights vector W = (W1 , W2 , . . .). Therefore, e, the total prediction error will also be a function of W 2

as e(W ) = Yj − Vj (W ) . For every weight W 1 , according to the gradient descent

2 Classification of Motor Imagery Events from Prefrontal …

19

algorithm the updating formula is considered as, Wnew = Wold +α(∂e/∂ W )Wold . Here α is the learning parameter and the value of α lies between 0 and 1.

4 Results and Discussions The activation due to imagery movement planning occurs in the prefrontal cortex which can be measured by the increased concentration of HbO2 . For the LH and RH imagery movement planning, the activation level is found to be increased in the concentration of HbO2 in the right hemisphere and left hemisphere, respectively. Since four different portions are selected to observe the actual concentration change in HbO2 of LH and RH movement planning. The variation of HbO2 concentration in LL, LM, RM, and RL portion of the prefrontal cortex due to planning movements of LH and RH of a randomly selected subject have been presented in Fig. 7. It is easily observable from Fig. 7 that, all the proposed portions on the PFC exhibit significant variations in the concentration of HbO2 .

(a)

4

(b) 15

LH Planning

LH Planning RH Planning

2

Micromole/liter

Micromole/liter

RH Planning

0

-2

-4 0

10

20

5

0

-5 0

40

30

10

10

(c)

6

20

30

40

Sample Number

Sample Number

5

(d)

Micromole/liter

Micromole/liter

4 2 0

0

-5

-2 LH Planning RH Planning

-4 0

10

20

Sample Number

30

40

-10 0

LH Planning RH Planning

10

20

30

40

Sample Number

Fig. 7 The comparison of the concentration change over time in HbO2 between LH and RH movement planning in LL (a), LM (b), RM (c), and RL (d) portion of the prefrontal cortex

20

Md. Asadur Rahman et al.

As a result, all these portions should be considered as a significant source to discriminate the imagery activities. Accordingly, features were extracted from the fNIRS signals of LL, LM, RM, and RL. The features were used to train kNN separately to check the feature-dependent accuracy of the predictive model. Several time domain features like variance, total summation, kurtosis did not return good classification accuracy except mean and slope. Therefore, the kNN-based predictive model was constructed by the feature mean and slope and tested the subject dependent classifying accuracy. The classifying accuracy for one subject for the slope and mean based kNN model can be found from the confusion matrix of their classification accuracy given in Fig. 8a and b, respectively. From the confusion matrix, it is found that both have the classifying accuracy of 77.5%. We hypothesize that with the help of both features prediction should be increased. To deal with the high-dimensional feature set, the ANN-based predictive model was constructed with the features of 10 trials and the predictive model was tested with the rest 20 trails. The results corresponding to the same subject’s data used in the kNN-based model are given by Fig. 9a. Figure 9b presents the classification accuracy of another subject. The classification performance of five subjects by kNN and ANN are given in Table 1. These results are showing the subject dependent classification accuracy. Along with the classification accuracy, the sensitivity and specificity of the ANN-based model have also been calculated as Sensitivity =

TN TP and Specificity = TP + FN TN + FP

(7)

Here, TP = true positive, TN = true negative, FP = false positive, and FN = false negative. For the subject independent approach, we have only used the ANN-based predictive model. From all data of five subjects, the features of LH and RH were mixed and independent to the subject. Therefore, we have 300 sets of features in which 150 trials for each class. From all these feature set, 100 trials (50 LH + 50 RH) were used to train the network. The rest trials were used for testing. From the results, we

(a)

(b)

Fig. 8 Confusion matrix of the classification performance based on slope (a) and mean (b) of the subject 1 by the kNN-based predictive model. This model is trained with 10 trials and tested with 20 trials of LH and RH fNIRS data

2 Classification of Motor Imagery Events from Prefrontal …

(a)

21

(b)

LH

18 45.0%

RH

2

90.0%

5.0%

10.0%

2

18

90.0%

5.0%

45.0%

10.0%

90.0%

90.0%

90.0%

10.0%

10.0%

10.0%

LH

LH

19

2

90.5%

47.5%

5.0%

9.5%

1

18

94.7%

2.5%

45.0%

5.3%

95.0%

90.0%

92.5%

5.0%

10.0%

7.5%

RH

RH

LH

RH

Fig. 9 Figures present confusion matrix of the classification performance of the ANN-based predictive model. These classification accuracies of the two subjects are given in (a) and (b) Table 1 Subject-dependent classification accuracy by kNN and ANN

Sub no.

Results of a kNN-based predictive model Accuracy (with slope) (%)

Accuracy (with mean) (%)

1

77.5

77.5

2

80

78 65

3

69.5

4

72

74.5

5

81.5

82

Sub no.

Results of the ANN-based predictive model Accuracy (%)

Specificity (%)

Sensitivity (%) 90

1

90

90

2

92.5

90

95

3

82.5

80.9

84.21

4

87.5

85.7

89.4

5

88.25

82.5

94

found that in the case of subject-independent approaches the classification accuracy is 75% with 80% sensitivity and 70% specificity. The results have been given by the confusion matrix in Fig. 10.

5 Conclusions This research work has verified the suitability of motor imagery event classification of prefrontal cortex fNIR data. From the total results, it is found that two class motor movement planning hemodynamics of the prefrontal cortex can be classified with

22

Md. Asadur Rahman et al.

Fig. 10 Presentation of the classification accuracy (%), sensitivity (%), and specificity (%) for the subject independent approach by the resulting confusion matrix

significant accuracy by the ANN-based predictive model considering both subject dependent and subject independent approach. Therefore, this research work suggests that motor imagery events can be classified from prefrontal hemodynamics and can be used to implement effective BCI. In this work, all subjects were young and healthy. The prefrontal hemodynamics due to movement planning task (or motor imagery task) can be different for the aged and paralyzed patient. Considering this aspect as a limitation, we are interested in research on the prefrontal hemodynamic pattern regarding the motor imagery tasks by the aged and paralyzed patients in the future. Acknowledgements In this paper, the fNIRS data were taken from five subjects: (i) Md. Al Noman, (ii) Md. Torikul Islam Piash, (iii) Md. Ahnaf Tahmid Ifty, (iv) Md. Sohidujjaman Sojib, and (v) Md. Abu Bakar Siddik Yusha. The authors would like to thank them for providing their kind consent to use their data in this research work. Moreover, special thanks to Md. Al Noman for giving his consent to use his picture in Fig. 2b. The data acquisition protocol and the consents of the participants were ethically approved by Data Acquiring Ethics Evaluation Committee (DAEEC) of Khulna University of Engineering & Technology (KUET), Bangladesh.

References 1. Mihara M, Miyai I (2016) Review of functional near-infrared spectroscopy in neurorehabilitation. Neurophotonics 3:031414-8 2. Burle B, Spieser L, Roger C, Casini L, Hasbroucq T, Vidal F (2015) Spatial and temporal resolutions of EEG: is it really black and white? A scalp current density view. Int J Psychol 97:210–220

2 Classification of Motor Imagery Events from Prefrontal …

23

3. Basic principles of magnetoencephalography. MIT Class notes. http://web.mit.edu/kitmitmeg/ whatis.html 4. Ayaz H (2010) Functional near infrared spectroscopy-based brain computer interface. PhD thesis, Derexel University, Philadelphia (2010) 5. Robinson N, Zaidi AD, Rana M, Prasad VA, Guan C, Birbaumer N, Sitaram R (2016) Real-time subject-independent pattern classification of overt and covert movements from fNIRS signals. PLoS ONE 11:1–21 6. Hong KS, Naseer N, Kim YH (2015) Classification of prefrontal and motor cortex signals for three-class fNIRS–BCI. Neurosci Lett 587:87–92 7. Batula AM, Kim YE, Ayaz H (2017) Virtual and actual humanoid robot control with four-class motor-imagery-based optical brain-computer interface. BioMed Res Int 2017:1–13 8. Janani A, Sasikala M (2018) Evaluation of classification performance of functional near infrared spectroscopy signals during movement execution for developing a brain–computer interface application using optimal channels. J Near-Infrared Spectrosc 26:209–221 9. Hong KS, Khan MJ, Hong MJ (2018) Feature extraction and classification methods for hybrid fNIRS-EEG brain-computer interfaces. Front Hum Neurosci 12:1–25 10. Carlson RN (1994) Physiology of behavior, 5th edn. US Allyn & Bacon, Needham Heights, MA 11. Rahman MA, Ahmad M (2017) Evaluating the connectivity of motor area with prefrontal cortex by fNIR spectroscopy. In: International conference on electrical, computer and communication engineering, pp 16–18 12. Rahman MA, Ahmad M (2016) Movement related events classification from functional near infrared spectroscopic signal. In: 19th international conference on computer and information technology, pp 207–212 13. Oldfield RC (1971) The assessment and analysis of handedness: the Edinburgh inventory. Neuro-psychologia 9:97–113 14. World Medical Association declaration of Helsinki-ethical principles for medical research involving human subjects, Adopted by 64th WMA General Assembly, Fortaleza, Brazil, Special Communication: Clinical Review & Education (2013) 15. Rahman MA, Ahmad M (2016) Identifying appropriate feature to distinguish between resting and active condition from FNIRS. In: International conference on signal processing and integrated networks, pp 1–5 16. Rahman MA, Ahmad M (2016) Lie detection from fNIR signal and neuroimage. In: International conference on medical engineering, health informatics and technology, pp 1–6 17. Rahman MA, Ahmad M (in press) Selecting the optimal conditions of savitzky-golay filter for fNIRS signal. Biomed Signal Process Control 18. Kabir MF, Islam SMR, Rahman MA (2018) Accuracy improvement of fNIRS based motor imagery movement classification by standardized common spatial pattern. In: International conference on electrical engineering and information and communication technology, pp 1–6 19. Heijden FV, Duin R. PW, Ridder D, Tax DMJ (2004) Classification, parameter estimation and state estimation: an engineering approach using MATLAB. Wiley

Chapter 3

Diabetic Retinopathy Detection Using PCA-SIFT and Weighted Decision Tree Fatema T. Johora, Md. Mahbub -Or-Rashid, Mohammad A. Yousuf, Tumpa Rani Saha and Bulbul Ahmed

1 Introduction As diabetes is a metabolic disease, therefore the body is unable to produce insulin which eventually increases the glucose level in the blood. When the glucose level of the blood vessel in retina is increased the vision becomes blurred and without proper treatment it can lead to complete blindness, this process of damage within the retina is called diabetic retinopathy. Excess amount of glucose in the blood vessel may lead to anomalies like micro aneurysms, hemorrhages, hard exudates and Cotton wool spots develop during the different phases of diabetic retinopathy [1, 2]. According to a study which was conducted by World Health Organization, it shows that over the next 20 years the number of diabetes patients will rise up to 350 million [3, 4]. In developed countries, one of the alarming causes of blindness is diabetic retinopathy [5] and for the developing countries this problem is even more dangerous as they do not have the proper screening technologies to overcome the prevention from this disease, however 75% of the people with diabetic retinopathy lives in the developing countries [6]. The number of patient with diabetic retinopathy is increasing, which will increase the workload for the ophthalmologist because most of their time will be spent to detect diabetic retinopathy. The two types of diabetic retinopathy are 1. NPDR (Non F. T. Johora (B) · Md. Mahbub -Or-Rashid · M. A. Yousuf · B. Ahmed Jahangirnagar University, Savar, Dhaka, Bangladesh e-mail: [email protected] Md. Mahbub -Or-Rashid e-mail: [email protected] M. A. Yousuf e-mail: [email protected] T. R. Saha Bangbabandhu Sheikh Mujibur Rahman Science and Technology University, Gopalganj, Bangladesh © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_3

25

26

F. T. Johora et al.

proliferative diabetic retinopathy) 2. PDR (proliferative diabetic retinopathy). Where NPDR (Non proliferative diabetic retinopathy) can be subdivided into 1. Mild non proliferative diabetic retinopathy 2. Moderate Non proliferative diabetic retinopathy 3. Severe non proliferative diabetic retinopathy. Through this work we have tried to develop such a system that will be able to detect the severity more efficiently and more accurately.

2 Background Study In recent years, many diabetic retinopathy detection methods are proposed. A computer-based approach was proposed to detect the diabetic retinopathy stage with the help of fundus images [7]. For classification they used Artificial Neural Network (ANN). In [8], their main task was to detect the retinal changes of diabetes patient’s eye using Convolutional Neural Network (CNN). In [9], the authors proposed a system which utilizes morphological operations such as opening, closing, erode and dilate for feature extraction and for blood vessel detection they have used kirsch’s edge detector. They have used fuzzy image processing for detection of diabetic retinopathy. In [10] at first they enhance the images and then curve let transformation is applied to equalize the images. These pre-processed images are used for extraction of the blood vessels. In [11], exudates in color fundus were detected as well as classify the severity of the lesions using Support Vector Machine (SVM). In [12], the authors also did the same type of work as that of [11] but in their work they have used ANN classifier. In [13], they used adaptive histogram equalization method for the efficient detection of diabetic retinopathy. In [14], the authors proposed a method which consist of a novel sparking process and a holo-entropy based decision tree for automatic diabetic retinopathy detection. Lahmiri and Boukadoum presented a new and simple automated system to detect exudates in retina digital images [15]. Here a PCA-SIFT based feature extraction approach is proposed in combination with Naive Bayes classifier for noise free images and prior and posterior probability based weighted decision tree for noisy images to detect diabetic retinopathy.

3 Proposed Model The proposed system consists of preprocessing of image, feature extraction using PCA based SIFT (or PCA-SIFT based feature extraction), and then classification (Naïve Bayesian classifier for noise free dataset and prior and posterior probability based decision tree for noisy dataset). Figure 1 shows the overall flow of our diabetic Retinopathy detection system.

3 Diabetic Retinopathy Detection …

27

Fig. 1 Proposed system model

A. Image Processing The proposed system starts with the conversion of RGB images into HSI images. The reason behind to use preprocessing is to reduce the noise level of images and to produce more accurate result. Image preprocessing includes filtering based on adaptive noise detector and then optic disc detection and elimination. (1) Filtering Then we have applied median filter because this filter is more robust method than the traditional filtering as it preserves the sharp edges. After applying the filter, we have checked the noise range. This is done by checking the histogram of the filtered image and the original image. We have separated those images whose noise range is greater than defined threshold (th) and rest of the images whose noise range is below the threshold are named as noise free images. Figure 2 shows the image preprocessing steps. (2) Optical Disk Detection and Elimination The marker-controlled watershed segmentation is used for optic disc detection as it is flexible and robust method to segment the objects with closed contours, where the boundaries are illustrated as ridges [16]. Later on, the watershed regions’ boundaries are organized on the desired ridges, thus separating each object from its neighbors. To eliminate the optic disc, the HDR image obtained after preprocessing is transformed into binary image as processing of colorful image is more complex. Every pixel with intensity value greater than or equal to 195 and less than 195 are converted to 0 and

28

F. T. Johora et al.

Fig. 2 Image preprocessing

Fig. 3 Optic disc eliminated image

1 respectively. After that all connected components having fewer than 20 pixels are removed from the binary image. Then the operation of flood-filling is performed by filling the ‘holes’ in the background image. Then we have suppressed the lighter structures with compare to their surroundings which are attached to the image border and inverted the image and computed the Euclidean distance transform of the binary image. We have neglected the pixels of the entire matrix, i.e., pixels with a value of 0 will remain unchanged but pixels with 1 change into −1 and assigned nonobject pixels to be—Infinity and eliminated regional maxima by specifying several threshold values such as 2, 4, 8 etc. Figure 3 shows the Optic disc eliminated image.

3 Diabetic Retinopathy Detection …

29

Fig. 4 Pictorial representation of DoG

B. Scale-Space Extrema Detection For scale-space Extrema detection it generates several octaves of the original image and each octave if half of the previous one. Within an image octaves are progressively blurred using Gaussian Blur operator. Expression for Gaussian Blur is determined by the Eqs. (1) and (2) L(x, y, σ ) = G(x, y, σ )∗ I (x, y) G(x, y, σ ) =

1 − x 2 +y2 2 e 2σ 2σ 2 π

(1) (2)

where, L represents the Blurred Image, G stands for Gaussian Blur operator, I is an image, (x, y) is the location coordinates, σ is the scale parameter, ∗ is the convolution operator. Using images obtained using the Eqs. (1) and (2) it creates another set of images called the difference of Gaussian (DoG). It’s similar to calculate the second order derivatives on that blurred image. It locates edges and corner on the image which is good for finding key points. Figure 4 demonstrates the pictorial representation of DoG. C. Key Point Localizations (1) Locate Maxima/Minima in DoG Image In this step, SIFT iterate through each pixel and check all its neighbor. The yellow circles mark the neighbors of each current pixel (X). This way, a total of 26 checks are made. X is marked as a “key point” if it is the greatest or least of all 26 neighbors. Key points of retinal image for training and test data is shown in Fig. 5.

30

F. T. Johora et al.

Fig. 5 Key points of retinal image for training data (a) and test data (b)

(2) Find Sub Pixel Maxima/Minima The values of sub pixel are generated using the available pixel data. This is done by the Taylor expansion of the image around the approximate key point using Eq. (3). D(x) = D + x

(3)

where, D ……, x…… The pixels that lie along an edge not having enough contrast at the necessary pixel (in the DoG image are rejected here). A. PCA-SIFT Key Point Descriptor (1) Computation of Eigen Space In SIFT, a 16 × 16 neighborhood around the key point is considered. It is separated into 16 sub-blocks with a size of 4 × 4. 8 bin orientation histogram is created for each of the sub-blocks. So, 128 bin values in total are available. It is represented as a vector to form key point descriptor and extracted 51 * 51 patches at the given scale that bring together over the key points. Then it is being rotated to align its orientation to an acceptable direction. Computation of eigenspace of each key point is shown in Algorithm 1. Algorithm 1: Computation of eigen space of each key point. Notations: Extr_volume = Total number of keypoints found D = contains all the Difference of the Gaussian (DoG) images Dx = difference in the x direction (horizontal gradient vectors) Dy = difference in the y direction (vertical gradient vectors) A = one dimentional array contains all gradients of the patches G = one dimentional array contains gradient of a single patch Flag = array contains keypoints position/index which are to be deleted

3 Diabetic Retinopathy Detection …

31

N = number of eigenvectors to be stored and used in projection Project = contains eigenspaces (projection matrix) Width = radius of the area to be extracted around the keypoint Initialize flag to empty array. Initialize A to empty array. For each i = 1 to extr_volume do Compute row of keypoint, rx from extrema Compute column of keypoint, ry from extrema Compute level of the matrix containing the keypoint from extrema If the keypoint is in the searchable area then Extract a 51x51 patch centered over the keypoint from D [Dx, Dy] = compute gradient of the extracted patch Initialize G to zeros Set j to 1 for each p = 1 to width do for each q = 1 to width do Assign G(j) to Dx(p, q) Assign G(j+1) to Dy(p, q) Increment j by 2 End End Add gradient, G of the extracted patch to array A Else Save the position/index of keypoint in flag array to be deleted End Delete the unsearchable keypoints from extrema using flag array Reassign extr_volume to the new length of extrema [p, q] = size of A Normalize A and Calculate mean vector, M using p, q and A Calculate convariance matrix using M, conv = M’*M V = Calculate Eigenspace from conv Init N to 20 Project = extract matrix containing top N eigenvectors from V

(3) Key Points Orientation Assignment To assign orientation SIFT collects magnitudes and gradient directions around each of the key points. That is why at first it collects magnitudes and gradient directions around each of the key points. Then finds out the most prominent orientation(s) in that region and assigns that orientation(s) to the key point. (4) Feature Representation To obtain the feature vector, at first 4802 element vector is calculated and then image gradient vector is normalized. Then the normalized value is projected into the feature space. After extraction of features we resized the input image into 128 * 128 pixels. The algorithm for key point’s descriptor is shown by Algorithm 2. Figure 6 shows the feature extraction by using PCA based SIFT descriptors.

32

F. T. Johora et al.

Fig. 6 Feature extraction using PCA-SIFT

_____________________________________________________________________ Algorithm 2: Key Point Descriptor _____________________________________________________________________ Notations: Descriptor = single keypoints descriptor vector Feature = array containing all descriptors of all the keypoints Initialize feature to empty array For each i = 1 to extr_volume do Initialize descriptor to empty array; Initialize sigma Initialize width to 51 Compute row of keypoint, x from extrema Compute column of keypoint, y from extrema Compute level of the matrix, z containing the keypoint from extrema If the keypoint is in the searchable area then Init Array sub_x to row pos from x-(width-1)/2 to x+(width-1)/2 Init Array sub_y to column pos from y-(width-1)/2 to y+(width-1)/2 Set matrix sub dimension to (2, length of sub_x * length of sub_y) and init to 0 Init j to 1 Foreach p = 1 to length of sub_x do Foreach q = 1 to length of sub_y do Assign each column of sub to [sub_x(p) - x; sub_y(q) - y] j++; End End

3 Diabetic Retinopathy Detection …

33

Compute distort matrix using the keypoints assigned orientation Distort the sub matrix using distort matrix and assign to sub_dis Add x to first row and y to second row and assign to sub Fix_sub = round all elements of sub matrix to 0 Take a width*width (51x51) patch matrix and init to 0 Set rowval to fix_sub (1,1) and columnval to fix_sub(2,1) Set octaveval to keypoints octave no Set levelval to keypoints matrix level in the octave Elementval = extract the element from D using octaveval, levelval, rowval, columnval Now set patch matrix each element to elementval Take a vector G of dimension (1, 2*width*width) Save all the elements of Dx and Dy in G thus building the gradient vector build the descriptor vector by projecting matrix, project and G, descriptor = project*G’ End Add the descriptor vector to feature array End

D. Classification Framework (1) Classification for Noise Free Dataset Naïve Bayesian (NB) classifier is a simple probabilistic classifier based on probability model, which can be trained very efficiently in a supervised learning. Bayes Theorem finds the probability of an event occurring given the probability of another event that has already occurred which is given by P(w|X ) = P(X |w)P(w)/P(X )

(4)

where P(w) is the prior of y, P(w|X ) is the posterior probability of X , w is class variable and X is a dependent feature vector (of size n) where X = (x 1, x 2 ……, x n ). And after simplification we have got P(w|x1 , x2 , ·, xn )∞P(w)

n 

P(xi |w)

(5)

i=1

In our NB classifier model, we find the probability of given set of inputs for all possible values of the class variable y and pick up the output with maximum probability. This can be expressed as follows: ˆ = argmaxw P(w) W

n  i=1

P(xi |w)

(6)

34

F. T. Johora et al.

where P(w) is the classification probability and P(xi |w) is conditional probability. For the noise free dataset that we have separated in the preprocessing step, NB based classifier is applied. (2) Classification for Noisy Dataset ID3 algorithm works by building decision tree. To build the decision tree this algorithm uses information theory that chooses spitted attributes with highest information gain for a particular dataset. And the attribute carrying maximum information indicates the probability of occurrences that is called entropy. All data set is said to be homologous when there is no uncertainty and entropy is zero. Calculation of entropy is given in Eq. (7). H ( p1 , p2 , . . . . . . ps ) =

S 

( pi log(1/ pi ))

(7)

i=1

where ( p1 , p2 , . . . . . . . . . ps ) are the probabilities for different classes in the dataset. Let consider the data set S where entropy is H(s), and the dataset is divided into new subsets called D, D = {S1, S2………. Ss }. Then again we will calculate the entropies of these dataset. For a particular subset of the given dataset when it is sequential then there is no need of further split (Belongs to the same class) Using Eq. (8) ID3 calculates information gain for any particular split Gain(S, D) = H (S) −

S  i=1

P(Si ) · H (Si )

(8)

3 Diabetic Retinopathy Detection …

35

4 Result Analysis The dataset used for this work have been collected form a competition site [17]. There have been almost 2.4 GB data which have been categorized into two groups named test dataset and training dataset. The main goal of this work is to propose such a Diabetic Retinopathy detection methodology, which would be simple in implementation nature and provides better result compared to previous work. To get the accurate result at first we have applied pre processing on data set then detect the optic disc and then feature extraction and in the last step we have applied different classification algorithm for noisy dataset and noise free dataset. For noisy dataset we have applied probability based weighted decision tree and for noise free dataset Naïve Bayesian algorithm is applied. As here the rate of feature extraction better than the previous methods shown in Table 1. Figure 7 is showing the classification rate of 5 categories data namely 1. Normal, 2. Mild DR, 3. Moderate DR, 4. PDR, 5. Severe DR. It can produce better classification both for the noisy data set and noise free dataset shown in Table 2.

5 Future Work and Conclusion Diabetic Retinopathy has become a leading problem throughout the world and many peoples are losing their vision because of this disease. Keeping this in mind we have proposed a system which will be able to detect diabetic retinopathy from the

36 Table 1 Feature extraction rate using PCA-SIFT

F. T. Johora et al.

Class

Feature extraction rate (%)

Normal

95

Mild DR

85

Moderate DR

80

PDR

89

85

Severe DR Fig. 7 Feature Extraction Rate using PCA-SIFT

Mean (%)

100 Feature Extraction Rate(%)

Feature Extraction Rate (%)

100 90 80 70 60 50 40 30 20 10 0 0

1

2

3

4

Class

Table 2 Classification rate using proposed method

Classification method

Classification rate (%)

Naïve Bayesian (Noise Free Dataset)

99

Probability based weighted decision tree (Noisy Dataset)

97

image of an eye of a patient. Compared to the standard representation, PCA-SIFT is more distinctive and more compact leading to significant improvements in matching accuracy (and speed) for both controlled and real world conditions. Moreover for noisy data it was very difficult to classify, to do so we have assigned weight to the decision node based on the prior probability thus having better classification rate. Although the system is able to provide better results but it has some limitations. To work properly images must be in high resulations. Moreover as we have categorized our data into noisy and noise free, sometimes it may take more times to classify the more noisy data. In future we want to develop an online based service where doctors can get yearly subscription and through which they can easily check for the diabetic retinopathy very easily without having to worry about program or its internal code. They will simply update the images into our online repository and our system will provide them with the details of every single image along with the level of severity. This will be a very convenient in which every single doctor can get access to our system through internet as in today’s world internet is accessible by almost everyone.

3 Diabetic Retinopathy Detection …

37

References 1. Frank R (2004) Diabetic Retinopathy. New Engl J Med 350(1):48–58 2. Ng K, Acharya UR, Rangayyan R, Suri J, Ophthalmological imaging and applications 3. Faust O, Acharya UR, Ng E, Ng K, Suri J (2010) Algorithms for the automated detection of diabetic retinopathy using digital fundus images: a review. J Med Syst 36(1):145–157 4. World Diabetes (1998) A newsletter from the World Health Organization, 4, 1998 5. Ong G, Ripley L, Newsom R, Cooper M, Casswell A (2004) Screening for sight-threatening diabetic reti-nopathy: comparison of fundus photography with automated color contrast threshold test. Am J Ophthalmol 137(3):445–452 6. Orbis, Orbis.org (2016). http://www.orbis.org. Accessed: 01 Apr 2016 7. Nayak J, Bhat P, Acharya UR, Lim C, Kagathi M (2007) Automated Identification of diabetic retinopathy stages using digital fundus images. J Med Syst 32(2), 107–115 (2007) 8. CS231n convolutional neural networks for visual recognition, Cs231n.github.io (2016). http:// cs231n.github.io/convolutional-networks/. Accessed: 01 Apr 2016 9. Rahim SS, Palade V, Shuttleworth J, Jayne C (2016) Automatic screening and classification of diabetic retinopathy and maculopathy using fuzzy image processing. Brain Inform 3(4):249–267. Epub 2016 Mar 16 10. Maher R, Panchal D, Kayte J (2015) Automatic diagnosis microaneurysm using fundus images. Int J Adv Res Comput Sci Softw Eng 5(10) 11. Gandhi M, Dhanasekaran D (2013) Diagnosis of diabetic retinopathy using morphological process and SVM classifier. In: International conference on communication and signal processing 12. Shahin E, Taha T, Al-Nuaimy W, Rabaie S, Zahran O, El-Samie A (2012) Automated detection of diabetic retinopathy in blurred digital fundus images 13. Prentasic P (2013) Detection of diabetic retinopathy in fundus photographs. University of Zagreb, Faculty of Electrical Engineering and Computing Unska 3, 10000 Zagreb, Croatia 6–8 14. Mane* VM, Jadhav DV (2016) Holoentropy enabled-decision tree for automatic classification of diabetic retinopathy using retinal fundus images. Biomed. Eng.-Biomed. Tech. aop 15. Lahmiri S Boukadoum M (2014) Automated detection of circinateexudates in retina digital images using empirical mode decomposition and the entropy and uniformity of intrinsic modefunctions. Biomed Eng-Biomed Tech 59:357–366 16. Lahmiri S, Gargour C, Gabrea M (2014) Automated pathologies detection in retina digital images based on the complex continuous wavelet transform phase angles. Health Technol Lett 1:104–108 17. www.kaggle.com

Chapter 4

GIS-Based Surface Water Changing Analysis in Rajshahi City Corporation Area Using Ensemble Classifier Mahbina Akter Mim and K. M. Shawkat Zamil

1 Introduction Humans, animals, plants are hugely dependent on water. Water is necessary for every aspect of our life. Surface water and groundwater are popular terms of the water body. Surface water can be found in terms of rivers, canals, streams, creeks, lakes, and reservoirs. On the contrary, groundwater also plays an important role in the water system. These types of water are basically stored in the spaces between rock particles and in the layers of the earth surface. Groundwater slowly moves underground because of gravity. For the daily needs of the human being, both the surface water and the groundwater can be used. In average situation, surface water mainly used in public supply, domestic, irrigation, aquaculture, industrial, mining, thermoelectric purpose. The groundwater is also used in those types of work. For excessive use of water, both the amount of surface water and groundwater is decreasing. But the surface water decreasing level turned into a massive alarm nowadays. Because earth pollution increasing day by day. The human waste and industrial waste directly dump in the river, canals, creeks, etc. River water has great importance in the global hydrological cycle and supplies fresh water to humankind. 75% of the earth’s population lives in countries and regions with a present extent of water resources use of more than 20–57% water withdrawal and 70% of global water consumption occur in Asia because the major irrigated lands are located in this region. The water distribution over the world is uneven. Water resources do not coincide with population spread and economic development. The greatest volume of is withdrawn from Southern Asia regions including India, M. A. Mim (B) · K. M. Shawkat Zamil Department of Computer Science and Engineering, Rajshahi University of Engineering & Technology, Rajshahi, Bangladesh e-mail: [email protected] K. M. Shawkat Zamil e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_4

39

40

M. A. Mim and K. M. Shawkat Zamil

Pakistan, Bangladesh, and Southeastern Asia, with the greatest part attributed to irrigated areas in China on the Asian continent [1]. Bangladesh has a huge population. The need for water is increasing with population growth. It is high time to find out a prediction result of the recent data. The study area is the Rajshahi City Corporation (RCC) area. The changes of water body can be measured with the help of Geographic Information System (GIS). The images are collected from the United States Geological Survey (USGS). The images are used to extract feature with the help of ArcGIS. Then, these data formed a dataset. Ensemble classifier, i.e., random forest is used in this research. As the number of tree increases in the algorithm, the accuracy also increases. It is possible to predict the presence of surface water in the future. Here prediction accuracy describes how accurately the algorithm finds out the prediction of the presence of the water.

2 Related Works The study area was covered with 4000 ponds. For the wastage dumping and urbanization, the number of ponds, i.e., surface water decreased day by day. The decreasing rate was about fifty percent. In the year 2000, the number of ponds was 729. At present, the number of the water bodies in Rajshahi city is 214 only [2]. The Padma River covered a large portion of the wate rbody in Rajshahi City Corporation area [3]. Reza et al. [4] describe that the hydrological condition of the Barind track, which is a part of our study area is an acute state of deforestation. This area is faced with low rainfall and lack of surface water. In the study of Ahmeduzzaman et al. [5], it is clear that groundwater fluctuates at the Barind area in Rajshahi. The surface water is linked with the groundwater and can create an impact on the water level. Rajshahi City Corporation is situated in Barind Tract which is in the northern portion of Bangladesh. Most of the soil of this area is yellowish brown and grayish loam to clay loam. There are mainly two seasons: monsoon season and dry season. Rainy season stays from June to September and dry season stays during the rest of the year. The average temperature varies from 4 to 44 °C. In January and February, the temperature falls below 8 °C. March is the hottest month, humidity rises from 60 to 85% and rainfall area varies from 1500 to 2000 mm [5]. In the Barind area the, amount of groundwater extraction causes by irrigation and it is rising day by day. As the population is increasing, the demand for crops is increasing. Irrigation is an important parameter to increase the production of crops. Rainfall does not occur equally in all the moth of the year in Bangladesh. In the Barind tract, rainfall is very little. Most of the soil of this area is hard clay and clay type which is less permeable to infiltration. So a little amount of rain water infiltrated the ground. Groundwater condition depends on geology, hydrologic parameter, soil properties, recharge and discharge and hydraulic characteristics of the aquifer. In Gopalpur mouza, the water level varies from 7.17 to 61.59 ft. where maximum level occurs 61.59 ft. in the year of 2001 (March) and minimum level occurs 7.17 ft in

4 GIS-Based Surface Water Changing Analysis …

41

1995 (August) and 1996 (August) [5]. In Sontospur mouza of Paba upazilla, available data for period 1995–2009 is analyzed. The water level varies from 5.25 to 38.00 ft. Maximum level occurs 38.00 ft. in the year 2007 (April) and minimum level occurs 5.25 ft in the year 1996 (October). Most of the groundwater abstractions take place in the dry months starting from January to May, also June in some dry season [5]. In the study of George et al. [6], it was found that the water quality of the study area reduced due to increasing population, aggregation, fast urbanization, industrialization, and pollution. The groundwater and surface water around the study area was affected due to effluents from the industry. The water samples were collected from different portions of the study area. They were analyzed in the laboratory, and affected areas were marked using GIS because GIS can capture data with lots of integrations.

3 Methodology The study area is Rajshahi City Corporation (RCC). The images of this area are collected from USGS EarthExplorer. The multispectral images generally consist of three to ten bands. By selecting proper bands and with the help of remote sensing radiometer, the area what needed is located. In ArcGIS, the images are classified to a maximum likelihood classification technique for extracting features. Then Ensemble classification technique used. Here, the Random forest algorithm used for classification and calculate the prediction accuracy.

3.1 Image Collection Images of the study area lie between the years 1987 and 2016. These images are collected from the United States Geological Survey. The shape file of the study area which is Rajshahi City Corporation (RCC), is collected from Rajshahi Development Authority (RDA) (Fig. 1).

3.2 Image Classification To extract features from the downloaded images, ArcGIS tool is used. There are some techniques for extracting features. The maximum likelihood classification is chosen. By generating signature file, the images are classified in that approach. This technique is a very popular method for Geographic Information System and remote sensing. Maximum Likelihood classification has different benefits [7] such as it can generate a variety of estimation situations and it is also used in mathematics.

M. A. Mim and K. M. Shawkat Zamil º

º

º

º

º

º

º

º

Rajshahi City Corporation (RCC) Area

º 20’30"N

º

º21’30"N

º

º22’30"N

º

º 23’30"N

º

º24’30"N

º

º

N

42

Legend

Scale 0 0.5 1

RCC Landuse º

º

º

º

º

2 º

3

4 º

Kilometers º

Fig. 1 The study area of Rajshahi City Corporation

GIS is considered for multicriteria analysis in resource evaluation. Remote sensing and GIS is proved to potential tools for integrating of numerous thematic maps based on topography, geomorphology, land use/land cover, slope and producing a conceptual model through Spatial Analyst in ArcGIS [8]. From different supervised classification techniques, maximum likelihood classification is selected. In supervised learning, the instances are given with known labels [9]. Maximum likelihood classification is done in ArcGIS. Four basic steps demonstrate the procedure [10]. In the beginning, the image analysis toolbar from ArcMap is enabled. After that, polygons are drawn to identify the areas. Then, the signature file is created. ArcGIS provides different classification techniques. Classification can be done using maximum likelihood classification, iso cluster, class probability, or principal component technique. Maximum likelihood classification is used from the above techniques. Different areas are detected by changing the color bands to different numbers. For many advantages, this technique has been used to get the values of water bodies, vegetation, urban area, and open space. In Fig. 2, 12 different colors denote 12 different classified areas. These areas then reclassified to get the desired classified images with data.

º

º

º

º

43 º

º

º

º

After Maximum Likelihood classification of RCC in 1998

º20’30"N

º

Legend 1 2 3 4 5 6 7 8 9 10 11 12

º

º21’30"N

º22’30"N

º

º23’30"N

º

º24’30"N

º

º

N

4 GIS-Based Surface Water Changing Analysis …

Scale 0 0.5 1

2

3

4

RCC Landuse

º

º

º

º

º

º

º

Kilometers º

Fig. 2 Study area in 1998 after using maximum likelihood classification

3.3 Feature Extraction For extracting the features, the classified images are gathered. Then images again reclassified for clearly finding the four features which are called water body area, vegetation area, urban area, and open space area. From these values, the percentage of water bodies calculated. All the data are counted in square kilometer unit then converted into acres (Fig. 3).

3.4 Dataset Preparation After reclassifying all the images from the year 1987 to 2016, a dataset is generated. This dataset has 25 numeric values. The dataset contains 25 rows and 6 columns, which represents the water body area, vegetation area, urban area, open space area, the percentage of the water body and the class value.

M. A. Mim and K. M. Shawkat Zamil º

º

º

º

º

º

º

º

Rajshahi City Corporation in 1998

Legend

º

º

º

Waterbody Vegetation Urban Area Open Space

º20’30"N

º21’30"N

º

º22’30"N

º

º23’30"N

º

º24’30"N

º

º

N

44

Scale 0 0.5 1 º

º

º

2 º

3

4

Kilometers

º

º

Fig. 3 Study area in 1998 after reclassification

3.5 Classifier For classifying this type of dataset, random forest technique has been used. • Random forest Random forest algorithm is one of the ensemble classification algorithms of data mining. It is a supervised classification algorithm. This algorithm is composed of a collection of independent decision trees. In the random forest algorithm, it selects particular features from the total feature of the feature space. This value should less than total feature value. From the features, it calculates the best splitting point. Then the child nodes are found using this best splitting approach. These processes are repeated to create “n” random trees which form a random forest. The feature selection depends on entropy measurement. For this, the information gain approach is used. H (x) = E X [I (x)] = −



p(x) log p(x)

(1)

x∈X

Here, p(x) is the attribute value which is known as feature values and H (x) is the calculated entropy.

4 GIS-Based Surface Water Changing Analysis …

45

One of the main concerns of using this algorithm is it takes decision better than the decision tree. It can also estimate missing values [9]. It can handle missing values in training data as well as testing data.

4 Result and Discussion We tested the dataset in random forest algorithm. The experiment uses fivefold crossvalidation, because dataset consists of only 25 values. To calculate the accuracy, split-ration is selected as 0.6 that means 60% of the data is used for training and 40% of the data is preserved for testing. The final output of the algorithm is given in Table 1. We can see that the accuracy of the algorithm is increasing with increasing the number of the tree. The final accuracy measured 92% in our dataset. The random forest algorithm can detect and predict the changes of water, 92% accurately. We can show this in graphical representation also (Fig. 4). We can also measure the precision, recall, and F1-score to justify the accuracy, the confusion matrix for three class is given in Table 2. We have three classes in the dataset. That’s why we get 3 × 3 confusion matrix. To calculate precession, the true positive value is divided by true positive value added with the false positive value of that corresponding class. To calculate the recall, the true positive value is divided by true positive value added with the false negative

Table 1 Accuracy measurement using random forest algorithm Algorithm

Accuracy (Tree 1)

Accuracy (Tree 5)

Accuracy (Tree 10)

Random forest

76%

80%

92%

100 80 Accuracy (Tree 1) 60 Accuracy (Tree 5)

40 20

Accuracy (Tree 10)

0 Random Forest Algorithm Fig. 4 Graphical representation of accuracy in random forest algorithm

46 Table 2 Accuracy measurement using random forest algorithm

Table 3 Precision, recall, and F1-score measurement

M. A. Mim and K. M. Shawkat Zamil

Algorithm

Confusion matrix

Random forest

TPA (5)

EAB (0)

EAC (0)

EBA (1)

TPB (16)

EBC (0)

ECA (1)

ECB (0)

TPC (2)

Algorithm

Precision

Recall

F1-score

Accuracy

Random forest

0.943

0.920

0.922

92%

value. From the result of precision and recall, we can also measure the F1-score. The calculated values are given in Table 3.

5 Conclusion In this research, we analyze the surface water changes by using a different approach. First, the images are collected from the USGS Earth Explorer and the shapefile collected from the Rajshahi Development Authority (RDA). Then we extract the features from those images using ArcGIS tool. From different technique, we choose the maximum likelihood classification for the feature extraction process. Because this method used is in developing a variety of estimation situation. Choosing the correct feature extraction technique was a challenging task. After reclassifying images, a dataset generated. Then we used the ensemble classification technique to analyses the surface water changes. Here the prediction accuracy measured 92%. We also calculate the precision, recall, and F1-score to justify the accuracy. This research work can be extended by enhancing dataset. Different classification methods can be used for further work. Acknowledgements The authors would like to give special thanks to their honorable Prof. Dr. Md. Shahid Uz Zaman, Department of Computer Science and Engineering, Rajshahi University of Engineering and Technology, Bangladesh for his valuable suggestion, encouragement, guiding, and his constant support.

References 1. Shiklomanov IA (2000) Appraisal and assessment of world water resources. Water Int 25(1):11–32 2. Rahman MdH (2014) Pond filling plagues Rajshahi city. DhakaTribune, 2 Sept 2014, https:// www.archive.dhakatribune.com/environment/2014/sep/02/pond-filling-plagues-rajshahi-city 3. Rajshahi City Corporation, BANGLAPEDIA, 9 Mar 2015. https://www.en.banglapedia.org/ index.php?title = Rajshahi_City_Corporation

4 GIS-Based Surface Water Changing Analysis …

47

4. Reza AHMS, Mazumder QH (2005) Evaluation of hydrogeological conditions of sapahar and porsha upazillas, barind tract, Bangladesh. J Life Earth Sci 1(1):15–20 5. Ahmeduzzaman M, Kar S, Asad A (2012) A study on ground water fluctuation at barind area, Rajshahi. Int J Eng Res Appl (IJERA). ISSN: 2248–9622 6. George GK et al, Study of ground water pollution around an industry using GIS 7. Natrella M (2010) NIST/SEMATECH e-handbook of statistical methods 8. Sajjad H, Iqbal M, Bhat FA (2014) Integrating geospatial and geophysical information for deciphering groundwater potential zones in Dudhganga catchment, Kashmir Valley, India. Am J Water Resour 2(1):18–24 9. Qi Y, Klein-Seetharaman J, Bar-Joseph Z (2005) Random forest similarity for protein-protein interaction prediction from multiple sources. Biocomputing 2005:531–542 10. Kotsiantis SB, Zaharakis I, Pintelas P (2007) Supervised machine learning: a review of classification techniques 3–24 11. Melgani F, Bruzzone L (2004) Classification of hyperspectral remote sensing images with support vector machines. IEEE Trans Geosci Remote Sens 42(8):1778–1790 12. Landsat Program (2017) Wikipedia, wikimedia function, 29 Nov 2017. https://www.en. wikipidea.org/wiki/Landsat_program 13. Breiman L (2001) Random forests. Mach Learn 45(1):5–32 14. Perlman (2016) USGS Howard. “Groundwater depletion.” Groundwater depletion, USGS water science, 9 Dec 2016. https://www.water.usgs.gov/edu/gwdepletion.html

Chapter 5

Leveraging Machine Learning Approach to Setup Software-Defined Network(SDN) Controller Rules During DDoS Attack Sajib Sen, Kishor Datta Gupta and Md. Manjurul Ahsan

1 Introduction The enormous increase in the number of network devices to the Internet has provided many useful solutions in different fields like information technology, medical science, etc. This large amount of increase in connectivity to the Internet has challenged the architecture of the traditional network. To face those challenges, a new network architecture, named as Software-Defined Network(SDN), had been implemented which separates the traditional data plane and control plane [1, 2] so that the network controller can be programmable directly. The main advantage of SDN architecture is that it provides better network management capability. Although SDN architecture has many advantages, it also has some vulnerable issues same as traditional network attacks. Although network admin may identify the potential attack in real time, it would be impossible to identify simultaneous attack efficiently. That is why similar to firewall rules, some rules for network security also need to implement on the SDN controller. However, the challenging task is to create controller rules which allow access to the benign user while restricting the malicious user or attackers. Abdou et al. [3] in their work presented the specific features to differentiate between benign and malicious users. One such common characteristic can be found in the attackers is coordinated attacks. As this problem is a classification problem, machine learning (ML) approach has significant potential in this regard [4]. In this paper, several S. Sen (B) · K. D. Gupta Department of Computer Science, University of Memphis, Memphis, TN 38152, USA e-mail: [email protected] K. D. Gupta e-mail: [email protected] Md. Manjurul Ahsan Department of Industrial Engineering, Lamar University, Beaumont, TX 77710, USA e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_5

49

50

S. Sen et al.

machine learning algorithms have been adopted to predict and classify malicious user or Denial of Service(DDoS) attack from a user, using private network data in SDNbased environment. The purpose of using ML in SDN environment is to leverage the ML approach so that security rules on SDN controller can be defined to block the potential attackers as well as their entire subnet. To accomplish the goal Mininet (an open-source emulation software) had been used for creating and simulating an SDN-based testbed and WEKA (an open-source machine learning software) had been used to build, test, and compare different machine learning models.

2 Literature Review Nowadays, machine learning algorithms are used in lots of unconventional fields such as Genetic algorithms are used in image processing and cloud computing [5]. To detect Distributed Denial of Service (DDoS) attack in SDN environment-based network Ashraf and Latif [2] adopted several machine learning (ML) based techniques such as Neural Networks, Genetic Algorithms, Bayesian Network, and Support Vector Machine. Their paper discussed how effective the algorithms are for anomaly detection as well as showed a clear distinction among all of these approaches. Ali et al. [6] in their paper provided a detailed survey about how to secure a network based on SDN. Their paper suggested that SDN network can be used as a service to handle network threats. Abdou et al. [3] in their paper has given a detailed analysis of SSH brute-force attacks using the Longtail project [7] dataset. Another paperwork for anomaly detection using k-Nearest Neighbors (kNN), Support Vector Machines, Bayesian Networks, Expectation Maximization (EM) with several attack scenarios in SDN environment presented by Sommer [8]. Kim and Feamster [9] in their work proposed different techniques of improving network management system on SDN. Among all of the techniques, three issues have been given importance: how to handle frequent changes in network conditions, how to support high-level language to configure a network, and how to provide a better interface for troubleshooting. Another unique approach adopted by Keller et al. [10] presented a system which can be used on the top of SDN layer to leverage the control over network switches via programmable code as well as can provide access to a user to set their own rules. In respect of security aspects of SDN using machine learning, Eskca et al. [1] in their paper presented a detailed discussion. In contrast to the above work, this paper uses an SDN-based testbed as a network emulator. Also, machine learning techniques have been applied to predict malicious users or unauthorized users early so that SDN controller can take the output of the classifier to set rules automatically.

5 Leveraging Machine Learning Approach to Setup Software …

51

3 Background 3.1 Machine Learning Techniques Machine learning techniques are categorized into supervised, unsupervised and semisupervised learning [11]. Supervised learning algorithms use labeled input as training data and predict the unknown cases with the trained model. Among all other techniques in supervised learning support vector machine(SVM) has been used by researchers widely for Network Intrusion Detection System (NIDS) [12] Unsupervised learning uses unlabeled data as training data and tries to predict the data cluster or structure in the dataset. When a new data has been given to the trained model, the model puts it in one of the clusters. Different clustering algorithms such as K-means and Expectation Maximization has been used widely for NIDS and anomaly detection [13, 14]. In semi-supervised algorithms, classification models are trained by supervised learning methods, but the given training data are both labeled and unlabeled. Among many techniques, Spectral Graph Transducer and Gaussian Fields techniques have been used widely for anomaly detection [15, 16]. Nowadays deep learning algorithms have gained much more popularity for intrusion detection, which can be applied both in supervised and unsupervised way.

3.2 Network Intrusion Detection System(NIDS) The primary purpose of an Intrusion Detection System (IDS) is to restrict malicious users and to track network stats. As shown in the Fig. 1 IDS is classified as host-based intrusion detection system(HIDS) and network intrusion detection system(NIDS). Moreover, IDS techniques are also classified anomaly-based and signature-based [17]. Between these two techniques anomaly-based detection is efficient for unknown threats and among the methods its uses, machine learning is one of them.

3.3 Software-Defined Network (SDN) Software-Defined Network has three layers: the data plane or infrastructure layer, controller plane, and application plane (see in Fig. 2). Controller plane works as an operating system in an SDN network and controls the network flows. To control the controller externally, controller plane uses the REST API. OpenFlow is used to communicate with switches and sends commands to them according to flow rules. Moreover, these flow rules can be sent to the controller with a program or can be set manually. The application layer is the highest level of SDN. This layer consists of end-user commercial applications like network monitoring. In short, when we send commands to controller what we can see is the application layer. Controller plane

52

S. Sen et al.

Fig. 1 Overview of intrusion detection system

translates the data from the application layer to the data plane layer. Controller plane can be termed as a man in the middle which translate data from the application layer to the data plane layer. The data plane layer is the layer where physical devices like switching and routing devices stay. This layer accepts the data from the application layer and accepts the new flow rules changing about how the network devices will act. To forward a packet, switches maintain a flow table as shown in Fig. 3 which consists of flow entries. From a flow record, a set of packets with common features can be summarized. When first data packet from a host comes to a switch, it executes a lookup operation, by using packets header features, in its flow table. If there is no matching in the flow table, then the packet flow is being transmitted to the controller for a decision through the OpenFlow protocol. After that, the flow is returned to the data path. Then the actions derived into a drop or forward packets to another port/controller (Fig. 4).

4 Tools Used to Implement the System A brief description of the tools used in this paper to implement the system is given below.

5 Leveraging Machine Learning Approach to Setup Software …

53

Fig. 2 Software-Defined Network (SDN) architecture indicating control plane, data plane and application plane [12]

4.1 OpenVSwitch This switch stays below the OpenFlow interface, which is an open-source virtual switch enabling programmable automation with supporting protocols and interfaces [18]. Moreover, this switch allows several network control systems such as sFlow, NetFlow to respond and adapt to environmental changes and supports migrating network states and configuration between instances. Port(s1-ethX) on the switch works as a virtual bridge to set up connections between different hosts and the switch, where s1 is the switch and s1-ethX is the name of the interfaces (see in Fig. 5).

4.2 Mininet It is a network emulator capable of creating virtual network environment such as virtual switches, OpenFlow controller, hosts, which can communicate through virtual links [19]. Mininet also supports OpenFlow protocol, which is a big advantage to set up the SDN environment. Another advantage of using mininet is that emulated

54

S. Sen et al.

Fig. 3 A pictorial representation of traffic flow graph in Software-Defined Network (SDN) architecture [12]

Fig. 4 A pictorial overview of packet flow in SDN environment

Fig. 5 An architectural overview of OpenVSwitch

5 Leveraging Machine Learning Approach to Setup Software …

55

Fig. 6 A pictorial overview of traffic monitoring mechanism using mininet and sFlow analyzer

Fig. 7 A pictorial overview of event handling mechanism using mininet and sFlow analyzer

network in mininet can run real Linux applications. It can also provide networking stack as well as Linux kernel for development. Because of this advantage, code which runs on mininet emulated network, can directly be transferred to real hardware switches with minimal changes.

4.3 sFlow-RT sFlow-RT [20], which is open-source software, has OpenFlow controller embedded in it. It allows flow insertion to OpenFlow switches and monitors flows to the switches. Besides, it can handle events of interest, apply traffic rules to a particular controller and can also raise triggers (Figs. 6 and 7).

5 Machine Learning Approach Used in DDoS Detection System To set up a test bed for SDN environment, mininet is installed on a VirtualBox on Ubuntu Linux with the real-time network analyzer sFlow-RT. To work simultaneously with existing network, sFlow-RT need to be configured before to monitor traffic changes. This process has been done by installing sampling agents and attaching them to those switches of interest. After that, a logical bridge port(sFlow) has

56

S. Sen et al.

Fig. 8 Packet flow during normal condition

Fig. 9 Packet flow during DDos condition

been created, which sends sFlow traffic from a particular switch (i.e., s1) to sFlow collector (127.0.0.1), sampled by sFlow agent/s (i.e., eth0). At first, a network topology has been created by the mininet emulator where the controller is remote. An example of creating virtual SDN topology is given below. The network can be created by using the following command in mininet: sudo mn – controller = remote,ip = 127.0.0.1,port = 6653 –topo = single,3. The command to connect OpenVSwitch to sFlow-RT analyzer is: sudo ovs-vsctl – – id = @sflow .. ¨ create sflow agent = eth0 target = 127.0.0.1 : 6343 s ampling = 10 polling = 20 – – set bridge s1 sflow = @sflow. Here the OpenVSwitch has been connected to sFlowRT with eth0 as agent and ow switch as s1. After the setup, sFlow analyzer needs to be started manually to monitor the collected data samples by./sFlow-rt/start.sh command. To check the topology of mininet network in sFlow-RT the given address http://localhost:8080/ui/pages/index.html need to visit. Moreover, to monitor realtime network flow Wireshark (a graphical utility for viewing packets) has been used to collect the network metrics data. This data has been stored manually as a dataset for further machine learning classification (Figs. 8 and 9). In DDoS attack, attacker uses a command to send traffic from a huge number of compromised host to a designated target to flood the target infrastructure so that the target host deny access to its legitimate users. In this paperwork, an ICMP DDoS flood

5 Leveraging Machine Learning Approach to Setup Software … Table 1 Dataset features

57

No.

Features

Description

1

Service

Network service on the destination, e.g., http, telnet, etc.

2

Header length

Length of header data

3

Flags

Normal (0) or error (1) status of the connection

4

TTL

Time to live

5

Protocol

Type of the protocol, e.g., tcp, udp, etc.

6

Data bytes

Bytes of data needed for certain protocol

7

Epoch time

Time to complete one epoch

8

Reply response time

Time to give response

9

Land

1 if connection is from/to the same host/port; 0 otherwise

attack had been initialized using ping -f 10.0.0.1 command from different MiniNet hosts. Besides to create flood attack manually, a payload had been created by python code using scapy library. During the DDoS flood attack Wireshark network data visualizer also used to capture the network data, and from these data, a dataset for the DDoS attack has been created. Furthermore, WEKA machine learning software has been used to build the model and to test the accuracy of the model. The dataset used to build the model includes data for normal and Dos attack for six types of protocols such as TCP, UDP, ICMP, ARP, IPv4, and SSH. A detailed explanation of the data set is given in Table 1. As the dataset instances are not very large, to reduce bias, 15% noisy data has been added to the dataset. A 20-Fold cross-validation measures have been taken to build the model. We used AdaBoosting machine learning algorithm taken decision stump as a weak classifier to build the classifier model for the network. After model had been built, this model was concatenated with the SDN controller where the controller only flows those traffic which is filtered first by the classifier model. This concatenation ensures two-step network traffic filtering. The first step is performed by SDN environment itself by providing user-defined network con figuration to filter out ordinary attack traffic. In the second step, network traffic passed by SDN environment filters out again through machine learning classifier which only passes those traffic which is labeled as benign for the network. So in this approach, we established SDN controller rules where trained model parameters are integrated with general SDN environment controller rules.

58

S. Sen et al.

6 Analysis To analyse the system performance, we need to test the performance of the machine learning model, as the machine learning model integrated with the system entirely determined the efficacy of the system. However, to calculate the model’s performance there are some statistical methods already established and largely used for machine learning classifiers named as accuracy, precision, recall, and F-measures. These statistical performance metrics had been used to measure the detection accuracy of our model. Moreover, to compare the efficacy of our model we also build several machine learning model using WEKA software with the same dataset. A comparison study of the performance metrics for different machine learning classifier given in Table 2. Figure 10 also shows the comparison graph of F-measure of different machine learning techniques for the given dataset and system.

Table 2 Performance metrics for different machine learning techniques No

Techniques

Precision Recall

Fmeasure

1 2

ROC area

Bayes net

0.889

0.885

0.885

0.863

Nave Bayes

0.731

0.705

0.693

0.707

3

Multilayer perceptron

0.836

0.836

0.836

0.834

4

Support vector machine (Kernel = 3)

0.853

0.852

0.852

0.853

5

AdaBoost (Decision stump as weak classifier)

0.934

0.934

0.934

0.887

6

J48 decision tree

0.903

0.902

0.901

0.880

7

Random forest

0.837

0.836

0.836

0.899

Fig. 10 Comparison graph of F-measure found for different machine learning techniques including AdaBoost

5 Leveraging Machine Learning Approach to Setup Software …

59

7 Conclusion and Future Work In this paper, a DDoS attack detection system has been presented leveraging machine learning classifier output in SDN controller rules. The main aim of this paper was to develop an SDN environment-based virtual network testbed to implement DDoS attack and to detect that attack leveraging machine learning model. Moreover, we also have prepared an SDN-based network traffic dataset to build and test our machine learning model. In this paper, we have found 93% DDoS attack detection accuracy with the dataset using Adaboosting with decision stump as a weak classifier. The clear advantage of identifying each attack type with machine learning model is that SDN controller only blocks those specific attack types by automatically defining rules in its controller taking advantage of classifier output which add an extra security over the firewall rules. Moreover, this technique reduces false-positive rate mostly with high accuracy rate. In the future, we have the plan to extend the approach to detect other network attacks besides DDoS such as spoofing, U2R, R2L, etc. Moreover, to handle with more network data features for large network deep learning method will be used for feature extraction in addition to the strong classifier.

References 1. Eskca EB, Abuzaghleh O, Bondugula S, Joshi P, Nakayama T, Sultana A (2015) Software defined networks security: an analysis of issues and solutions. Int J Sci Eng Res 6:1270–1275 2. Ashraf J, Latif S ( 2014) Handling intrusion and DDoS attacks in software defined networks using machine learning techniques. In: Software engineering conference (NSEC), pp 55–60 3. Abdou A, Barrera D, van Oorschot PC (2016) What lies beneath? analyzing automated SSH bruteforce attacks. In: International conference on PASSWORDS 2015: technology and practice of passwords, pp 72–91 4. Qazi ZA, Jin T, Lee J, Bellala G, Arndt M, Noubir G (2013) Application awareness in SDN. ACM SIGCOMM Comput Commun Rev 43:487–488 5. Gupta KD, Sen S (2018) A genetic algorithm approach to regenerate image from a reduce scaled image using bit data count. BRAIN. Broad Res Artif Intell Neurosci 9:34–44 6. Ali ST, Sivaraman V, Radford A, Jha S (2015) A survey of securing networks using software defined networking. IEEE Trans Reliab 64:1086–1097 7. LongTail (2018) Longtail log analysis. Accessed 24 Nov 2018. Retrieved from http://longtail. it.marist.edu/honey/ 8. Sommer V (2014) Anamoly detection in SDN control plane. Masters thesis, Technical university of Munich, Munich, Germany 9. Kim H, Feamster N (2013) Improving network management with software defined networking. IEEE Commun Mag 51:114–119 10. Keller E, Drutskoy D, Rexford J (2013) Scalable network virtualization in software-defined networks. IEEE Internet Comput 17:20–27 11. Atkinson RC, Bellekens XJ, Hodo E, Hamilton A, Tachtatzis C (2017) Shallow and deep networks intrusion detection system: a taxonomy and survey, pp 1–43 12. Niyaz Q, Sun W, Javaid AY, Alam M (2016) A deep learning approach for network intrusion detection system. In: International conference wireless networks and mobile communications (WINCOM), pp 258–263

60

S. Sen et al.

13. Syarif I, Prugel-Bennett A, Wills G (2012) Unsupervised clustering approach for network anomaly detection. In: International conference on networked digital technologies NDT 2012: networked digital technologies, pp 135–145 14. Bennett KP, Demiriz A (1999) Semi-supervised support vector machines. In: Neural information processing systems, pp 1–7 15. Chen C, Gong Y, Tian Y (2008) Semi-supervised learning methods for network intrusion detection. IEEE Int Conf Syst Man Cybern 2603–2608 16. Hinton G, LeCun Y, Bengio Y (2015) Deep learning review. Weekly journal of science in nature international. Nature 521:436–444 17. Reza MBI, Aburomman AA (2016) Survey of learning methods in intrusion detection systems. In: International conference on advances in electrical, electronic and systems engineering (ICAEES), pp 362–365 18. OVS-ofctl (2018) Open flow switch management commands. OpenVSwitch.org, Last Retrieved on 24 Nov 2018. http://OpenVSwitch.org/support/dist-docs/ovs-fctl.8.txt 19. Mininet Team (2018) Mininet overview. Last Retrieved on 24 Nov 2018. Retrieved from http:// mininet.org/overview/ 20. Phaal P, Lavine M (2018) sflow version 5. sFlow.Org, Last Retrieved on 24 Nov 2018. Retrieved from https://sflow.org/sflow_version_5:txt

Chapter 6

A Fuzzy-Based Study for Biomedical Imaging Applications Fahmida Ahmed, Tausif Uddin Ahmed Chowdhury and Md. Hasan Furhad

1 Introduction Magnetic resonance imaging (MRI) is cutting-edge technology in biomedical field and an application of big data that has been extensively used to obtain detailed images of human brain and body and provides the opportunity to investigate the difference between normal and abnormal tissues in different brain disorders [1, 2]. MRI imaging technology is more progressive compared to Computed Tomography (CT) and X-rays for detecting brain abnormalities and various early-stage brain diseases. These technologies are particularly useful in tracking early signs and symptoms of gray matter diseases such as Alphers disease, Alzheimer diseases, as well as, white matter diseases such as multiple sclerosis and leukodystrophy [3]. For investigating early diagnosis through computer-aided techniques, image segmentation plays a substantial role in biomedical image processing research. Image segmentation technique works out by dividing image different regions which are homogenous with respect to each other based on some image features called intensity and texture [4]. Image segmentation can be categorized based on clustering, thresholding, and region extraction. Thresholding segments an image by converting it into a binary pattern considering the threshold values that have been inputted to. Otsu’s method is one the popular clustering-based segmentation approach and it is considered here in one part of the work. Also, some of the other methods for image segmentation are based on image compression, histogram analysis, dual clustering [5]. The authors in [18] F. Ahmed (B) University of Chittagong, Chittagong, Bangladesh e-mail: [email protected] T. U. A. Chowdhury Data soft Systems Bangladesh, Dhaka, Bangladesh Md. H. Furhad University of New South Wales, UNSW, Sydney, Australia © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_6

61

62

F. Ahmed et al.

proposed a border detection method which is comprised of filtering, detection, and tracing. Fuzzy clustering (FCM) is an unsupervised classification technique, which is extensively used in image segmentation applications. It works by dividing the data into several regions or clusters based on the membership value by Euclidean distance calculation between the image pixels [6]. FCM calculates the Euclidean distance between pixels by considering the same equality of each feature. However, it is observed that the image features are not given priority in most of the real-world applications and it affects the clustering performance. Also, the presence of noise and low contrast degrades the segmentation accuracy. Hence, to address this limitation several techniques have been reported, such as FCM incorporated with smoothing technique [19], modification of objective function or membership function [20], combining hard c-means (HCM) and FCM into one method [21] for image segmentation. Suppressed fuzzy c-means (SFCM) takes care of the FCM weakness by enhancing its convergence speed and clustering performance [7]. However, the main limitation of these algorithms is the selection of suppression factor α, because its inappropriate selection produces some flat portions of the histogram curves resulting in inaccurate image segmentation and sometimes cause divergence from the classification. Thus, we propose a modified suppression factor based SFCM algorithm to achieve a better image segmentation. We have organized the remaining paper as follows. The proposed framework is described in Sect. 2. The computational setup with results is presented in Sect. 3. We have concluded our discussion in Sect. 4.

2 Proposed Framework Thresholding and region-based segmentation have been extensively studied in computer vision and pattern-based investigations to improve the compactness of the regions due to its ease and clustering validity [8]. However, it is observed that these techniques are not efficient in the accurate segmentation while the complex image processing tasks take place. Hence, combined methods offer a better and flexible solution to achieve better and accurate segmentation results in this context. Considering this, we propose a combined framework, which is described in this section. Figure 1 shows the proposed framework. First, we implement Vector Median Filtering (VMF) to preprocess the images. This is a vector processing operator, which is an extended version of the scalar median filter and used to denoise the image if there are any blur exists and removes the boundary shifting. The details of this can be found in [9]. Usually, the noise in MRI images are randomly distributed over the whole image and maintains an uncorrelated relation with the image pixels. Also, uncertainty is extensively existed because of the impulsive noise which is originated from the low resolution of sensors [10]. Due to this reason, the noise in brain MRI images degrades the result of image

6 A Fuzzy-Based Study for Biomedical Imaging Applications

63

Fig. 1 Proposed framework

segmentation performance. Hence, we exploit the vector median filter (VMF) in this paper during the preprocessing stage. Next, the Otsu thresholding technique is used to exploit the global information of the image that provides a better estimation of the image prior to implementing the proposed clustering algorithm. This algorithm is based on discriminant analysis that divides an image into background and objects classes and calculates the optimum threshold by separating them into variance values. For instance, the intra-class variance is achieved through the minimal value and the inter-class is achieved through the maximal value. The details of this calculation can be found in [11]. However, for the ease of reader the otsu thresholding is summarized here: First, two regions in the image are divided by intensity threshold calculation. This is determined by inter-class variance through minimization or maximization. The  L method considers the pi i = 0 − 1 as probabilities of image histogram at gray level, and L is the range. The calculation of the probabilities for the background (PB ) and object (PO ) with a threshold can be obtained as follows: PB (t) =

t i=0

PO (t) = 1 − PB =

Pi

L−1 i=t+1

(1) Pi

(2)

The image background and object mean is calculated by considering the subsequent equations. μB (t) =

t  i*Pi P (t) i=0 B

L−1  i*Pi μO (t) = P (t) i=t+1 O

(3)

(4)

64

F. Ahmed et al.

After obtaining the values on mean values from, the variance is computed by the following: σB (t) = σO (t) =

t  Pi (i − μB )2 ∗ PB (t) i=0 t 

(i − μO )2 ∗

i=0

Pi PO (t)

(5) (6)

The inter-class variance is considered as the weighted variance of the cluster and obtained as follows: σinter−class (t) = PB (μB (t) − μ)2 + PO (μO (t) − μ)2

(7)

where μ is image global mean. The intra-class variance can be obtained as follows: σintra−class (t) = PB (t) ∗ σB (t) + PO (t) ∗ σO (t)

(8)

The optimal threshold value, topt is obtained by maximizing or minimizing the class variances as follows: topt = arg max1≤i≤L−1 (σintra−class (t)) = arg min1≤i≤L−1 (σinter−class (t))

(9)

Lastly, we implement the proposed clustering algorithm to segment the image. The standard FCM sometimes provides lower performance in segmentation accuracy. Modified version of SFCM is proposed by improving its suppression factor here. The performance of SFCM depends on the selection of suppressed factor α. The factor α determines the rigidity of the clusters through the degree of inter-class separation strength. This works as, the more distance between the cluster centers, the more rigid the cluster is. Thus, a proper selection of α denotes a good clustering performance. The algorithm works as follows. Let us consider that the clustering algorithm operates on the image data set X = {X1 , X2 … Xn }, where n is the number of data points or image pixels. The clustering algorithm is initialized by setting the cluster numbers c, fuzzy parameter m, initial cluster center vi , and terminating threshold ε. The parameters are initialized to minimize the following cost function which is the goal of the clustering algorithm. In the following function, uik represents the membership degree of the data x k that belongs to the kth cluster. C FC M (U, V ) =

c  n  i=1 k=1

m 2 u ik d (xk , vi )

(10)

6 A Fuzzy-Based Study for Biomedical Imaging Applications

65

The clustering algorithm produces the ideal number of clusters cost function minimization C FC M (U, V ) and by updating the values of uik and vi considering the subsequent equations: ⎡ ⎤−1 c   2   1/(m−1) ⎦ d (xk , vi )/d 2 xk , v j u ik = ⎣ j=1

n m k=1 u ik ∗ x k vi =  n m k=1 u ik

(11) (12)

As the cost function is retrospectively minimized, the cluster centroid vi gains more stability. At this stage, we propose a new suppression factor to update the partition as follows:

2 −1 α = min vi − v j /m

(13)

Here vi denotes the ith cluster center, vj represents the jth cluster center, and m is the fuzzification degree. It is essential that proper selection of α indicates the efficiency of the algorithm such as higher values of α indicates the superiority of FCM over HCM and lower values of α indicates the opposite. It is obvious that better clustering performance depicts good clusters and the greater distance between the cluster centers reflects the better segmentation results. Another important aspect of this proposed algorithm is that both m and α has an impact on the learning rate of the algorithms, which improves the segmentation performance. The proposed algorithm automatically updates the value of α, which is considered as an optimal method to produce a better clustering performance for the data set provided. Next, the partition and the cluster centroid are updated by the subsequent equations and considering the new suppression factor α. 

1 − α + αu ik , i = p αu pk , i = p n m k=1 u pk ∗ x k vi = n m k=1 u pk

u pk =

(14) (15)

where u pk denotes the data point, xk fits the cluster with the maximum data points (largest cluster) p and α is the proposed modified suppression factor. The algorithm is iteratively updated by minimizing the cost function until the following terminating condition (ε is the predefined threshold) is satisfied. maxvi+1 − vi  < ε

(16)

66

F. Ahmed et al.

3 Computational Evaluation The implementation, evaluation procedure, and the results are discussed in this section. The proposed framework is implemented on MATLAB 2013b version running in a machine with processor Intel Core i7-4770 CPU, operating frequency 3.40 GHz, memory 16 GB and hard disk of 2 TB.

3.1 Computational Setup Empirical parameters are considered to analyze the proposed algorithm’s performance. In this paper, the number of vectors N vector = 5 are selected for processing the vector median filtering in the preprocessing stage. The authors in [6], has experimentally determined that the fuzzy parameter (m) intervals can be ranged from 1.1 to 5 and the terminating threshold (ε) from 0.01 to 0.0001, respectively. Hence, m = 1 and ε = 0.001 are selected for algorithm evaluation in this paper. In addition, the qualitative metrics, SNR, MSE, PSNR, and segmentation accuracy (SA) are considered to analyze the efficacy of the proposed method [12]. For quantitative evaluation, we have considered segmentation accuracy in this work. This gives the result in terms of accuracy measurement such as if the features of the image can segment the image appropriately from each other and between the background and foreground then the algorithm will show better performance. It is observed that the requirements for segmentation accuracy are harder if we need to extract specific information from the images such as spatial or spectral properties or any temporal features. For instance, the gray matter features are very sensitive to the segmentation accuracy and it indicates that capturing this accurately portrays the success of this quantitative analysis. In addition, we have included white matter results to show more quantitative analysis below. Moreover, FCM [13], FCMT [14], and SFCM [15] are considered as standard algorithms for analyzing the effectiveness of the proposed approach.

3.2 Results and Discussion The image samples are collected from the Internet Brain Segmentation Repository (IBSR) [16], those depict the real brain MRI images with ground truth. Fig. 2a, b illustrates two samples for gray and white matter regions provided by the IBSR where those have been labeled manually, and the remaining images illustrate the results for the clustering algorithms for (c) FCM, (d) FCMT, (e) SFCM, and (f) proposed SFCM (PrSFCM). The red mark indicates that the algorithms miss the segmentation which introduces some classification error. However, the proposed approach can avoid this classification error. Moreover, the results in terms of qualitative analysis exhibit better results

6 A Fuzzy-Based Study for Biomedical Imaging Applications

67

Fig. 2 a T1-weighted image originally collected from IBSR, b (GM) labeling and (WM) slice region c FCM, d FCMT, e SFCM, f PrSFCM

for the proposed framework over the other standard algorithms which are described below. First, the quantitative performance is shown through segmentation accuracy and the qualitative results are discussed using the related metrics. Segmentation accuracy is considered here for performance evaluation for all the algorithms [17]: Segmentation Accuracy(S A) =

c  Ai ∩ ci c j=1 c j i=1

(17)

where c represents the cluster numbers, Ai pixels set of the ith cluster obtained by the proposed method, and C i pixels set for the ith cluster that belongs to the reference image. To analyze the effectiveness of the proposed technique and considering the noisy environment, various types of noise ranging from different amount has been added to the MR image collected from the IBSR repository, as shown in Fig. 3. The images were added with 11% and 15% salt and pepper (sp) noise, speckle noise, and 10% Gaussian noise. Figure 3a depicts the results obtained considering the different algorithms implemented on image degraded synthetically with 11% sp noise. Figure 3b shows the results obtained considering the different algorithms implemented on image degraded synthetically with 15% sp noise. Figure 3c shows the results obtained considering the different algorithms implemented on degraded synthetically with speckle. The results obtained considering the algorithms implemented on image degraded synthetically with 10% gaussian noise is shown in Fig. 3d. Figures 4 and 5 represents the quantitative results for the proposed algorithm considering the metrics segmentation accuracy. It is observed from Fig. 4, that the proposed approach improves the segmentation accuracy by 3.55% for GM under all noisy environments compared to other approaches, and improves by 3.62% for WM in Fig. 5. This is a good achievement in this domain of research. The proposed algorithm shows better performance over other standard algorithms in Figs. 6, 7, 8. We compute SNR, MSE, and PSNR for the segmented image. The proposed approach depicts higher SNR, lower MSE, and higher PSNR. Hence, it can be stated that the proposed approach plays a significant contribution to biomedical image segmentation studies as presented in this paper.

68

F. Ahmed et al.

Fig. 3 a and b 11% and 15% salt and pepper (sp) noise, c speckle noise, d 10% guassian noise

Segmentation Accuracy (GM) 1.2 1

11% salt & pepper noise

0.8 0.6

15% salt & pepper noise

0.4

speckle noise

0.2

10% guassian noise

0

FCM

FCMT

SFCM

PrSFCM

Fig. 4 Segmentation accuracy results for algorithms FCM, FCMT, SFCM, and PrSFCM considering GM

6 A Fuzzy-Based Study for Biomedical Imaging Applications

69

Segmentation Accuracy (WM) 1.2 1

11% salt & pepper noise

0.8 0.6

15% salt & pepper noise

0.4

speckle noise

0.2

10% guassian noise

0

FCM

FCMT

SFCM

PrSFCM

Fig. 5 Segmentation accuracy results for algorithms FCM, FCMT, SFCM, and PrSFCM considering WM

SNR Evaluation 20 18 16

11% salt & pepper noise

14 12

15% salt & pepper noise

10 8

speckle noise

6

10% guassian noise

4 2 0

FCM

FCMT

SFCM

PrSFCM

Fig. 6 SNR evaluation for FCM, FCMT, SFCM, and PrSFCM

MSE Evaluation

0.8 0.7

11% salt & pepper noise

0.6 0.5

15% salt & pepper noise

0.4

speckle noise

0.3 0.2

10% guassian noise

0.1 0

FCM

FCMT

SFCM

PrSFCM

Fig. 7 MSE evaluation for FCM, FCMT, SFCM, and PrSFCM

70

F. Ahmed et al.

PSNR Evaluation 40 35 11% salt & pepper noise

30 25

15% salt & pepper noise

20

speckle noise

15 10

10% guassian noise

5 0

FCM

FCMT

SFCM

PrSFCM

Fig. 8 PSNR evaluation for FCM, FCMT, SFCM, and PrSFCM

4 Conclusion An enhanced segmentation-based framework has been proposed in this paper. First, vector median filtering is implemented in image preprocessing to remove the noise from the MRI images. Next, Otsu thresholding is used to provide better segmentation results for the proposed clustering algorithm. Lastly, a modified suppression factor based fuzzy c-means algorithm is employed to obtain better segmentation results. The suppression factor is modified through which better clusters and compactness in clusters can be achieved. Different noisy environments by adding noise to the T1-weighted MRI images are considered for experimental evaluation. Finally, computational experimental results demonstrate the efficiency of the proposed approach over other algorithms. Acknowledgement We acknowledge the authors of the paper [22], for their valuable suggestions while performing this study. Also, the idea helped us to obtain better results by implementing our modified suppression factor and it can be carried out to a bigger solution in biomedical image processing research.

References 1. Sarraf S, Ostadhashem M (2016) Big data application in functional magnetic resonance imaging using apache spark. In: IEEE international conference on future technologies. IEEE Press, San Francisco, pp 281–284 (2016) 2. Zhang R, Wang H, Tewari R, Schmidt G, Kakrania D (2016) Big data for medical image analysis: a performance study. In: IEEE international parallel and distributed processing symposium workshops. IEEE Press, Chicago, pp 1660–1664 3. Dudley AJ, Lee JH, Durling M, Strakowski SM, Eliassen JC (2015) Age-dependant decreases of high energy phosphates in cerebral gray matter of patients with bipolar disorder: a preliminary phosphorus-31 magnetic resonance spectroscopic imaging study. J Affect Disorders 175:251–255

6 A Fuzzy-Based Study for Biomedical Imaging Applications

71

4. Sungwoong K, Nowozin S, Kohli P, Chang DY (2013) Task-specific image partitioning. IEEE Trans Image Process 22:488–500 5. Fu KS, Mui J (1981) A survey on image segmentation. J Pattern Recognit 13:3–16 6. Bezdek JC, Keller J, Krisnapuram R, Pal NR (1999) Fuzzy models and algorithms for pattern recognition and image processing. The handbooks of fuzzy sets 7. Fan JL, Zhen WZ, Xie WX (2003) Suppressed fuzzy C-means clustering algorithm. Pattern Recognit Lett 24:1607–1612 8. Taneja A, Ranjan P, Ujjlayan, A (2015) A performance study of image segmentation techniques. In: IEEE international conference on reliability, infocom technologies and optimization. IEEE Press, Noida, pp 1–6 9. Astola J, Haavisto P, Neuvo Y (1990) Vector median filters. Proc IEEE 78:678–689 10. Kumar SR, Karnan M (2014) Review of MRI image classification techniques. Int J Res Stud Comput Sci Eng 1:21–28 11. Otsu N (1979) A threshold selection method from gray-level histograms. IEEE Trans Syst Man Cybern 9:62–66 12. Gaubatz M, Hemami S Metrix MuX visual quality assessment package, http://ollie-imac.cs. northwestern.edu/~ollie/GMM/code/metrix_mux/ 13. Bezdek JC (1981) Pattern recognition with fuzzy objective function algorithms. Advanced applications in pattern recognition. Springer, Berlin 14. Chaabane SB, Sayadi M, Fnaiech F, Brassart E (2008) Color image segmentation using automatic thresholding and the fuzzy C-means techniques. In: IEEE mediterranean electrotechnical conference. IEEE Press, Ajaccio, pp 857–861 15. Ameer A, Karmakar GC, Dooley LS (2004) Fuzzy image segmentation using suppressed fuzzy C-means clustering (SFCM). In: International conference on computer and information technology. Dhaka 16. Internet brain segmentation repository (IBSR), http://www.cma.mgh.havard.edu/ibsr/ 17. Chen L, Chen LP, Lu M (2011) A multiple-kernel fuzzy C-means algorithm for image segmentation. IEEE Trans Syst Man Cybern 41:1263–1274 18. Law T, Itoh H, Seki H (1996) Image filtering, edge detection, and edge tracing using fuzzy reasoning. IEEE Trans Pattern Anal Mach Intell 18:481–491 19. Cai W, Chen S, Zhang D (2007) Fast and robust fuzzy C-means clustering algorithms incorporating local information for image segmentation. Pattern Recognit 40:825–838 20. Krinidis S, Chatzis V (2010) A robust fuzzy local information C-means clustering algorithm. IEEE Trans Image Process 19:1328–1337 21. Chaabane SB, Sayadi M, Fnaiech F (2008) Color segmentation using automatic thresholding and the fuzzy C-means techniques. In: 14th IEEE mediterranean electrotechnical conference, pp 857–861 22. Alamgir N, Kang M, Kwon YK, Kim CH, Kim JK (2012) A hybrid technique for medical image segmentation. BioMed Res Int

Chapter 7

Meta Classifier-Based Ensemble Learning For Sentiment Classification Naznin Sultana and Mohammad Mohaiminul Islam

1 Introduction With the widespread usage of social networks like Facebook, Twitter, Orkut, blogs, and forums, peoples get chance and flexibility to express their opinion about any products, support systems, or services at any time on these platforms. These opinions sometimes burst out as their emotional express. So such reviews appear as a key factor and useful resource for people’s sentiment analysis. Due to this reason, researchers started to focus on these review data to grab emotional information and automatically categorize them into different polarity levels and also some sentiment levels [1]. Sentiment classification techniques can help researchers and decision makers to have an insight about customers’ feeling, opinions, and satisfactions or resentment about any commodity or services. Today, the enormous data obtainable from social media can have substantial value for society when they are assessed as part of opinion mining analysis. Therefore, finding the right techniques and models for the sentiment analysis of big data has become a crucial activity to obtain greater value from the available data. The objective of our study is to maximize the potential usage of these kinds of data found in the web, as sentiment can be analyzed from this data source in order to determine trends and other useful information on a diverse domain [2]. Some researchers use meta-level features of base classifiers while others use ensemble learning, but few works has been found that integrate both in the model. The main concern of this paper was to explore the effectiveness of meta classifier in conjunction with ensemble learning for sentiment classification. In this paper, seven popular supervised-based machine learning algorithms have been chosen for N. Sultana (B) · M. M. Islam Department of CSE, Daffodil International University, Dhaka, Bangladesh e-mail: [email protected] M. M. Islam e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_7

73

74

N. Sultana and M. M. Islam

sentiment classification from text and ensemble approach (bagging, boosting, and stacking) were employed on them. Finally, the individual prediction accuracy of these classifiers were compared with the proposed ensemble approach. The algorithms we have chosen are: Naïve Bayes (Multinomial and Bernoulli), Logistic Regression, Support Vector Machine (SVM), Stochastic Gradient Descent (SGD), K-Nearest Neighbor (KNN), and Multilayer Perceptron (MLP). The ensemble classification takes into account the classification results of these classifiers and uses the Majority voting (stacking) method for deciding the final sentiment class prediction. Bagging and boosting approaches were applied on SVM and SGD classifiers in order to improve the result of these learners. Bagging refers to bootstrap aggregation, which can reduce the variance of an estimate by averaging together multiple estimates; boosting is an ensemble technique that is able to convert weak classifiers to a strong one by an iterative approach and stacking method combines multiple classification or regression models via meta classifier [3]. We have considered only two sentiment polarities for classification purpose, i.e., positive and negative. The rest of this paper is organized as follows: Sect. 2 outlines related works from literature; Sect. 3 presents the methodology; experimental setup and results are provided in Sect. 4, and finally Sect. 5 addresses the conclusion and potential extension of the work.

2 Related Works Sentiment analysis deals with the identification and extraction of sentiment related information from mining the available data source for decision-making and knowledge gathering. The area concerns with the use of computational linguistics, text analysis and natural language processing (NLP). Many studies on sentiment classification can be found in the literature, but few works have been done so far that use ensemble learning for text analysis. Nie et al. [4] developed an embedded and blended framework for supervised, unsupervised, and semi-supervised feature reduction approaches. Aue and Gamon [5] proposed a customized sentiment classification techniques based on Expectation Maximization Approach. Tsutsumi et al. [6] investigated MCS (Multiple Classifier System) for movie reviews which outperformed single classifiers. Abbasi et al. [7] analyzed the effect of happiness, sadness, and fear like sentiment by using SVRCE (Support Vector Regression Correlation Ensemble) technique, which were superior than regression and some other models being tested. A hybrid technique by using multiple classifiers and a combination of SVM and rule-based classifiers was proposed by Prabowo and Thelwall [8]. Whitehead and Yaeger [9] investigated ensemble methods for sentiment mining. They used Random Subspace and bagging ensemble methods and achieved the best performance. Lu and Tsou [10] proposed a combination of lexicon and SVM-based machine learning algorithm. Xia et al. [11] proposed ensemble-based feature sets and classifiers considering component weight estimation technique to achieve better performance. Li et al. [12] used ensemble learning for Chinese language using BKS (Behavior–Knowledge Space)-based ensemble technique and proved that their model outperformed other

7 Meta Classifier-Based Ensemble Learning …

75

existing techniques. Su et al. [13] investigated ensemble learning approach based on stacking approach which was more effective than other existing methods. Bootstrapped ensemble model was provided by Hassan et al. [14]. Wang et al. [15] used three ensemble approaches: bagging, boosting and Random Subspace method. They demonstrated that Random Subspace performs better than other existing methods. Bayesian Model Averaging ensemble method was proposed by Fersini et al. [16] and the result showed that the Bayesian model outperformed other approaches. Although different authors proposed different models using ensemble method for sentiment classification, however, there is no ultimate terminology of ensemble learning. Jain et al. [17] proposed a framework consisting of 18 classifiers and was able to show that the ensemble model outperforms any single classifier; Witten and Frank [18] also worked on multiple models in details, i.e., bagging, boosting, stacking and error correcting output codes; bagging with sub-bagging, boosting using AdaBoost and stumping algorithm and the mixture of experts method was proposed by Marsland [19], Alpaydin [20], in his book on machine learning provides seven methods using multiple learners such as ensemble-based bagging, boosting and voting; errorcorrecting output codes; cascading; stacking and mixtures of experts. Most of the approaches discussed above are widely used in sentiment classification, however, they suffer from some limitations in case of feature selection, dimensionality reduction and cost of searching out optimal methods. Our work addresses these concerns. We have compared different ensemble approaches and used seven base learners in addition with four ensemble (bagged and boosted) learners. Two highdimensional and balanced datasets have been used and fivefold cross-validation were performed on the experiment. Moreover, we performed feature selection within each run and each fold of the cross-validation process to avoid overfitting of the built classification models. We have used Chi2 dimensionality reduction approach to reduce run time. In addition, stacking with voting scheme was applied for the final prediction of the model.

3 Methodology This section introduces our sentiment classification model and presents the different ensemble methods we have employed on our model.

3.1 Proposed Sentiment Classification Model Our sentiment classification model is based on vote ensemble classifier utilizes from 11 individual classifiers: Two-class Multinomial NB, Bernoulli NB, LR, Linear SVM, SGD, Bagging (Linear SVM and SGD), Boosting (Linear SVM and SGD), KNN, and MLP. In order to optimize the parameters of SVMs, we used CV Parameter Selection and analyzed the result for various values of C to get the optimal accuracy. Figure 1

76

N. Sultana and M. M. Islam

Fig. 1 Proposed sentiment classification model

shows the framework of our proposed model. After collection and preprocessing of data, the datasets were trained and tested with all the classifiers. The output of these classifiers are fed into the ensemble classifier where we used Majority Voting combination rule. In addition, the bi-grams (Bag of Words model) feature selection technique was employed in our experiment because it provides the best performance. We conducted lots of combinations of different classifiers with different parameter sets and combination rules to build our model. Though the use of ensemble methods in the model improves the performance of the classifier, however, the disadvantage is the time it takes to complete the training phase. So our concern was to build a time effective model whose performance would be better in an average than the other individual classifiers in the model. In our experiment, the same datasets were used to train all the classifiers individually and the whole process of data preparation to model training and classification task was accomplished by using Python. Our model not only uses the ensemble method (Voting), but also applies meta classifier (Bagging and Boosting) as the classifier component of some selected classifiers to improve the model’s prediction accuracy. Each classifier in our model learns some parts of the sentiment classification problem and ensemble method combines these hypotheses and predict the polarity level of the data being tested in respect of sentiment.

7 Meta Classifier-Based Ensemble Learning …

77

3.2 Bagging Bagging also known as bootstrap aggregation is one of the effective methods of ensemble learning used for classification and regression analysis that produce individual models in a way for its ensemble by training each classifier on a randomly distributed training subset, and thereby improve accuracy and stability of the associated learning algorithm. This process generates different random sampling of the training dataset for each individual classifier. The main idea of bootstrapping is sampling with replacement and hence, some instances are represented multiple times while others are left out. Usually, the prediction for regression analysis is performed by averaging the results of the applied models and for classification problem majority voting approach is used. Mainly, bagging approach is effective for inconsistent learning algorithms, which means a minor change in the training set can produce have a greater impact on the learner’s prediction capability [3, 21].

3.3 Boosting Boosting is also called as model averaging method. It is a meta algorithm-based ensemble learning method that is most widely used to improve the performance of weak machine learning algorithms by some iterative approach. Though it is originally designed for the classification task, however, it can also be successfully applied to the regression problem as well. This technique works well on classifiers which is weak in terms of prediction accuracy. This process starts with the assumption that the accuracy on the training set is only slightly better than random guessing. So, a number of successive models are built iteratively, each one being trained on the data set for which some points are misclassified (for classification problem) or predicted poorly (for regression problem) by the previous model and are given more weights. This way all of the successive models are weighted according to their success and after the iterations (based on the limit of the base learners) stops, the outputs are combined using voting or averaging technique. By this process, the final model gains the ability to generate quite better result than the initial model [3, 21].

3.4 Voting Voting is a meta algorithm that combines several classifiers and generates prediction decision by analyzing the output of those classifiers using some combination rules. The Vote algorithm follows some basic rules based on probability such as maximum probability, minimum probability, multiplication of probabilities, average of probabilities, and majority voting. For the majority voting technique, the class labels predicted by each classifier are considered and the majority of the class label

78 Table 1 Dataset statistics

N. Sultana and M. M. Islam

Dataset

Positive

Negative

Total

Amazon

30,000

30,000

60,000

Airline

22,098

19,298

41,396

predicted by the classifiers are taken into account to label the class of a test data. For example, let us consider a classification problem consisting of three classifiers and two classes, i.e., Class A and Class B. If the prediction results for a sample are like that (Classifier1 predicts Class B, Classifier 2 predicts Class B, and Classifier 3 predicts Class A), then the sample data would be categorized to Class B as per the majority voting rule. For the case of tie, the label can be assigned based on the ascending sort order of the classifiers. In case of average of probabilities approach, the maximum value of the average of predicted probabilities is considered to determine the class label. There are some other techniques which combine the decision of the learners (rather than separately) and obtains the weights by maximizing the performance of the whole set [3, 21].

4 Experimental Evaluation This section consists of dataset description, preprocessing, experimental setup, and result analysis.

4.1 Dataset Description To evaluate the performance of Bagging, Boosting, and Stacking approach used in our model, we considered two labeled datasets. One is book review dataset obtained from most popular Amazon web site which consists of 60k reviews and another is Twitter US Airline Sentiment dataset, consists of 41k tweets collected from Dream to Learn 1001 data repository site. These data sets were selected because they came from realworld problems and varied in size and characteristics. Our datasets consists of about 9173 unique books and 362 different airlines data. Table 1 gives the characteristics of our data sets. In airline dataset of Table 1, the number of positive and negative reviews were unbalanced, so we performed re-sampling with replacements for this dataset to avoid biasing the classifiers toward one specific class.

7 Meta Classifier-Based Ensemble Learning …

79

4.2 Preprocessing Different preprocessing techniques were applied to remove the noise and redundant information from the datasets. Preprocessing also help to reduce the dimension of the data set significantly and hence can build more accurate classifier in less time. Two tasks associated with this is data preprocessing and feature selection process. The following preprocessing operations were performed on our datasets: Data preprocessing. Data preprocessing operation consisted of the following steps: • • • • • •

Dropped null values from review text and other columns of the datasets. Pulled out all the numeric values from the review text. Converting all capital letters of the review texts into lower case letter. Removed all the punctuations from text data. Removed Stop words from text data. Lemmatized the text data (Lemmatization was performed using WordNet’s built-in morphy function of NLTK package). • Stemming the text data (used Python NLTK Package’s implementation of stemming function by using Porter stemming algorithm). • Tokenized the text data. • For Amazon, dataset threshold value of Helpfulness attribute was set to 0.5 and dropped all rows with helpfulness values < threshold. Feature selection. This section consisted of the following steps: • Selected the review text field as a feature for Airline dataset and review text and summary field as features for the Amazon dataset. • Converting text into bigram. • 80–20% split for training and testing. • Vectorized the text data with tf–idf and Count Vectorizer separately. • Used Chi2 feature selection method for selecting top 50,000 features from text data. Experimental Setup. We conducted our experiment using Python with its robust NLP libraries for natural language processing. Following are the implementation of different classifiers where parameter optimization was performed to achieve better accuracy: -MultinomialNB(with alpha=2.0,Additive(Laplace/Lidstone) smoothing parameter (0 for no smoothing). -BernoulliNB(with alpha=2.0,Additive (Laplace/Lidstone) smoothing parameter (0 for no smoothing). -LogisticRegression(with C=1.0,Inverse of regularization strength; must be a positive float). -LinearSVM(With C=1.0,selected the C values from fivefold cross-validation).

80

N. Sultana and M. M. Islam

-BaggedLinearSVM(With C=1.0,selected the C values from fivefold cross-validation). Bagged model was implemented with BaggingClassifier(with n_estimators = 50,oob_score =False). -BoostedLinearSVM(With C=1.0, selected the C values from fivefold cross-validation). Boosted model was implemented with AdaBoostClassifier(With n_estimators =50,algorithm =“SAMME”). -SGD(With tol=0.0001, max_iter=10000,learning_rate= ‘constant’,eta0=0.1). -BaggedSGD(With tol=0.0001,max_iter=10000, learning_rate =‘constant’,eta0 =0.1).Bagged model was implemented with -BaggingClassifier(with n_estimators = 50, oob_score =False). -BoostedSGD(Withtol=0.0001,max_iter=10000, learning_rate = ‘constant’,eta0 =0.1) Boosted model was implemented with AdaBoostClassifier(With n_estimators =50, algorithm=“SAMME”). -KNN(n_neighbors=12,weights=’uniform’, algorithm=’auto’, leaf_size=30,p=2,metric=’minkowski’,metric_params=None, n_jobs=1). -Multi-LayerPerceptron(With solver=‘lbfgs’,alpha=1e-5, random_state=1). -Stacking was implemented with the StackingClassifier() method from Python mlxtend classifier. Three different classifiers were used as meta classifier with different hyper parameters: -StackingClassifier(with meta_classifier=SGDClassifier (tol=0.0001,max_iter=10000,learning_rate=‘constant’,eta0=0.1), use_features_in_secondary =True). -StackingClassifier(with meta_classifier= Logistic Regression(C=1.0),use_features_in_secondary =True). -StackingClassifier(with meta_classifier=LinearSVC ()). Result Analysis. The metric we have chosen for evaluating the model performance is classification accuracy. The accuracy of a classifier is the ability of a given classifier to correctly predict the label of a newly unseen data measured as the ratio of total number of correct predictions by the total number of input samples. In our proposed approach, at first, the 11 base classifiers are constructed individually to obtain a very good generalization performance. Second, the ensemble method voting is applied for the final prediction. We used the fivefold validation to evaluate the machine learning classification task and two combination rules such as tf–idf + bigram and count vectorizer + bigram was applied separately during the feature selection process in order to investigate in which combination the model could perform well in terms of overall prediction accuracy. Test results for both combinations

7 Meta Classifier-Based Ensemble Learning …

81

Table 2 Classification accuracy of different classifiers tf–idf + Bigram (Amazon) %

Count vectorizer + bigram (Amazon) %

tf–idf + bigram (Airline) %

Count vectorizer + bigram (Airline) %

MNB

90.91

90.99

85.33

85.4

BNB

88.86

88.14

84.25

84.13

LR

91.62

93.03

88.61

87.27

L. SVM

93.11

91.64

89.03

84.32

L. SVM (Bagged)

93.11

91.64

89.03

84.32

L. SVM (Boosted)

93.11

91.64

89.03

84.32

SGD

91.51

88.2

88.61

85.13

SGD (Bagged)

91.49

91.24

88.59

87.97

SGD (Boosted)

91.49

91.24

88.59

87.97

KNN

68.11

69.04

54.96

68.74

MLP

92.71

92.24

86.13

86.31

Majority voting (SGD)

92.71

91.64

86.13

85.09

Majority voting (LR)

92.73

92.07

86.29

84.38

Majority voting (L. SVM)

92.71

92.24

86.06

85.99

are shown in Table 2. From the table, it has been found that all the classifiers generate almost similar accuracy, except KNN which got the lowest accuracy in comparison with other classifiers for the same datasets. The accuracy of KNN is 68.11% for Amazon dataset and 54.96% for Airline dataset. In an average, linear SVM got the highest accuracy for both the datasets. The classification accuracy of all classifiers with tf–idf and count vectorizer for Amazon and Airline dataset is shown graphically in Figs. 2 and 3, respectively. Despite the prediction accuracy for both tf–idf and count, vectorizer was found almost the same in the experiment, however, the impact of bagging and boosting approach was found mostly when we applied count vectorizer as a feature selection method. According to the result, we found that the proposed model (count vectorizer + bigram + majority voting) shows a significant improvement in classification accuracy rather than using single classifier or some other combinations. Also, the results are found to be statistically significant in terms of classification accuracy. It is to be mentioned here that we have investigated using three meta classifiers during voting process to see the impact and found that the prediction capability of linear SVM as a meta classifier produce a consistent result and can provide the best accuracy than others. Though some single classifier generated better accuracy in some aspects but, in general, our model has the advantage of producing a good result when dataset

82

N. Sultana and M. M. Islam

Accuracy (Amazon Dataset) 100.00% 80.00% 60.00% 40.00% 20.00% 0.00% ƞidf+Bigram

tfidf+Bigram

Count vectorizer+Bigram

Fig. 2 Accuracy of different classifiers (Amazon dataset)

Accuracy (Airline Dataset) 100.00% 80.00% 60.00% 40.00% 20.00% 0.00% ƞidf+Bigram

tfidf+Bigram

Count vectorizer +Bigram

Fig. 3 Accuracy of different classifiers (Airline dataset)

varies in size, dimension, and scale, i.e., ensembles are appropriate when no particular model is highly likely to be correct for any one point in the input space.

5 Conclusion This study investigates thoroughly the effectiveness of ensemble approach for sentiment classification from review data. The main concern of this paper was to present an extensive analysis of Bagging, Boosting, and Stacking ensemble learning for sentiment analysis. From the comparison, it is clear that the proposed model shows signif-

7 Meta Classifier-Based Ensemble Learning …

83

icant improvement over the single classifiers. Our results demonstrate that ensemble learning (Bagging, Boosting, and Stacking) can outperform a single classifier in some cases and shows a more consistent result. However, for some data sets, Boosting may show zero gain or even a decrease in performance from a single classifier. Analysis of our results suggests that the performance of both Bagging and Boosting methods (Ada-Boosting) is at least partly dependent on the data set being examined. We believe that our approach for sentiment analysis can be applicable to other social media analysis and any other classifier could easily be integrated into the proposed framework. Future work may include the inclusion of more datasets, especially those for more specific purposes and using some unbalanced and unlabeled dataset to observe if our findings remain valid. In addition, the classification task could be expanded by considering the neutral text and other ensemble approaches (weighted average scheme) by considering other lexicon resources in sentiment analysis.

References 1. Agarwal A, Xie B, Vovsha I, Rambow O, Passonneau R (2011) Sentiment analysis of twitter data. In: Proceedings of the workshop on languages in social media. ACL, pp 30–38 2. Alnashwan R, O’riordan A, Sorensen H, Hoare C (2016) Improving sentiment analysis through ensemble learning of meta-level features. In: KDWEB, pp 1748 3. Smolyakov V (2017) Ensemble learning to improve machine learning results. Stats & Bots 4. Nie F, Huang H, Cai X, Ding C (2010) Efficient and robust feature selection via joint L2,1norms minimization. In: Proceedings of advances in neural information processing systems, pp 1813–1821 5. Aue A, Gamon M (2005) Customizing sentiment classifiers to new domains: a case study. In: Proceedings of recent advances in natural language processing, vol 3(3), pp 16–18 6. Tsutsumi K, Shimada K, Endo T (2007) Movie review classification based on a multiple classifier *. Learning 2005:481–488 7. Abbasi A, Chen H, Thoms S, Fu T (2008) Affect analysis of web forums and blogs using correlation ensembles. IEEE Trans Knowl Data Eng 20(9):1168–1180 8. Prabowo R, Thelwall M (2009) Sentiment analysis: a combined approach. J Inf 3(2):143–157 9. Whitehead M, Yaeger L (2010) Sentiment mining using ensemble classification models. Innovations and advances in computer sciences and engineering. Springer, Berlin, pp 509–514 10. Lu B, Tsou BK (2010) Combining a large sentiment lexicon and machine learning for subjectivity classification. In: Proceedings of international conference on machine learning and cybernetics, ICMLC 2010, vol 6, pp 3311–3316 11. Xia R, Zong C, Li S (2011) Ensemble of feature sets and classification algorithms for sentiment classification. Inf Sci (Ny) 181(6):1138–1152 12. Li W, Wang W, Chen Y (2012) Heterogeneous ensemble learning for Chinese sentiment classification . J Inf Comput Sci 9(15):4551–4558 13. Su Y, Zhang Y, Ji D, Wang Y, Wu H (2013) Ensemble learning for sentiment classification. In: Proceedings 13th Chinese conference on chinese lexical semantics, pp 84–93 14. Hassan A, Abbasi A, Zeng D (2013) Twitter sentiment analysis: a bootstrap ensemble framework. In: Proceedings - SocialCom/PASSAT/BigData/EconCom/BioMedCom 2013, pp 357–364 15. Wang G, Sun J, Ma J, Xu K, Gu J (2014) Sentiment classification: the contribution of ensemble learning. Decis Support Syst 57(1):77–93 16. Fersini E, Messina E, Pozzi FA (2014) Sentiment analysis: bayesian ensemble learning. Decis Support Syst 68:26–38

84

N. Sultana and M. M. Islam

17. Jain AK, Duin RPW, Mao J (2000) Statistical pattern recognition: a review. IEEE Trans Pattern Anal Mach Intell 22(1):4–37 18. Witten IH, Frank E (1999) Data mining: practical machine learning tools and techniques with Java implementations. Morgan Kaufmann, Massachusetts 19. Marsland S (2015) Machine learning: an algorithmic perspective. MIT Press, Cambridge 20. E. Alpaydin (2014) Introduction to machine learning 21. Sujatha G, Rani KU (2013) An experimental study on ensemble of decision tree classifiers 2(8):300–306. https://ijaiem.org/

Chapter 8

Mining Periodic Patterns and Accuracy Calculation for Activity Monitoring Using RF Tag Arrays Md. Amirul Islam and Uzzal Kumar Acharjee

1 Introduction In numerous applications, it is vital to monitor activities in thick arenas. It is needed to monitor behaviors in the doctor’s facility, industrial workshops, concoction plants or distribution center. Intuitively, the regular trajectories about transferring objects repeatedly follow ordinary patterns. When we bear it patterns, strange activities of moving objects can be effortlessly recognized via pattern matching [1]. Monitoring with video cameras poses some limitations and restrictions to use in wide area. Initially, the goal trajectories have to lie predefined. The cameras should be redeployed when the trajectories are changed. Also, it is hard to check different fields aside from target trajectories with camera. Automatically inspecting the pictures from numerous cameras and detecting irregular activities includes a great deal of processing activities [2]. In this paper, we talk about an utilization of the RFID innovation to give a modest and moderately exact way to deal with activity monitoring of periodic behaviors. Instead of using a series of video cameras, we use an array of RF tags and a few RF readers and also apply the data mining techniques to detect and analyze periodic activities. Our proposal is more adaptable and much cost-proficient than the video monitoring arrangements in the investigation of periodic activity monitoring. The remaining part of this paper is organized as follows. In Sect. 2, the related works are described while some preliminary concept is discussed in Sect. 3. We briefly discuss our problem statement in Sect. 4. Our Proposed Method and the periodic mining algorithm of activity monitoring using RF tag arrays are discussed in Sect. 5. Experimental evaluation is shown in Sect. 6 and in Sect. 7, concludes our work with future research directions. Md. Amirul Islam (B) · U. K. Acharjee Department of Computer Science and Engineering, Jagannath University, Dhaka, Bangladesh e-mail: [email protected] U. K. Acharjee e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_8

85

86

Md. Amirul Islam and U. K. Acharjee

2 Related Works Recently, periodic pattern mining has been considered extensively. Han et al. [3] first proposed fractional periodic pattern mining algorithm in timeseries databases utilizing halfway pattern mining properties, for example, Apriori property. Tanbeer et al. [4] presented finding periodic frequent patterns in value-based databases. Ma and Hellerstein [5] proposed a similar, Apriori-based method containing two level wise algorithms in unknown period value. Yang et al. [6] presented a novel non-concurrent sort of intermittent pattern mining algorithm to discover all patterns inside the scope of information arrangement with a most extreme number of interruption permitted. Huang and Chang [7] proposed another asynchronous method which validates segment and sequence through a minimum number of repetitions of patterns. In [8], a prediction method based on the periodic pattern is proposed. To answer predictive queries efficiently, a trajectory pattern tree is proposed to index the periodic patterns. One significant use of periodic pattern is for future area prediction. Most existing strategies focus on not so distant future development expectation, for instance, one minute from now, one hour from now, following day or one month after. Periodic patterns can enable better to anticipate future development, particularly for a removed question time. In this manner, these actualities spurred us to propose a Periodic patterns mining strategy that is more adaptable and much cost-effective than the video checking arrangements in the study of periodic activity monitoring. This procedure is time and memory productive.

3 Preliminaries Periodic pattern mining researches are classified as structured and unstructured. Trajectory systems are the case of organized information, show charts over circumstances. Diagram vertices show the tag location and chart edges express among the trajectory among tag locations [9]. Definition 1 (Periodic Pattern (PP)): Given a trajectory and an arbitrary pattern P. The P P is a pair of < P, Sp(P) >, where P is closed pattern over a periodic support set S p (P) with |L p (P)| > μ and S p (P) is maximum for P [9]. It maintains subsumption property in Property 1. Property 1 Subsumption: Two periodic pattern P1 and P2 , their support set S p (P1 ) = (t1 , p1 , s1 ) and S p (P2 ) = (t2 , p2 .s2 ). P1 support set < P1 , S p (P1 ) > contains or subsumed P2 support set < P2 , S p (P2 ) > if and only if the following condition hold. i. P2 ⊆ P1 ii. t2 ≥ t1 iii. c2 mod c1 = 0 and c1 < c2

8 Mining Periodic Patterns and Accuracy Calculation …

87

iv. t2 + c2 (o2 − 1) ≤ t1 + p1 (o1 − 1) v. (t2 − t1 ) = c. j for some integer j > 0.

4 Problem Statement A typical methodology in finding periodic movement patterns is to apply the mining on successions of areas. The objects pursue similar routes over ordinary time interims. The finding of covered up hidden periodic patterns in the dataset could express imperative data to the information investigator. The RF labels conveyed are stationary; their spatial areas are known to the server. The data mining assignment comprises two stages: the training stage and the monitoring stage. In the training stage, we accumulate the RF tag signal sequences over n periods, where n is a client determined time allotment. The arrangements in the training stage will be utilized to discover periodic activities as the model of the ordinary activities in the field. In the monitoring phase, activities are detected and compared with the periodic behaviors. If an activity matches behaviors, it is viewed as normal. Otherwise, an alert will be issued. In the rest of the paper, we focus on the periodic activities mining problem (i.e., the training phase), activities are detected and compared with the periodic behaviors (i.e., the training phase). In summary, the problem of mining periodic activities patterns from RF tag sequences is to explore the trajectories happening at least Th times in the training phase, where Th is a user-specified frequency threshold.

5 Proposed Method The principal contribution of the proposed method is that we use an array of RF tags and a few RF readers and also apply the data mining techniques to detect and analyze periodic activities. Essentially, here we discuss the possible object positions and periodic activities mining algorithm.

5.1 Identifying Possible Object Positions We identify the probable locations of objects using the spatial map of the stationary tags. Normally, the areas of objects are parts of the border that a reader can see. Various objects may exist in a similar range. In our periodic activities mining algorithm, we might utilize such ranges to gather the possible periodic behaviors. Another essential introduction is that a few objects may hide behind other objects. To recognize those hidden objects, we apply the accompanying two techniques. First,

88

Md. Amirul Islam and U. K. Acharjee

we employ multiple readers. Second, we conduct fault-tolerant method. However, it also two major challenges. Challenge 1: Accurate detection of object’s position. RFID data is very noisy and tags often have various characteristics. Some of them are very sensitive, i.e., their signal is not stable even when there is no activity. The magnitude of the RF tags also varies. Different RF tags may also give different signal changes even if they have the same interference. Challenge 2: Detection of the periodic activity’s in trajectories. The RF tags do not send their signals synchronously; some activities may skip one or a few tags.

5.2 The Periodic Activities Mining Algorithm The periodic activities are mined in the following two steps. Finding Periodic Positions of Objects Those trajectories of a periodic movement might occasionally trigger a tag in the item area segments. Eventually filtering the object area segments on the whole periods once, we might find the tags that would in the segments over in any event Th periods for readers. Algorithm 1: Find periodic positions of objects.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18

Data: RF tag signal sequences Sig(μ)iR , periodic threshold Th . Result: The set of periodic positions of objects w.r.t reader R; for each tag μ do Create a counter Countμ = 0 and a flag Flagμ = 0 end for each period j do for each tag μ do if Sig(μ)iR = 1 then Countμ = Countμ + 1; end if μ is at the border of interfered tags then Flagμ = 1; end end end for each tag μ do if Countλ ≥ Th AND Flagμ = 1 then Output μ as a periodic position; end end

Finding Set of Periodic Activities Let, N = (x1 , y1 , t1 ), (x2 , y2 , t2 ), . . . be the original movement database for a moving object. The interpolated sequence is denoted as P O S I = posi 1 posi 2 . . . posi n , where posi i is a spatial point represented as a pair

8 Mining Periodic Patterns and Accuracy Calculation …

89

( posi i .x, posi i .y). Given a location sequence POSI, our problem aims at mining all periodic behaviors. RS = r s1 , r s2 , . . . , r sd is the set of all reference spots, where d is the number of reference spots. A period T is a regular time interval in the (partial) movement. Algorithm 2 represents a framework to find set of periodic activities.

Algorithm 2: Algorithm to Find Set of Periodic Activities.

1 2 3 4 5 6 7 8 9 10 11 12

Data: A movement sequences P O S I = posi 1 posi 2 . . . posi n . (Algorithm 1) Result: A set of periodic activities. /* Stage 1: Detect periods */ Find reference spots R S = {r s1 , r s2 , . . . , r sd }; for r si ∈ RS do Detects periods in r si and the store the periods in Periodi ; Periodset ← Periodset ∪ Periodi ; end /* Stage 2: Mine periodic activities */ for T ∈ Periodset do R ST ={ r si |T ∈ Periodi }; Construct the symbolized sequence S using R ST ; Mine periodic activities in S. end

5.3 Time Complexity We analyze the computational complexity of the periodic activities mining. So as with would this, we infer a upper bound for the Algorithm 1: To find periodic activities position objects and also calculate upper bound for Algorithm 2: to find set of periodic activities of T timesteps. We substantiate that this upper bound may be a polynomial work of the number for timesteps and the least help quality. We show that the upper bound will be sharp toward constructing a ’worst-case’ periodic activities. Time Complexity of Proposed Method Here, we first calculate the time complexity of Algorithm 1. – For first loop executes runtime is O(n) times. – Outer loop executes period O(n) times and for each of those times, inner loop executes O(m) times. This is the nested loop. So, the body of the inner loop is executed O(mn) times. – For last loop executes runtime is O(n) times. Consecutive statements: This is just add all runtime O(n) + O(mn) + O(n) = O(mn). So, our first algorithm time complexity is O(mn). Where n and m is the total number of tags. Now, we calculate the time complexity of Algorithm 2. – For first loop executes runtime is O(m) times. – For last loop executes runtime is O(n) times.

90

Md. Amirul Islam and U. K. Acharjee

Consecutive statements: This is just add all runtime O(m) + O(n) = O(m + n). So, our second algorithm time complexity is O(m + n). Memory Complexity of Proposed Method The memory complexity of an algorithm is the amount of memory it uses. (Just like the fact that the time complexity of an algorithm is the amount of computing time it uses.) The memory complexity of proposed Algorithm 1 is (n). The memory complexity of proposed Algorithm 2 is (m + n).

6 Experimental Evaluation In this experimental study, we observe our periodic activities mining algorithm on a real implementation of 80 RF tags and 1 reader. Toward applying our mining algorithm on the readings for each RF tag gained starting with those readers, we report the accuracy and effectiveness of identifying periodic behaviors. On measure those identify accuracy, we use Eq. 1. We start those tests with single activities with one direction and one route for particular case object. Finally, we inspect the complex activities with multiple objects.

6.1 Experiments Environments A comprehensive performance study has been conduct for real, synthetic datasets to expose the performance of proposed Periodic Activities mining methods. The two algorithms are implemented in the JAVA and Data generator implemented in C++, and all the experiment carried out on 2.70 GHz Intel Core i3 with 2 GB memory. The system runs Microsoft Windows 7 and visual studio 2010 and NetBeans IDE 8.2. In these implementations, Google dense/sparse hash library has been used because of it is time and memory efficiency. In our experimental results, the sum of users and CPU time is reported as computational time. Memory usage is measured by the maximum used space that showed by C++ memory usage function.

6.2 Datasets A data set is a collection of data. To check the validity of any algorithm datasets are most important factor. There are many spatiotemporal datasets which have been classified two categories: real datasets and synthetic datasets. We have used both categories of data in our experiment analysis. The description of these two categories is given below.

8 Mining Periodic Patterns and Accuracy Calculation …

91

Real Datasets Real spatiotemporal datasets are collected from various sources like as GPS datasets and Buffalo Datasets. In the experimental analysis, to show the effectiveness of our proposed system we use one real datasets that are flows. GPS Datasets: GPS datasets contains 35044 objects evolving 10332 distinct timestamps at seven locations. Buffalo Datasets: This datasets includes 421 buffaloes traveled 12 locations over 1225 different timestamps from year 2010 to the year 2016. Synthetic Datasets To show the performance study of our algorithms, we also generate three synthetic datasets. O400T100L50-1000 [10]: We use 400 objects (|Odb | = 400) for 100 timestamps (|Tdb | = 103 ) at 50 locations (|L db | = 50), where total 1000 location and timestamps pair are generated. Each pair of location and timestamps, the number objects generates randomly between 1 and 400. O300T110L150-2000: We generate 200 objects (|Odb | = 300) for 110 timestamps (|Tdb | = 110) at 50 locations (|L db | = 150), where total 2000 location and timestamps pair are generated. Each pair of location and timestamps, the number objects generates randomly between 1 and 150. O200T120L150-2000: We generate 200 objects (|Odb | = 200) for 120 timestamps (|Tdb | = 120) at 100 locations (|L db | = 50), where total 2000 location and timestamps pair are generated. Each pair of location and timestamps, the number objects generates randomly between 1 and 150. The datasets are shown in the flowing Table 1.

6.3 Result and Performance Analysis In this section, we calculate the results of Accuracy and Runtime with respect to support threshold Th . To measure the detection accuracy, we use Eq. 1. After calculating the accuracy we compare this result with existing method and also comparison runtime for the different experiments with respect to threshold value Th .  σ =

L Rl

 (1)

where, σ = Accuracy, L= The Length of correctly detected trajectory of periodic activities, Rl = The length of the real periodic route Table 1 Comparison parameters of various datasets

Dataset

|Odb |

|Tdb |

|L db |

GPS Buffalo O400T100L50-1000 O300T110L150-2000 O200T120L150-2000

35044 421 400 300 200

10332 1225 100 110 120

7 12 50 150 150

92

Md. Amirul Islam and U. K. Acharjee

Accuracy and Runtime Comparison for Single Activities Periodic, frequent trajecotry accuracy and runtime comparison for single activities by the Eq. 1. And also here use the synthetic datasets O400T100L50-1000. Accuracy and runtime represent in the table. From the above Table 2, we comparison the accuracy of single activities vs. periodic activities with respect to different support threshold Th . From the above Table 3, we calculate the runtime of frequent trajectory and periodic activities. This section shows the comparison between the frequent trajectory and periodic activities the runtime based on the Threshold Th . Accuracy and Runtime Comparison for Community Activities Periodic, frequent trajecotry accuracy and runtime comparison for community activities by the Eq. 1. And also here use the synthetic datasets O200T120L150-2000. Accuracy and runtime represent in the graph. From the Fig. 1 show that the Accuracy with respect to support Threshold Th for Community Activities. From above Fig. 2 show that, the Runtime with respect to support Threshold Th for Community Activities. Accuracy and Runtime Comparison for Complex Activities Periodic, frequent trajecotry accuracy and runtime comparison for complex activities by the Eq. 1. And also here use the synthetic datasets O300T110L150-2000. Accuracy and runtime represent in the graph. Table 2 Accuracy comparison

Table 3 Runtime comparison

Threshold Th

Existing frequent trajectory accuracy (%)

Proposed periodic activities accuracy (%)

2 3 4 5 6

85 60 40 30 20

93 75 60 40 30

Threshold Th

Existing frequent Proposed periodic trajectory runtime (s) activities runtime (s)

2 3 4 5 6

0.40 0.35 0.040 0.45 0.67

0.30 0.20 0.25 0.30 0.40

8 Mining Periodic Patterns and Accuracy Calculation …

93

Fig. 1 Accuracy comparison with respect to support threshold value Th for community activities

Fig. 2 Runtime comparison with respect to support threshold value Th for community activities

From the Fig. 3 show that, the Accuracy with respect to support Threshold Th for Complex Activities. From above Fig. 4 show that, the Runtime with respect to support Threshold Th for Complex Activities. Our experimental evaluation study using the RFID implementation confirms that proposed data mining techniques can detect periodic set of activities, when the activities are not exactly muddled over space and the accuracy is high.

94

Md. Amirul Islam and U. K. Acharjee

Fig. 3 Accuracy comparison with respect to support threshold value Th for complex activities

Fig. 4 Runtime comparison with respect to support threshold value Th for complex activities

7 Conclusions and Future Works Proposed algorithm can retrieve maximal periodic patterns from the frequent region which is more effective and efficient. Experimental evaluation study using real RFID datasets and synthetic datasets verifies the effectiveness of the proposed method. Currently, we are considering the cross-validation method using multiple readers, and a more thorough test in real application fields. Proposed method should be efficient and less costly to investigate the optimal deployment of RF tags and readers in a field. We do not consider the performance of RFID technology in multiple trials.

8 Mining Periodic Patterns and Accuracy Calculation …

95

Since RFID provisions frequently make a considerable measure for data, we trust the individual’s provisions will pose new challenges and chances for data mining and pervasive computing innovative work.

References 1. Lian X, Chen L, Yu JX, Wang G, Yu G (2007) Similarity match over high speed time-series streams. In: 2007 IEEE 23rd international conference on data engineering, ICDE 2007. IEEE, pp 1086–1095 2. Liu Y, Zhao Y, Chen L, Pei J, Han J (2012) Mining frequent trajectory patterns for activity monitoring using radio frequency tag arrays. IEEE Trans Parallel Distrib Syst 23(11):2138– 2149 3. Han J, Dong G, Yin Y (1999) Efficient mining of partial periodic patterns in time series database. In: Proceedings of the 15th international conference on data engineering. IEEE, 1999, pp 106– 115 4. Tanbeer S, Ahmed C, Jeong B-S, Lee Y-K (2009) Discovering periodic-frequent patterns in transactional databases. Adv Knowl Discov Data Min 242–253 5. Hellerstein JL, Ma S, Perng C-S (2002) Discovering actionable patterns in event data. IBM Syst J 41(3):475–493 6. Yang J, Wang W, Yu PS (2004) Discovering high-order periodic patterns. Knowl Inf Syst 6(3):243–268 7. Huang K-Y, Chang C-H (2004) Asynchronous periodic patterns mining in temporal databases. In: Databases and applications, pp 43–48 8. Jeung H, Liu Q, Shen HT, Zhou X (2008) A hybrid prediction model for moving objects. In: 2008 IEEE 24th international conference on data engineering, ICDE 2008. IEEE, pp 70–79 9. Halder S, Samiullah M, Lee Y-K (2017) Supergraph based periodic pattern mining in dynamic social networks. Expert Syst Appl 72:430–442 10. https://sites.google.com/site/sajalhalder/research/suarw. Accessed 05 Aug 2017

Chapter 9

Can the Expansion of Prediction Errors be Counterbalanced in Reversible Data Hiding? Hussain Nyeem and Sultan Abdul Hasib

1 Introduction Reversible Data Hiding (RDH) is being widely investigated in last two decades as a key covert-communication technology [6, 18]. Given a cover image as input, an RDH scheme invisibly embeds data in it to output an embedded image allowing complete extraction of the data and lossless recovery of the cover image afterward [14]. Development of such schemes generally aims to improve the embedding capacity with invertible and minimum possible distortion. For example, embedding capacity of the pioneering difference expansion (DE)-based RDH scheme [20] was later improved with generalized expansion [1], sorting and prediction [5, 19], and adaptive embedding [8]. Besides, the histogram shifting (HS)-based scheme [13] was improved for lower distortion using the difference-histogram [7] and multiple histogram [11]. Other potential developments include the RDH schemes with prediction error expansion (PEE) [3, 4, 9, 10, 15–17], vector quantization [12], interpolation [22, 23], and encryption [2]. Among different RDH principles mentioned above, PEE attracted much attention for its impressive combined-features of DE and HS. A prediction error is usually expanded either by embedding a bit in it or by shifting it by a “suitable” value. Unlike the use of pixel-histogram in classic HS, it deals with the prediction errors to obtain a much sharper histogram with a set of higher peak beans resulting in higher embedding capacity. Additionally, unlike the direct change of pixels in classic DE, it expands the prediction errors to offer minimum possible changes in the pixels resulting in H. Nyeem (B) · S. A. Hasib Department of Electrical, Electronic and Communication Engineering, Military Institute of Science and Technology (MIST), Mirpur Cantonment, Dhaka 1216, Bangladesh e-mail: [email protected] S. A. Hasib e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_9

97

98

H. Nyeem and S. A. Hasib

higher quality embedded images. Successive developments of PEE can be tracked with pixel value ordering (PVO) on different size pixel-group [4, 10, 15–17]. The PVO forms a PEE embedding to be directional and has been an easy solution to minimize the prediction errors. Although the pixel value grouping and sorting principle was initially utilized in [5], PVO was well established by Li et al. [10] for PEE-based embedding. Li et al.’s scheme predicted the maximum pixel from a pixel-pair and was verified to minimize distortion. Peng et al. [16] redefined prediction error with spatial order of maximum and second-maximum pixels in a block. Ou et al. [15] extended the classic PVO to PVO-k for adaptive embedding in blocks. Unlike the block-wise prediction in the original PVO, Qu et al. [17] then extended the original PVO-scheme with pixel-wise prediction using sorted context pixels for larger capacity and better image fidelity. Jung [4] lately demonstrated a minimum PVO scenario to embed two bits in every three pixels with classic PEE. Although the existing PVO-based RDH schemes generally improved the rate-distortion performance, no effort can be tracked to counterbalance the expansion of prediction errors to further minimize the distortion and increase the embedding capacity. In this paper, we therefore investigate how the expansion of prediction errors can be counterbalanced in a simple embedding scenario with PVO and classic PEE. Particularly, we introduce a new embedding technique that first embeds with PEE and PVO in every non-overlapping image block of size 1 × 3. This allows the lowest and highest valued pixels in each block are predicted from the middle valued pixel. Subsequently, all the predicted pixels are grouped into two sets, namely, the minimum-set and maximum-set. The minimum-set only contains the lowest predicted pixels, while the other contains only the largest predicted pixels of each block. Backward prediction based embedding is then applied separately on these sets to restore the predicted pixels to their original values for higher embedding rate and better embedded image quality. Rest of this paper is structured as follows. We provide a general framework of PVO-based RDH scheme in Sect. 2 and develop computational model of the proposed embedding in Sect. 3. Early experimental results are presented in Sect. 4 to demonstrate the rate-distortion performance of our scheme over the baseline RDH scheme. Concluding remarks are given in Sect. 5.

2 A PVO-Based RDH Framework In this section, we present a general framework for PVO-based RDH scheme and its use in representing the Jung’s scheme [4] that captures a simple embedding scenario with classic PVO and PEE. We will also use these notations and framework to define the computational processes of our proposed embedding in Sect. 3.

9 Can the Expansion of Prediction Errors be Counterbalanced …

99

2.1 General Framework A PVO- and PEE-based embedding generally start with a set of image blocks. An input image is partitioned into a set of non-overlapping blocks. For each block, pixels are sorted and expanded upon the prediction errors and embedding conditions. Computational modeling of this embedding process is defined as follows. An input cover image, I of size M × N is partitioned into a set of non-overlapping blocks, i.e., I = [X k ]. Each block thus contains n pixels, ı.e, X = (x1 , x2 , . . . xn ) . Once the blocks are computed, their pixand total number of blocks is k = M×N n els are sorted using a sorting function. For example, a block, X containing pixels, (x1 , x2 , . . . xn ) are now sorted in ascending order using a sorting function σ (·) to output (xσ (1) , xσ (2) , . . . xσ (n) ). Here, σ : {1, 2, . . . n} → {1, 2, . . . n} is a unique oneto-one mapping function such that xσ (1) ≤ xσ (2) ≤ · · · ≤ xσ (n) with σ (i) < σ ( j) if xσ (i) = xσ ( j) and i < j. The sorted block pixels are then used for computing prediction errors followed by their expansion with a suitable PEE-based embedding. Existing PEE-based embedding techniques rely on different embedding conditions and image partitioning rules. Since we are motivated to verify how the expanded pixels can be counterbalanced in an additional level of embedding, we reasonably consider a simple embedding scenario with PVO and classic PEE that is recently reported in the Jung’s scheme. Without loss of generality, we illustrate the Jung’s PEE-based embedding conditions below for a single data bit, b ∈ {0, 1}.

2.2 Jung’s PVO-Based RDH Scheme Jung [4] recently proposed a PVO-based RDH scheme that partitions an image into a set blocks of size 1 × 3. The scheme ideally embeds 2 bits in each block. For example, }. This means an image I is partitioned such that I = [X k ] with k = {1, 2, . . . M×N 3 that, with our above general PVO framework, Jung’s scheme operates on each block X with the number of block-pixels, n = 3. Thus the sorting function, σ (·) is used to sort the block-pixels (x1 , x2 , x3 ) to be (xσ (1) , xσ (2) , xσ (3) ), where xσ (1) and xσ (3) are the minimum and maximum block-pixels, respectively. For each block, a pair of prediction errors, emin and emax is calculated from the middle block-pixel, xσ (2) according to the Eqs. (1a)–(1b). These errors are expanded either for embedding of a data-bit, b in the error or shifting the error by a value of “1” using Eqs. (2a)–(2b). The minimum and maximum block-pixels are then predicted from the middle block-pixel and the expanded errors with Eqs. (3a)–(3b). emax = xσ (3) − xσ (2) emin = xσ (1) − xσ (2)

(1a) (1b)

100

H. Nyeem and S. A. Hasib

eˆmax

eˆmin



if emax = 0 if emax = 1 if emax > 1

(2a)

if emin = 0 if emin = −1 if emin < −1

(2b)

xˆσ (3) = xσ (2) + eˆmax

(3a)

xˆσ (1) = xσ (2) + eˆmin

(3b)

eˆmax − 1, −eˆmin − 1, ⎧ ⎪ ⎨xˆσ (3) , = xˆσ (3) − b, ⎪ ⎩ xˆσ (3) − 1,

b=

xσ (3)

⎧ ⎪ ⎨emax , = emax + b, ⎪ ⎩ emax + 1, ⎧ ⎪ ⎨emin , = emin − b, ⎪ ⎩ emin − 1,

if 1 ≤ eˆmax ≤ 2 if − 2 ≤ eˆmin ≤ −1 if eˆmax = 0 if 1 ≤ eˆmax ≤ 2 if eˆmax > 2

xσ (2) = xˆσ (2) ⎧ ⎪ if eˆmin = 0 ⎨xˆσ (1) , xσ (1) = xˆσ (1) + b, if − 2 ≤ eˆmin ≤ −1 ⎪ ⎩ xˆσ (1) + 1, if eˆmin < −2

(4a)

(4b) (4c) (4d)

Extraction of the embedded data and recovery of the original block-pixels follow the inverse PEE embedding principle of the Jung’s scheme in Eq. (4) like other PVObased RDH scheme. With recovery of the maximum and minimum block-pixels of all expanded pixels, the original image is recovered. At the same time, the data-bits are extracted from each embedded blocks and concatenated to get the original data. With a single reference pixel in a block, Jung’s scheme presented a classic PVObased embedding scenario, where embedding rate is improved with a reasonable embedded image quality. In this paper, a more effective use of the Jung’s PVO is investigated, and thus the development of a new RDH scheme for higher embedding rate and better embedded image quality is presented in the section below.

3 A New PVO-Based Embedding A PVO-based embedding has evolved to utilize image correlations for better possible rate-distortion performance. With classic PVO, pixel values in a block are kept unchanged or expanded (either for embedding or shifting) centering the reference pixel(s). This principle of embedding has been better utilized with the adaptive block

9 Can the Expansion of Prediction Errors be Counterbalanced …

101

size or multilevel embedding in recent schemes for better rate-distortion performance, as mentioned in Sect. 1. However, no effort has been made yet to counterbalance the expanded pixels to partially restore them to their respective original pixel values. In this section, we attempt to utilize this reverse expansion property of “backward PEE” in Jung’s scheme to show that such reverse expansion of the expanded pixels in an additional phase of embedding can further improve the rate-distortion performance. Our RDH scheme constitutes two embedding phases, namely, (i) forward and (ii) backward. In the first phase, all the image pixels are considered for embedding. But, the second phase only considers the expanded pixels for embedding. These two phases of embedding are expected to improve both the visual quality of the embedded image and the embedding rate. Extraction of our scheme, on the other hand, follows the inverse processing of these two phase-embedding. Rest of this section captures more computational details of our RDH scheme’s embedding and extraction processes.

3.1 Forward Embedding As mentioned above, for embedding with forward PEE, we employ the Jung’s PVO-based scheme that starts with partitioning an input image, I into a set of nonoverlapping of blocks of size 1 × 3. This is discussed in Sect. 2.2. Each block-pixels (x1 , x2 , x3 ) are sorted to obtain (xσ (1) , xσ (2) , xσ (3) ), where xσ (1) and xσ (3) are the minimum and maximum block-pixels, respectively. With the computation and expansion of the pair of prediction errors, emin and emax using Eqs. (1a) and (2b), either a data-bit, b is embedded or error-value is shifted by 1. The minimum and maximum block-pixels are then predicted from the middle block-pixel and the expanded errors with Eqs. (3a) and (3b). Given an input image, I and a set of data-bits, D f , with this forward embedding, we thus obtain the embedded image, Iˆ. We pre-process the input image to avoid the overflow and underflow problem using the conventional process of recording a location map. For example, the location map L Mb is constructed by recording a “1” for each boundary pixel and “0’s” for the other pixels. With an input image of bit-depth 8-bit, for example, a boundary pixel, x is then updated using Eq. (5). For simplicity, we omit the notational difference between the original input image and its pre-processed version with modified boundary pixels. ⎧ ⎪ ⎨x − 1, if x = 255 x = x + 1, if x = 0 ⎪ ⎩ x, otherwise

(5)

102

H. Nyeem and S. A. Hasib

3.2 Backward Embedding Backward embedding operates only on the predicted pixels in Iˆ to counterbalance the expansion made in the earlier phase. This means that the maximum and minimum pixels of a block, i.e. xˆσ (1) and xˆσ (3) , can have experienced a maximum expansion of value 1 either for embedding of a data-bit “1” or for shifting the pixel by the value 1. Thus, all the predicted pixels that experience this expansion are considered for minimum and maximum groups, Xˆ min and Xˆ max , respectively as in Eq. (6). Additionally, we locate the predicted pixels that remain unchanged in the first phase of embedding by recording their location map in L Mo . Once Xˆ min and Xˆ max are obtained, we separately and pairwise expand pixels of each set using backward PEE. In other words, both Xˆ min and Xˆ max are pairwise partitioned, sorted and used for embedding. For example, a pixel-pair (xˆ1 , xˆ2 ) ∈ Xˆ min with sorting becomes [xˆσ (1) , xˆσ (2) ]. These partitioning and sorting also apply to Xˆ max . With backward PEE, we predict xˆσ (2) from xˆσ (1) for Xˆ min using Eq. (7), and predict xˆσ (1) from xˆσ (2) for Xˆ max using Eq. (8). Thereby, we compute the set of expanded pixels, {xˆˆσ (2) } for Xˆ min and {xˆˆσ (1) } for Xˆ max to generate the expanded minimum and maximum groups, Xˆˆ min and Xˆˆ max , respectively. The final embedded image, Iˆˆ is obtained by updating Iˆ with Xˆˆ and Xˆˆ . min

max

Initialize: Xˆ min ← ∅ & Xˆ max ← ∅ For each block: Xˆ min ← Xˆ min ∪ xˆσ (1) if xˆσ (2) − xˆσ (1) > 1

(6a)

Xˆ max ← Xˆ max ∪ xˆσ (3) if xˆσ (3) − xˆσ (2) > 1

(6b)

exmin = xˆσ (2) − xˆσ (1) |(xˆσ (1) , xˆσ (2) ) ∈ Xˆ min ⎧ ⎪ if exmin = 0 ⎨exmin , eˆxmin = exmin + b, if exmin = 1 ⎪ ⎩ exmin + 1, if exmin > 1

(7a)

xˆˆσ (2) = xˆσ (1) + eˆxmin |xˆˆσ (2) ∈ Xˆˆ min exmax = xˆσ (1) − xˆσ (2) |(xˆσ (1) , xˆσ (2) ) ∈ Xˆ max ⎧ ⎪ if exmax = 0 ⎨exmax , eˆxmax = exmax − b, if exmax = −1 ⎪ ⎩ exmax − 1, if exmax > −1 xˆˆσ (1) = xˆσ (2) + eˆxmax |xˆˆσ (1) ∈ Xˆˆ max

(7b)

(7c)

(8a) (8b)

(8c)

9 Can the Expansion of Prediction Errors be Counterbalanced …

103

3.3 Decoding Decoding comprises the data extraction and image recovery processes that follow the inverse process of embedding of our proposed RDH scheme. This means, data is first extracted with the inverse embedding process with backward PEE followed by the inverse process of embedding with forward PEE. The input image to the decoder is partitioned into a non-overlapping block of size 1 × 3 and each block pixels are sorted. From the reserved pixels, the location map, L Mo is extracted. From the sorted pixels of each block and using L Mo , the sets of maximum and minimum expanded pixels, Xˆˆ max and Xˆˆ min , respectively are determined using Eq. (6). and Xˆ are recovered from Xˆˆ The embedded data bits are extracted and Xˆ min

max

min

and Xˆˆ max , respectively using Eq. (9). We start with computing the errors from each pixel-pair in Xˆˆ min and Xˆˆ max using Eqs. (9a) and (9b), respectively. Embedded bits are extracted using the error-values and conditions in Eq. (9c). The higher pixel, xˆˆσ (2) of are restored to xˆ using Eq. (9d). Similarly, xˆ ∈ Xˆ each pixel-pair in Xˆˆ min

σ (2)

σ (1)

max

is restored from xˆˆσ (1) ∈ Xˆˆ max using Eq. (9e). Thereby, we can restore Xˆ max and Xˆ min from Xˆˆ and Xˆˆ , respectively to finally compute Iˆ from Iˆˆ. max

min

eˆˆxmin = xˆˆσ (2) − xˆˆσ (1) |(xˆˆσ (1) , xˆˆσ (2) ) ∈ Xˆˆ min eˆˆxmax = xˆˆσ (1) − xˆˆσ (2) |(xˆˆσ (1) , xˆˆσ (2) ) ∈ Xˆˆ max  eˆˆxmin − 1, if 1 ≤ eˆˆxmin ≤ 2 b= ˆ −eˆxmax − 1, if − 2 ≤ eˆˆxmax ≤ −1 For all xˆσ (2) ∈ Xˆ min & xˆˆσ (2) ∈ Xˆˆ min : ⎧ ˆ ⎪ if eˆˆxmin = 0 ⎨xˆσ (2) , xˆσ (2) = xˆˆσ (2) − b, if 1 ≤ eˆˆmin ≤ 2 ⎪ ⎩ˆ xˆσ (2) − 1, if eˆˆmin > 2 For all xˆσ (1) ∈ Xˆ max & xˆˆσ (1) ∈ Xˆˆ max : ⎧ ˆ ⎪ if eˆˆxmax = 0 ⎨xˆσ (1) , xˆσ (1) = xˆˆσ (1) + b, if − 2 ≤ eˆˆmax ≤ −1 ⎪ ⎩ˆ xˆσ (1) + 1, if eˆˆmax < −2

eˆmax = xˆσ (3) − xˆσ (2) eˆmin = xˆσ (1) − xˆσ (2)

(9a) (9b) (9c)

(9d)

(9e)

(10a) (10b)

104

H. Nyeem and S. A. Hasib

The original image is finally restored with the inverse of our first phase embedding, which follows the extraction principle of Jung’s scheme (see Sect. 2.2). Upon the extraction of all data-bits, they are marged to the data-bits extracted in earlier phase. Similarly, the original image I is obtained by updating Iˆ with the restored values of xˆσ (3) and xˆσ (1) of each block.

4 Experimental Results We now present our early results to analyze and validate the performance of our RDH scheme. Test-images of size 512 × 512 × 8 are used from the USCSIPI database [21] for this performance evaluation. The embedding-capacity and embedding-rate are determined in terms of total embedded bits and bit per pixel (BPP), respectively. BPP is calculated using Eq. (11). For embedding, a set of pseudorandom bits is generated as data. The performance of our scheme is compared with the Jung’s scheme (2017) [4]. Implementations are carried out using MATLAB R2016b with a 1.3 GHz Intel core i5 CPU, 4 GB memory. BPP =

Total Capacity (bits) M×N

(11)

The embedded image quality is evaluated in terms two popular objective visual quality metrics, peak signal to noise ratio (PSNR) defined in Eq. (12) and structural similarity (SSIM) [24] defined in Eq. (13). Here, M × N is the image size, and I (i, j) and Iˆˆ(i, j) are the pixel values of position (i, j) in an original image and its embedded version, respectively. In Eq. (13), μx and μxˆˆ are the average values of ˆˆ where x ∈ I and xˆˆ ∈ Iˆˆ are the pixels of original and embedded images, x and x,

ˆˆ respectively; σ ˆ is respectively. Similarly, σx2 and σ ˆ2 are the variance of x and x, x xˆ xˆ ˆ the covariance of x and x; ˆ c1 and c2 are two regularization constants, and L is the dynamic range of the pixel values. MSE =

2 N M  ˆ ˆ j=1 i=1 I (i, j) − I (i, j) M×N PSNR = 10 log

SSIM =

L2 MSE

(12b)

(2μx μxˆˆ + c1 )(2σx,xˆˆ + c2 ) (μ2x + μ2ˆ + c1 )(σx2 + σ ˆ2 + c2 ) xˆ

(12a)

xˆ

(13)

9 Can the Expansion of Prediction Errors be Counterbalanced … Table 1 Performance comparison for USC-SIPI images Images Jung’s scheme [4] Our scheme Kbits BPP PSNR SSIM Kbits BPP Airplane Baboon Barbara Boat Elaine Lake Lena Peppers Tiffany Zelda Average

29.53 14.21 27.16 25.64 24.08 27.39 33.48 30.76 24.79 33.64 27.07

0.11 0.05 0.10 0.10 0.09 0.10 0.13 0.12 0.09 0.13 0.10

50.83 50.27 50.68 50.64 50.60 50.71 50.90 50.79 50.81 50.88 50.71

0.9996 0.9996 0.9994 0.9994 0.9994 0.9995 0.9992 0.9993 0.9993 0.9992 0.9994

55.67 17.83 33.88 33.62 30.27 34.68 40.55 38.77 30.83 40.42 35.65

0.21 0.07 0.13 0.13 0.12 0.13 0.15 0.15 0.12 0.15 0.14

105

PSNR

SSIM

53.99 53.09 53.30 53.20 53.26 53.30 53.48 53.34 53.40 53.48 53.38

0.9994 0.9998 0.9996 0.9996 0.9996 0.9996 0.9995 0.9995 0.9995 0.9995 0.9996

Overall embedding rate-distortion performance of our scheme is evaluated and presented in Table 1 for the test images. Our scheme demonstrates an average PSNR (dB) value of 53.38 with embedding rate (BPP) of 0.14, whereas Jung’s scheme has image quality of 50.71 (dB) for an average embedding rate of 0.10. We observe that with higher number of pixels in an input image, the number of counterbalanced pixels in second phase of embedding becomes higher, which eventually improves the embedded image quality and embedding rate. Additionally, to visualize the trend of overall performance of our scheme, embedding rate-distortion curve is presented in Fig. 1. We observe that our scheme has a trend to improve embedded image quality over the higher embedding rates. This is because, higher rate of embedding requires higher number of pixels to be embedded. Once the number of pixels becomes higher, the sizes of X min and X max in the second phase of embedding tend to be higher resulting more counterbalanced pixels and better embedded image quality. This trend of improvement is also evidenced by the average performance in Fig. 1. Example of embedded images and their decoded versions are illustrated in Fig. 2. A decoded image, by definition of our decoding principle, should be identical to the input image, which is verified for all the test images and can be roughly witnessed with the given examples in this paper. In summary, considering overall rate-distortion performance, our proposed RDH scheme outperforms its baseline scheme as illustrated in Table 2. An average of 31% higher embedding rate and 5% better PSNR are obtained for the SIPI test images.

106

H. Nyeem and S. A. Hasib Boat

Airplane Ours Jung (2017)

PSNR (dB)

PSNR (dB)

Ours Jung (2017)

70

70

65

60

65

60

55 55 0

1

2

3

4

5

0

6

1

104

Embedding capacity (bits)

3

4

Embedding capacity (bits)

2

104

Elaine

Lena 70

Ours Jung (2017)

Ours Jung (2017)

70

PSNR (dB)

PSNR (dB)

65 65

60

60

55

55

0

1

2

3

4

0

5

1

104

Embedding capacity (bits)

3

4

Embedding capacity (bits)

2

104

Average

Peppers 70

65

PSNR (dB)

PSNR (dB)

70

Ours Jung (2017)

60

Ours Jung (2017)

65

60 55

55 0

1

2

3

Embedding capacity (bits)

4

5 104

0

0.5

1

1.5

2

2.5

Embedding capacity (bits)

Fig. 1 Overall embedding rate-distortion performance comparison with the Jung’s scheme

3 104

9 Can the Expansion of Prediction Errors be Counterbalanced …

(a) embedded (ours)

(b) embedded (ours)

(c) embedded (Jung)

(d) embedded (Jung)

(e) decoded

(f) decoded

107

Fig. 2 Example of embedded and decoded versions of test images of size 512 × 512 × 8 for our scheme and Jung’s scheme: a, c, e Airplane and b, d, f Tiffany. Decoded images in e, f are identical to respective original input images and are thus the same for the both schemes (Original test images are from USC-SIPI database [21])

108

H. Nyeem and S. A. Hasib

Table 2 Performance improvement of our scheme over the Jung’s scheme BPP%

Airplane Baboon Barbara Boat 89

PSNR% 6

Elaine

Lake

Lena

Peppers Tiffany

Zelda

Average

25

25

31

26

27

21

26

24

20

31

6

5

5

5

5

5

5

5

5

5

5 Conclusions In this paper, we defined backward embedding with PEE and investigated its use to counterbalance the expansion made in forward embedding with PEE in developing a new PVO-based RDH scheme. With two phases embedding, our RDH scheme first applies classic PVO-based PEE on each non-overlapping image block of size 1 × 3. The second phase embedding with backward PEE is designed to partially restore the pixels predicted in the first phase to their original values. This counterbalance of expansion is verified to achieve better embedded image quality and higher embedding capacity. Experimental results demonstrated a promising performance of our proposed scheme and its significant improvement over the baseline RDH scheme. Better embedding rate-distortion performance, particularly at higher embedding rates mean to have more potential for the applications that usually require high embedding capacity like electronic patient record hiding in medical images. In addition to the study of our proposed scheme in specific application scenarios, future investigation on the backward PEE for its generalized RDH framework and information theoretic analysis may create a new avenue in data hiding research.

References 1. Alattar AM (2004) Reversible watermark using the difference expansion of a generalized integer transform. IEEE Trans Image Process 13:1147–1156 2. Cao X, Du L, Wei X, Meng D, Guo X (2016) High capacity reversible data hiding in encrypted images by patch-level sparse representation. IEEE Trans Cybern 46(5):1132–1143 3. Hasib SA, Nyeem H (2017) Developing a pixel value ordering based reversible data hiding scheme. In: Proceedings of EICT’17. IEEE 4. Jung KH (2017) A high-capacity reversible data hiding scheme based on sorting and prediction in digital images. Multimed Tools Appl 76(11):13127–13137 5. Kamstra L, Heijmans HJ (2005) Reversible data embedding into images using wavelet techniques and sorting. IEEE Trans Image Process 14:2082–2090 6. Khan A, Siddiqa A, Munib S, Malik SA (2014) A recent survey of reversible watermarking techniques. Inf Sci 279:251–272 7. Kim KS, Lee MJ, Lee HY, Lee HK (2009) Reversible data hiding exploiting spatial correlation between sub-sampled images. Pattern Recognit 42:3083–3096 8. Lee CC, Wu HC, Tsai CS, Chu YP (2008) Adaptive lossless steganographic scheme with centralized difference expansion. Pattern Recognit 41(6):2097–2106 9. Li X, Yang B, Zeng T (2011) Efficient reversible watermarking based on adaptive predictionerror expansion and pixel selection. IEEE Trans Image Process 20:3524–3533

9 Can the Expansion of Prediction Errors be Counterbalanced …

109

10. Li X, Li J, Li B, Yang B (2013) High-fidelity reversible data hiding scheme based on pixelvalue-ordering and prediction-error expansion. Signal Process 93(1):198–205 11. Li X, Zhang W, Gui X, Yang B (2015) Efficient reversible data hiding based on multiple histograms modification. IEEE Trans Inf Forensics Secur 10(9):2016–2027 12. Lin CC, Liu XL, Yuan SM (2015) Reversible data hiding for vq-compressed images based on search-order coding and state-codebook mapping. Inf Sci 293:314–326 13. Ni Z, Shi YQ, Ansari N, Su W (2006) Reversible data hiding. IEEE Trans CSVT 16:354–362 14. Nyeem H, Boles W, Boyd C (2014) Digital image watermarking: its formal model, fundamental properties and possible attacks. EURASIP J Adv Signal Process 2014(1):1–22 15. Ou B, Li X, Zhao Y, Ni R (2014) Reversible data hiding using invariant pixel-value-ordering and prediction-error expansion. Signal Process Image Commun 29(7):760–772 16. Peng F, Li X, Yang B (2014) Improved PVO-based reversible data hiding. Digit Signal Process 25:255–265 17. Qu X, Kim HJ (2015) Pixel-based pixel value ordering predictor for high-fidelity reversible data hiding. Signal Process 111:249–260 18. Shi YQ, Li X, Zhang X, Wu HT, Ma B (2016) Reversible data hiding: advances in the past two decades. IEEE Access 4:3210–3237 19. Thodi DM, Rodríguez J (2007) Expansion embedding techniques for reversible watermarking. IEEE Trans Image Process 16:721–730 20. Tian J (2003) Reversible data embedding using a difference expansion. IEEE Trans CSVT 13:890–896 21. USC-SIPI: Image database. http://sipi.usc.edu/database/, [Online; last accessed 23-Nov-2013] 22. Wahed MA, Nyeem H (2017) Efficient LSB substitution for interpolation based reversible data hiding scheme. In: Proceedings of ICCIT 2017. IEEE 23. Wahed MA, Nyeem H (2018) Reversible data hiding with interpolation and adaptive embedding. Multimed Tools Appl 1–25 24. Wang Z, Bovik AC, Sheikh HR, Simoncelli EP (2004) Image quality assessment: from error visibility to structural similarity. IEEE Trans Image Process 13(4):600–612

Chapter 10

Drowsiness Detection Using Eye-Blink Pattern and Mean Eye Landmarks’ Distance Abdullah Arafat Miah, Mohiuddin Ahmad and Khatuna Zannat Mim

1 Introduction In today’s world, one of the vital reasons behind death is a road accident. Distracted driving is one of the most common cases for a road accident. Drowsiness of the driver is one of the most important reasons behind distracted driving. According to [1] a study was done about the sleep habits in Argentina where they got about 43.7% drivers who slept frequently while driving. A study was done among the lorry drivers in Brazil in [2]. The study shows that 16.4% road accidents had happened due to drowsiness of the drivers. 75% Thai bus or truck drivers experienced drowsiness while driving [3]. To detect drowsiness of a driver and avoid road accidents by cautioning the drowsy drivers, some of the methods had been introduced. The data from the sensor mounted on the steering lever for measuring steering wheel angles is an important tool which is used to sense the drowsiness of the driver [4]. The variation of Electrocardiogram (ECG), Electroencephalogram (EEG), Heart rate (HR) due to drowsiness is studied to detect the drowsiness of drivers. A significant difference can be observed in R peaks and the R-R interval along with the difference between R and S peaks in awake and drowsy state of the driver which can be used for the detection purpose [5]. In [6] Steady State Visually Evoked Potentials (SSVEPs) is computed from the EEG signal. Later on it is used to discriminate between the open and close eye. In [7] the wavelet A. A. Miah (B) · M. Ahmad · K. Z. Mim Department of Electrical & Electronic Engineering, Khulna University of Engineering & Technology, Khulna 9200, Bangladesh e-mail: [email protected] M. Ahmad e-mail: [email protected] K. Z. Mim e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_10

111

112

A. A. Miah et al.

transform of the Heart Rate Variability (HRV) was used for the same purpose. The use of this like ECG, EEG, and HRV shows a better accuracy in comparison with the steering wheel measuring system but the main problem here is the processing time. If the detection system needs more processing time to detect the drowsy driver then it would not be very much useful to prevent road accidents. Another useful tool to detect a drowsy driver is the Local Binary Pattern (LBP) [8]. In [9] IR illuminator is used for the purpose of capturing images of the drivers. Here Percent eye closure (PERCLOS) and numerous different parameters are characterized utilizing fuzzy classifier to identify the driver’s cautiousness progressively. The contribution of the paper is to develop a system which can recognize or detect the drowsy condition of a driver from real-time video monitoring. In this paper, we had introduced a technique where image processing techniques, as well as machine learning algorithms, are utilized to distinguish the drowsiness of the driver. Mainly our algorithm is based on the eye blink pattern of the drivers and also on the vertical eye distance calculation of the drivers. If drowsy then its eye blinking frequency will be low and eye close duration will be high. From the eye blinking pattern, the drowsy or sleeping state of the drivers from their normal state can be differentiated easily. The rest of the paper is sorted out as pursues: In Sect. 2, we present the proposed method to detect a drowsy diver briefly with the details of the algorithm. Section 3 examines the exploratory outcomes. Section 4 finishes up the paper.

2 Proposed Method For detecting the drowsiness of the driver, the total procedure is performed with mainly four steps. They are (i) face and eye detection, (ii) facial landmarks detection, (iii) mean eye landmarks distance calculation, and (iv) blink detection.

2.1 Face and Eye Detection Firstly the face and eye of the driver ought to be distinguished appropriately and precisely. For this detection purpose, Haar cascade classifier is used. This classifier is developed by the Paula Voila and Michael Jones [10]. In this classifier, Haar features are mainly used to detect a face in an image. These features can be in different shapes. They can be horizontal or vertical. Edge and line features are different types of Haar features that are shown in Fig. 1a. Edge features can detect edges effectively. One portion’s pixels of a feature are completely black and another portion’s pixel is completely white. Now for a real face, there is no perfectly black and white region but we can assume that for a grayscale image the eyebrow has comparatively darker pixels then the forehead contains brighter pixels. It is very similar to the edge features. The line feature is combined with the white pixels’ portion then black and again white pixels’ portion.

10 Drowsiness Detection Using Eye-Blink Pattern and Mean Eye …

113

Fig. 1 a The most relevant Haar features of a face, b detected face and eye with haar cascade algorithm

It can be also as first black pixels’ portion then white and again black pixels’ portion. If the nose of a human face is considered then firstly one side of the nose has the darker pixels then the middle portion of the nose has the brighter pixels then again the other side of the nose has the darker pixels. It is very similar to the line feature. If we assign the relevant haar-like features of a human face then it appears in Fig. 1a. So firstly we detect the most relevant features of a face like eyes, lips, nose etc. Then we can detect the face. In ideal Haar features the pixel intensities in the white side are 0 and in the black side is 1 (in the scale of 0–255). But for real life, it will have larger pixel intensities in the darker side and lower pixel intensities in the brighter side. Not perfectly 0 or 1 (in the scale of 0–255). To find a Haar-feature, we sum up the white pixel intensities and average them and do the same for the black pixel intensities. Then delta is calculated which is the difference between the average of the white pixel intensities and the average of the black pixel intensities and compare the value to 1 because for the ideal Haar features the value of delta is 1. The closer the value to 1, the more likely a Haar feature is found. That’s mean we have found an eye or a nose etc. and finally the face is detected as well as the eye. To reduce the computational complexity integral image concept is used. =

1 1 Id (x) − Ib (x) n n

(1)

In Eq. (1) n is the total amount of pixels for a Haar feature where Id (x) and Ib (x) are the pixels intensities of white and black side respectively in image I. Delta () is the difference between the average values for the white and dark parts. Finally, we get a bounding box around the face and eyes as shown in Fig. 1b.

114

A. A. Miah et al.

2.2 Facial Landmarks Detection After detecting the face and eye our further progress is dependent on the different facial features like the corner of the eyes, eyebrows etc. That’s why facial landmarks detection is now necessary. Till now we have a bounding box around the face and eyes. We have to find the (x, y) coordinates of the facial landmarks within the bounding box we achieved in step 1. To find these facial landmarks many algorithms is developed till now. This algorithm can be divided into three categories; they are The Holistic Method, Constrained Local Method, and Regression-Based Methods. The performance of the holistic method and the Constrained Local Method is poor in comparison with the Regression-Based Methods [11]. Among the different types of Regression Methods, we used the Cascaded Regression Methods [12, 13] to obtain our desired facial landmarks. To detect the coordinates of the 68 facial landmarks on driver’s face Dlib library is used. Dlib library is an implementation of the [13]. Tree based regressor is used where an ensemble of regression trees are trained. Then the positions of the facial landmarks are estimated. The indexes of the 68 coordinates we achieved can be visualized from Fig. 2.

2.3 Mean Eye Landmarks’ Distance Calculation In this step, 68 facial landmarks are available on our face. But we just need the portion of the two eyes. So we cropped the portion of the two eyes along with facial landmarks which is shown in Fig. 3. Now from the landmarks of the eyes, the eye landmarks’ distance is calculated. In Fig. 3, the eye landmarks as A, B, C, D, E, F are indicated. Then, the horizontal distances between A, E and B, D are calculated. Then we calculated the vertical

Fig. 2 Facial landmarks detection

10 Drowsiness Detection Using Eye-Blink Pattern and Mean Eye …

115

Fig. 3 Cropped eye image with landmarks

distance between F, C. The distances between these landmarks are calculated from the Eq. (2).  2  D A,E = (A x − E x )2 + A y − E y   2 (2) D B,D = (Bx − Dx )2 + B y − D y    2 D A,E = (Fx − C x )2 + Fy − C y Here Dx,y represents the distance between the two landmarks of x and y. Here, (Ax , Ay ) represents the location of the landmark A. Similarly, locations of landmark are obtained for B, C, D, E, and F. Now after calculating between the horizontal and vertical distances, the eye landmarks distances are calculated. Then, the eye landmarks’ distances for both eyes are calculated by using Eq. (3). L=

D A,E + D B,D D F,C

(3)

Here, L is representing the eye landmarks’ distance. Now the eye landmarks’ distance of the left and right eye is calculated. For the left eye, the distance is l.l and l.r for the right eye. From these data, mean eye landmarks’ distance, L is calculated from Eq. (4). L=

l.r + l.l 2

(4)

2.4 Blink Detection Figure 4 shows that the mean eye landmarks’ distance varies with eye open and closed condition. A threshold distance 0.35 is assumed to differentiate between an open and close eye. The duration of the eye closing state (T) is considered. Table 1 shows the values of the parameters of the various condition of the driver while driving.

116

A. A. Miah et al.

Fig. 4 a Mean eye landmarks’ distance for open eyes, b mean eye landmarks’ distance for closed eyes Table 1 Conditions of the driver with the variation of different parameters

Conditions of driver

Description

Awake

L > 0.35

Normal eye blink

L < 0.35, T3s

Sleeping

L < 0.35, T>>3s

3 Experimental Results We have started our experiment by capturing the video from the webcam of our laptop. We processed each frame of the captured video to determine the duration of the closed eye or to determine the blinking rate as well as the processing time needed to process per frame. A single frame processing is shown in Fig. 5. Figure 6 shows the mean eye landmarks distance, the processing time per frame, and the blinking time. From this data, we have determined whether the driver is in drowsy condition or not. In this figure, mean eye landmarks’ distance versus the elapsed time is shown.

10 Drowsiness Detection Using Eye-Blink Pattern and Mean Eye …

117

Fig. 5 The processed frame of the capturing video

Fig. 6 Eyeblink pattern for the awake or lively state

Figure 7 shows the mean eye landmarks’ distance go under 0.35 several times within 25 s of surveillance. But the elapsed time for the lower mean eye landmarks’ distance is very short. So the eye blinking frequency is high here and the driver is in a normal state. Figure 7a shows the elapsed time during the lower mean eye landmarks’ distance is higher than the previous one. The eye blinking frequency is very low here. So here the driver is drowsy. On the other hand, in Fig. 7b the mean eye landmarks’ distance goes low for a very long time. So the driver’s eye is in a closed state for a very long time. Here the driver is sleeping. Figure 8 shows the processing time needed to process per frame is very low and almost the same for the entire surveillance. Figure 9 shows a typical blink of the driver. From this plot, we can easily determine when the blink started as well as its duration. The drowsiness detection system is running on a personal computer which has a 2.90 GHz and has a RAM of 8 GB. The system is tested on YawDD. YawDD is a video dataset of many drivers having different facial characteristics [14].

118

Fig. 7 Eyeblink pattern for a drowsy condition and b sleeping state

Fig. 8 Processing time for each frame versus elapsed time

Fig. 9 Visualization of a typical blink

A. A. Miah et al.

10 Drowsiness Detection Using Eye-Blink Pattern and Mean Eye …

119

YawDD dataset is divided into two different parts. In the first part, video was monitored of the drivers from the camera beneath the front mirror. And in the second part, the video was monitored from the driver’s dash. For the purpose of testing our results, we used the second dataset which contains about 29 videos having both female and male drivers and having glasses and without glasses. We experimented whether eye blinking pattern of the driver can be determined correctly or not. About 93% accuracy has been achieved by experimenting with the dataset. Table 2 shows the results we achieved in YAWDD dataset. Table 2 shows that most of the time our algorithm is capable to detect the eye blink as well as the duration of each eye blink. This method shows wrong results most of the times for the people with the glasses. For the failure of detecting the eye, the eye blink pattern can’t be determined correctly. Although in most of the cases the eye blink pattern and the duration of the closed eye state is determined correctly. After the measurement of the duration of the eye blink, the final task is generating an alarm for the driver. The alarm continues if the drowsiness is of the driver is not being recovered. When the driver comes to the normal state, the alarm becomes stop. Figure 10 shows sample processed frames from the dataset we have used for the proposed research purpose. We have contrasted our proposed outcome and some other existing techniques in drowsiness detection purpose. The correlation results appear in Table 3.

Table 2 The results achieved in the YAWDD dataset

Video classes

Total no. of videos

Correctly determined videos

Male without glasses

04

04

Male with glasses

08

06

Female without glasses

07

07

Female with glasses

04

04

Fig. 10 Processed frames from the YawDD

120 Table 3 A comparison with the proposed strategy and the other existing techniques

A. A. Miah et al.

References

Drowsiness measure

Precision (%)

[15]

EEG, ECG, EoG

95–97

[16]

ECG

90

[17]

EoG, EMG, EEG

90

[18]

EEG

84

[19]

Pupil

92

[20]

Yawning

91.97

Proposed method

Eye blink pattern

93

4 Conclusion and Future Works Throughout the paper, we introduced a noble automatic technique in order to determine the drowsy or non-drowsy state of the vehicles’ drivers. We showed that from facial landmarks’ position we can calculate the mean eye landmarks’ distance which is useful to determine the eye blinking pattern. from the eye blinking pattern, we can easily determine the drowsy and non-drowsy driver. The whole driver drowsiness detection system is processed in our personal computer. It’s a real-time drowsiness detection system. We tested our system on the YawDD on about 29 drivers and had achieved the accuracy of 93%. Now after classifying the drowsy driver, an alarm system had been developed to warn the driver. In this process, numerous lives from being the victim of the road accident due to the drowsy drivers can be saved. In future, we want to elaborate the system to detect not only the drowsiness but also the other incidents that can lead to a road accident such as talking to mobile phones or the person sitting beside the driver. We want to monitor the hand movements of the drivers in order to check any unusual movements. Acknowledgements For testing the proposed system, we used the YawDD dataset. The Figs. 1, 2, and 4 are not taken from the dataset. These are one of author’s own images taken for testing in different processing steps.

References 1. Perez-Chada D, Videla AJ, O’Flaherty ME, Palermo P, Meoni J, Sarchi MI (2005) Sleep habits and accident risk among truck drivers: a cross-sectional study in Argentina. Sleep: 1103–1108 2. Canani SF, John AB, Raymundi MG, Schonwald S, Menna Barreto SS (2005) Prevalence of sleepiness in a group of Brazilian lorry drivers. Public Health: 925–929 3. Leechawengwongs M, Leechawengwongs E, Sukying C, Udomsubpayakul U (2006) Role of drowsy driving in traffic accidents: a questionnaire survey of Thai commercial bus/truck drivers. Chotmaihet Thangphaet J Med Assoc Thailand 1845–1850 4. Li Z, Li SE, Li R, Cheng B, Shi J (2017) Online detection of driver fatigue using steering wheel angles for real driving conditions. Sensors 17:495

10 Drowsiness Detection Using Eye-Blink Pattern and Mean Eye …

121

5. Takalokastari T, Jung S-J, Lee D-D, Chung W-Y (2011) Real time drowsiness detection by a WSN based wearable ECG measurement system. J Sens Sci Technol 20(6):382–387 6. Hashemi A, Saba V, Resalat SN (2014) Real time driver’s drowsiness detection by processing the EEG signals stimulated with external flickering light. Basic Clin Neurosci 3(1: Winter) 7. Li G, Chung WY (2013) Detection of driver drowsiness using wavelet analysis of heart rate variability and a support vector machine classifier. Sensors 13:16494–16511 8. Huang D, Shan C (2011) Local binary patterns and its application to facial image analysis a survey. IEEE 41:765–781 9. Bergasa L, Nuevo J, Sotelo M, Barea R, Lopez M (2006) Real-time system for monitoring driver vigilance. IEEE Trans Intell Transp Syst 7(1) 10. Viola P, Jones M (2001) Rapid object detection using a boosted cascade of simple features. Proc CVPR 1:511–518 11. Yue W, Qiang J (2018) Facial landmark detection: a literature survey. Int J Comput Vis: 1–28 12. Dollár P, Welinder P, Perona P (2010) Cascaded pose regression. In: CVPR, pp 1078–1085 13. Kazemi V, Sullivan J (2014) One-millisecond face alignment with ensemble of regression trees. In: IEEE conference on computer vision 14. Abtahi S, Omidyeganeh M, Shirmohammadi S, Hariri B (2014) YawDD: a yawning detection dataset. In: Proceedings of ACM multimedia systems, Singapore 15. Khushaba RN, Kodagoda S, Lal S, Dissanayake G (2011) Driver drowsiness classification using fuzzy wavelet-packet-based feature-extraction algorithm. IEEE Trans Biomed Eng 58:121–131 16. Patel M, Lal SKL, Kavanagh D, Rossiter P (2011) Applying neural network analysis on heart rate variability data to assess driver fatigue. Exp Syst Appl 38:7235–7242 17. Hu S, Zheng G (2009) Driver drowsiness detection with eyelid related parameters by support vector machine. Exp Syst Appl 36:7651–7658 18. Liu J, Zhang C, Zheng C (2010) EEG-based estimation of mental fatigue by using KPCA-HMM and complexity parameters. Biomed Signal Process Control 5:124–130 19. Shen W, Sun H, Cheng E, Zhu Q, Li Q (2012) Effective driver fatigue monitoring through pupil detection and yawing analysis in low light level environments. Int J Digit Technol Appl 6:372–383 20. Xiao F, Bao CY, Yan FS (2009) Yawning detection based on Gabor wavelets and LDA. J Beijing Univ Technol 35:409–413

Chapter 11

Routing Protocol Selection for Intelligent Transport System (ITS) of VANET in High Mobility Areas of Bangladesh Md. Kamrul Hasan and Orvila Sarker

1 Introduction Vehicular Ad hoc Networks (VANETs) are an outgoing subclass of Mobile Ad hoc Network (MANETs) where mobile nodes comprise both movable vehicles and permanent infrastructure. VANETs allow for the provision of Intelligent Transportation Systems (ITS) that help to avoid congestion on and to provide safer roads [1]. Network topology in VANETs changes frequently, but the changes are sometimes predictable with vehicle velocity and position partly constrained by roads, traffic congestion, driver behavior, and traffic signals. Building blockage and inference of signal are also included in the challenges for VANETs. This is the subject of our study. The Accident Research Institute (ARI) [2] of Bangladesh University of Engineering and Technology (BUET), Bangladesh, provides a data compilation of road accident from January to March in 2018 which witnessed a total of 789 road accidents in which 1050 individuals were dead and 2015 people were injured. As in the developed countries, Bangladesh still lacks in the section of automated Intelligent Transportation System (ITS). We are aiming to implement VANET practically. As a prerequisite, we need routing protocols for the simulation of VANET so that we can find a best-fitted protocol for the transportation system. In this paper, we evaluate four different routing protocols (AODV, AOMDV, OLSR, and DSDV) in a VANET urban environment of Bangladesh so that we can select the best protocol for VANET. The simulation area is taken from the OPENSTREETMAP which covers the area of 1000 * 1000 m. The performances are analyzed on the principle of Packet Delivery Ratio (PDR), delay, jitter, throughput, and normalized routing overhead. Md. K. Hasan (B) · O. Sarker Comilla University, Cumilla 3506, Bangladesh e-mail: [email protected] O. Sarker e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_11

123

124

Md. K. Hasan and O. Sarker

The results of simulation show that OLSR protocol provides maximum PDR in load network and also performs well in higher network area. DSDV and OLSR show quite similar PDR as compared to others. AODV and AOMDV gain maximum PDR in higher congestion area but provide poor performances in lower load area. In terms of delay, AODV and AOMDV are outperformed by OLSR and DSDV where DSDV takes less time (as well as lower jitter) to transfer packets. The OLSR protocol gains maximum average throughput in low network load but AODV outperforms OLSR and DSDV in higher traffic area. Routing overhead in OLSR protocol is much less than other protocols in both high and low traffic areas except AODV provides best result in higher number of mobile nodes.

2 Related Works In [3], Vyas and Chopra explored and analyzed the performance of AODV and AOMDV routing protocols. In their simulation, two protocols for the measured parameter; packet delivery ratio, throughput, average end-to-end delay, and residual energy are AODV and AOMDV. Their evaluation shows that AOMDV is the best fit for traffic scenario. In [4], Feiz and Movaghar analyzed three routing protocols: AODV, DSR, and DSDV in fully and partially connected VANET. The issue of transient network fragmentation and the issue of broadcast blast also convoluted the design of routing protocols in VANETs. Though DSDV provides lower delay in their simulation, it shows higher routing overhead comparatively in high traffic area. In [5], the authors used a heterogeneous network; they have mentioned the density of vehicle node to improve the performance of OLSR and AODV protocol beneath two scenarios. They do so with fading propagation model for VANETs; however, the realistic MAC protocol was absent in the simulation. Their simulation shows that the packet delivery ratio is improved in OLSR than other but the throughput is still lower in higher mobility area as compared to AODV. In [6], the authors took the real map of US Census Bureau and employ different number of moving nodes up to 120 nodes, and they configured the design with IEEE 802.11p and a realistic fading model that mirror the influence of barrier on radio signal. In [7], Spaho and Barroli present the performance evaluation of AODV, OLSR, and DYMO routing protocols, and they configured a simple propagation model called two ray ground which does not mirror the influence of an urban environment on wireless signal. The simulation outcomes have displayed that DYMO is dissimilar to OLSR which attempts to use more hops to maintain the communication between vehicle nodes. In case of AODV, there is no buffer. In [8], Haerri and Filali et al. present the performance evaluation of VANETs between AODV and OLSR in urban environments with realistic mobility model. They only feature on artificial mobility map and some factors are missed out that

11 Routing Protocol Selection for Intelligent Transport System …

125

they have a direct impact on the performance of network like VANET application traffic and propagation model. In [9], the authors experiment the influence and importance of adequately considering the effects of physical layer to exactly quantify the performance. They investigated the impact of the radio propagation modeling on the performance and operation estimation of unicast and broadcast position-based wireless vehicular ad hoc routing protocols. The results which are obtained from the simulation have shown that if the modeling is not properly done then the radio channel effects can considerably influence not only on the behavior and operation of routing protocols but also on the performance. In [10], Mohapatra and Kanungo have used the random waypoint propagation model for the evaluation of AODV, DSR, OLSR, and DSDV routing protocols for VANET simulation. The simulation shows that DSR protocol gives the best average PDR but it is shown for maximum 50 nodes.

3 Intelligent Transport System: Vehicular Ad Hoc Network (VANET) A VANET is an exceptional kind of ITS that authorize devices (e.g., cellular) within moving vehicles to easily and dynamically receive and transmit information from other moving vehicles without the help of infrastructure devices. For the vehicle, the awareness simplifies more proficient vehicle agility, advanced safety features, and minimizing emersions through more identical speeds [11]. In VANET, vehicles make communication with other vehicles through V2I or V2V.

3.1 Vehicle-to-Vehicle (V2V) Communication Vehicle-to-Vehicle (V2V) communication is a technology which permits vehicles to communicate with one another in a network [12]. V2V organizes actively safety which comprises vehicle stability system, collision avoidance system, automated emergency braking system, and driver warning system. Dedicated Short Range Communication (DSRC) technology is responsible for establishing V2V communication. This technology can provide information of roadway and particular information about other vehicles which is also equipped with the same technology. The main challenging section of V2V technology for the application is the simulation and filtering of the receiving information, and it also handles the volume of receiving data and bogus warnings.

126

Md. K. Hasan and O. Sarker

3.2 Vehicle-to-Infrastructure (V2I) Communication Vehicle-to-Infrastructure (V2I) is a communication model which permits moving vehicles to share information with the components that facilitate the highway system of a country [13]. Those components comprise cameras, streetlights, parking meters, RFID readers, traffic lights, streetlights, and lane markers. The Road Side Units (RSUs) periodically broadcast vehicles speed limit message to determine whether a speed limit alarm engages to any vehicles in the vicinity. A warning broadcast message will be distributed to the vehicle if a vehicle oversteps the speed limit so that driver reduces the speed of the vehicle.

4 Design Methodology The block diagram of VANET simulation steps is shown in Fig. 1.

Start

Extract map from OpenStreetMap

Create Route file with no. of mobile nodes

Run Net convert Map.net.xml

Mobility Pattern

TraNS

Run SUMO

Map.rou.xml

Map.sumo.cfg

Create TCL with required parameters and mobility pattern

NS2 for Simulation NS2 generates NAM & trace file

Analyze trace file using Perl and AWK script

Stop Fig. 1 Flowchart of VANET simulation and analysis of protocols

Routing protocols AODV, AOMDV, OLSR, DSDV

11 Routing Protocol Selection for Intelligent Transport System …

127

4.1 OSM File Importing We can easily select the area of the map which is to be downloaded from the “www.openstreet.org” website [14]. If we want to exclude some paths to simplify the vehicles network, we can easily use Java OpenStreetMap (JOSM) editor which is used for adding or removing to and from the map.

4.2 Creating Network File We have used Ubuntu operating system where the OSM file is converted to network file by using the command “netconvert”. This command creates a network (Fig. 2a) in SUMO [15]. A network file consists of edge ID which defines the unique ID of every lane.

4.3 Configuration of Network Before configuration, we have employed the number of vehicles in the route which works as the mobile nodes in the network. A route file consists of route ID and vehicle departure ID. Then, the configuration of the network is made of combining the network file and route file. Figure 2b shows the movement of mobile nodes, i.e., vehicles. If we want to employ different number of nodes in the route, we will have to create separate route file for each configuration of the network.

(a) Zoomed small portion of Paltan area map Fig. 2 Paltan area map view using SUMO

(b)Vehicles movement in a small section of Paltan area.

128

Md. K. Hasan and O. Sarker

4.4 Mobility The command “Python” runs a script, named as “traceExporter.py”, and creates the mobility of the vehicles along with the TCL file of the network.

4.5 TraNS Traffic and Network Simulation Environment (TraNS) is a Graphical User Interface (GUI) tool that integrates traffic and network simulators (SUMO and NS2) to propagate realistic simulations of VANETs. The wireless TCL file is integrated with mobility TCL file in this section. The required parameters (shown in Table 1) are concluded in the scripts together with routing protocols (AODV, AOMDV, OLSR, and DSDV) for the simulation in NS2.

4.6 Simulation in NS2 The integrated TCL scripts are run into NS2 simulator [16] which generates the Network AniMator (NAM) scripts and the trace scripts. The NAM scripts show the graphical representation (Fig. 3) of the mobile nodes communicating with each other. The trace file is used to analyze the performance of different routing protocols in different scenarios.

Table 1 The parameters used in the VANET simulation

Communication model

MAC, 802.11 g

Connection type

CBR, TCP

Signal propagation model

Two ray ground reflection

Mobility model

Random waypoint

Antenna model

Omnidirectional

Nodes type

Mobile nodes

Node speed

10 m/s

Node scenarios

50, 100, 150

Simulation area

1000 * 1000 m2

Simulation duration

150 s

Layer type

Link layer

Interface queue type

Drop tail/priority queue

Interface queue length

50 packets

11 Routing Protocol Selection for Intelligent Transport System …

129

Fig. 3 Communication between nodes in Network AniMator in selected path

5 Performance Metrics 5.1 Packet Delivery Ratio Packet Delivery Ratio (PDR) is the ratio of total packets received at destination and total packets sent from the source. The PDR is defined by Eq. (1). PDR =

Total Packets Received Total Packets Sent

(1)

5.2 Throughput Throughput defines how fast a mobile node can actually transfer packets through a network. Throughput is calculated by the following Eq. (2): Throughput (bits/s) =

No. of delivered packets ∗ Pkt size ∗ 8 Total Simulation Period

(2)

130

Md. K. Hasan and O. Sarker

5.3 End-to-End Delay In most cases, End-to-End Delay (EED) is measured as the sum of processing delay (PD), queueing time (QT), transmission time, (TT) and propagation time (PT). EED is calculated using (3). EED = PD + QT + TT + PT

(3)

5.4 Jitter Jitter (generally measured in milliseconds) is the variation in time delay between data packets communicating over a network. The packets roam in same intervals over a strong network. This will permit the destination computer for processing the data. These intervals between packets become broke up when there is a state with jitter.

5.5 Routing Overhead Routing protocols produce tiny sized packets called routing packets to retain updated information about network paths. An instance of such kind of packets is a hello packet which is used to supervise whether the node in neighborhood is active. Routing packets carry only route information where data packets carry data. In most of the times, both data and routing packets have to take part the same network bandwidth and the routing packets are taken into account being an overhead in the network and it is called routing overhead. A good routing protocol should tie lower routing overhead. The Routing Overhead Ratio (ROR) is calculated as in (4). ROR =

Routing Packets Data PAckets

(4)

6 Results and Discussion 6.1 Packet Delivery Ratio (PDR) Figure 4 shows the graphical representation of PDR comparison where we can observe that OLSR protocol has achieved the maximum PDR than other routing protocols in lower and higher number of mobile nodes. DSDV also shows a performance which is very close to OLSR. AODV and AOMDV result in a good PDR in higher load but present poor performance in lower load.

11 Routing Protocol Selection for Intelligent Transport System … Fig. 4 Comparison of PDR among routing protocols

131

50

0.9917 0.9933 0.992 0.9873

0.9725 0.9898 0.9917 0.9889

1 0.98 0.96 0.94 0.92 0.9 0.88 0.86

0.9346 0.9121 0.9816 0.9725

PDR

Packet Delivery Ratio AODV PDR AOMDV PDR OLSR PDR DSDV PDR

100

150

NETWORK LOAD

6.2 End-to-End Delay Figure 5 shows the graphical representation of comparison among the routing protocols: AODV, AOMDV, OLSR, and DSDV. It shows that the average end-to-end delay is much lower in DSDV protocol than other routing protocols. The reason behind lower EED is the lower number of packets transmission. OLSR also provides lower delay as compared to AODV and AOMDV.

6.3 Jitter Analysis Jitter of AODV. As we have mentioned earlier that AODV protocol takes longer time for packet transmission that means it also has the longer jitter. Figure 6 shows the graphs for jitter w.r.t packet ID where we can observe that variation occurs in packets (ACK and TCP) time interval. Jitter of AOMDV. As compared to AODV, the jitter in AOMDV is much lower but there is huge variation in the packet interval as shown in Fig. 7. Jitter of OLSR. The average end-to-end delay in OLSR is much lower, i.e., much lower jitter than AODV and AOMDV. Figure 8 shows two network load scenarios for jitter variation in OLSR protocol in VANET simulation.

Fig. 5 Comparison of average end-to-end delay among routing protocols

292.07

194 175.06 149.58

200

284.05

300

126.84 130.78 123.12

EED (MS)

400

276.94 230.4 105.88 101.12

AVERAGE EED

AOMDV OLSR DSDV

100 0

AODV

50

100

NETWORK LOAD

150

132

Md. K. Hasan and O. Sarker

TCP

TCP

ACK

ACK

(a) 100 nodes

(b) 150 nodes

Fig. 6 Jitter between packets for AODV protocol TCP

TCP

ACK

ACK

(a) 100 nodes

(b) 150 nodes

Fig. 7 Jitter between packets for AOMDV protocol

TCP

TCP

ACK

ACK

(a) 100 nodes

Fig. 8 Jitter between packets for OLSR protocol

(b) 150 nodes

11 Routing Protocol Selection for Intelligent Transport System …

133

TCP

TCP

ACK

ACK

(a) 100 nodes

(b) 150 nodes

Fig. 9 Jitter between packets for DSDV protocol

Jitter of DSDV. In Fig. 9, the jitter in DSDV protocol is shown. Assimilating with OLSR, DSDV gives the lower jitter in higher network load.

6.4 Throughput

Fig. 10 Comparison of throughput among AODV, AOMDV, OLSR, and DSDV

300 200 100 0

50

134.76 202.83 196.63 182.27

400

78.12 47.98 175.66 150.88

THROUGHPUT

THROUGHPUT

370.77 350.8 229.25 182.06

We can observe from Fig. 10 that AODV gains the highest throughput in terms of high network load but provide poor throughput performance in low load network. AOMDV also results in higher data transfer as compared to OLSR and DSDV but poor in lower congestion. When it comes to OLSR, the throughput gradually increases with raising the mobile nodes. During the session of 50 mobile nodes, OLSR outperforms other protocols. Figure 10 shows the graphical representation of throughput.

AODV AOMDV OLSR DSDV

100

NETWORK LOAD

150

134

Md. K. Hasan and O. Sarker

Fig. 11 Comparison of routing overhead ratio among routing protocols

16.486

Routing Overhead Ratio Comparison 20

AODV-ROR

50

0.2 2.867 0.736 3.008

0

AOMDV-ROR

2.276 3.665 1.661 1.909

5

2.244 2.443

10

2.618

ROR

15

100

150

OSLR-ROR DSDV-ROR

NETWORK LOAD

6.5 Routing Overhead Analysis Figure 11 shows the comparison of Routing Overhead Ratio (ROR) among routing protocols. We can observe that OLSR protocol provides a better result than other protocols. AODV results in lower ROR in 150 nodes scenario, but if we want to select a protocol for VANET then OLSR is the best fit because the ROR in other two scenarios (50 and 100 nodes) for OLSR is the lowest assimilating to AODV, AOMDV, and DSDV.

7 Conclusion In this work, VANET simulation is performed in an urban area of Bangladesh with the routing protocols: AODV, AOMDV, OLSR, and DSDV. These protocols are applied with varying number of network sizes with a constant node speed. The simulation is carried out using NS2 and SUMO where the simulation area is selected from OPENSTREETMAP. With all of these parameters, we have evaluated the five performance measures, i.e., Packet Delivery Ratio (PDR), End-to-End Delay (EED), Jitter, and Throughput and Routing Overhead Ratio (ROR). In terms of PDR, OLSR protocol gives the best result in network sizes. AOMDV (99.33%) surpasses OLSR (99.2%) in higher load but it gives poor result (91.21%) in small network size. DSDV outperforms AODV and AOMDV in terms of end-to-end delay as it takes only 149.58 ms in 150 node scenario. But OLSR also results in lower delay (105.88 ms, 130.78 ms, and 175.06 ms for 50, 100, and 150 nodes scenario, respectively) which is very close as in DSDV in all network size. When the delay is small, the transmission duration between packets, i.e., jitter is also small. The OLSR protocol gains maximum average of 175.66 kbps in terms of throughput in low network load but AODV (370.77 kbps) outperforms OLSR and DSDV in higher congestion area. If throughput is prime criteria, AODV and AOMDV are better solutions in high mobility network but OLSR is best fit for low mobility network because it out sailed other protocols. If we only consider control/routing overhead for high mobility area then AODV is supreme choice for VANET implementation because it gives lower ROR as 0.2. But if we consider the combination of both small

11 Routing Protocol Selection for Intelligent Transport System …

135

and high mobility area, OLSR is the best protocol to implement. As the network size increases (50–150 nodes) for OLSR protocol, the ROR gradually decreases (2.244–0.736) more than other three protocols. Overall, the performance of OLSR and DSDV protocol is much better than AODV and AOMDV. If we have to select a protocol on the basis of overall performances, then OLSR is the best-fitted protocol for ITS of VANET. In future, some hybrid protocols (e.g., TORA, ZRP, etc.) will be implemented in this simulation and we are working to produce a hybrid protocol by the aggregation of OLSR and DSDV.

References 1. Zeadally S, Hunt R, Chen Y-S, Irwin A, Hassan A (2010) Vehicular ad-hoc networks (VANETS): status, results, and challenges. Telecommun Syst 50(4):217–241 2. https://www.ari.buet.ac.bd. Last accessed 07 May 2018 3. Vyas U, Chopra K, Lakkadwala P (2016) Performance enhancement of routing protocols for VANET with variable traffic scenario. IJRTER 2:332–338 4. Feiz SM, Movaghar A (2011) Comparison of DSR, AODV and DSDV routing protocols in fully and partially connected VANET. In: Proceedings of 2011 world congress in computer science, computer engineering and applied computing, Las Vegas, Nevada, 18–21, July 5. Zuo J, Wang Y, Liu Y, Zhang Y (2010) Performance evaluation of routing protocol in VANET with vehicle-node density. In: 2010 6th international conference on wireless communications networking and mobile computing WiCOM 6. Khan T, Qayyum A (2009) Performance evaluation of AODV and OLSR in highly fading vehicular ad hoc network environments. In: INMIC 2009–2009 IEEE 13th international multitopic conference, pp 1–5 7. Spaho E, Barolli L, Mino G, Xhafa F, Kolici V, Miho R (2010) Performance evaluation of AODV, OLSR and DYMO protocols for vehicular networks using CAVENET. In: 2010 13th international conference on network based information system, pp 527–534 8. Haerri J, Filali F, Bonnet C (2006) Performance comparison of AODV and OLSR in VANETs urban environments under realistic mobility patterns. In: Proceedings of 5th IFIP mediterranean ad-hoc networking workshop, no 1, pp 14–17 9. Bauza R, Gozalvez J, Sepulcre M (2008) Operation and performance of vehicular ad-hoc routing protocols in realistic environments. In: Proceedings of the 2nd IEEE international symposium on wireless vehicular communications (WiVeC), Calgary (Canada), pp 1–5 10. Mohapatra S, Kanungo P (2012) Performance analysis of AODV, DSR, OLSR and DSDV routing protocols using NS2 simulator. In: ELSEVIER international conference on communication technology and system design, pp 69–76 11. https://www.quora.com/What-is-the-purpose-of-VANET. Last accessed 03 June 2018 12. Shah K, Parentela EM (2015) A case study on potential benefits of V2V communication technology on freeway safety 13. Dey KC, Rayamajhi A, Chowdhury M, Bhavsar P, Martin J (2016) Vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) communication in a heterogeneous wireless network—performance evaluation. Transp Res Part C 68:168–184 14. Haklay M, Weber P (2008) OpenStreetMap: user-generated street maps. IEEE Pervas Comput 15. Behrisch M, Bieker L, Erdmann J, Krajzewicz D (2011) SUMO—simulation of urban mobility—an overview. In: Proceedings of 3rd international conference on advances in system simulation, no c, pp 63–68 16. Issariyakul T, Hossain E (2012) Introduction to network simulator 2 (NS2). In: Introduction to network simulator NS2. Springer, Boston, MA

Chapter 12

An Intelligent Children Healthcare System by Using Ensemble Technique Nishargo Nigar and Linkon Chowdhury

1 Introduction Health care is one of the most vital issues present in human life. But it is a matter of regret that people do not care enough about maintaining their health until they suffer from serious diseases. Various types of digital healthcare services are becoming more common these days. Technologies like mHealth are aiding in such processes. An astonishing fact is that patients are being treated in a smart home or even in smart car environments at present. Digital healthcare services offer more accuracy, real-time facilities, emergency care and obviously cost reduction [1]. Implementing IoT in healthcare application is not a new practice. We have noticed the trend of smart healthcare [2] already. More examples include monitoring of elderly in a remote environment [3], lightweight mHealth platform with biometrics [4] and wearable systems [5]. Designing such systems involves a lot of risks and threats. Still, researchers have found their ways to dive into the practical execution of such incredible healthcare services. The necessary strategy of health care for children is different from adults. They need to be monitored appropriately when they are developing their physical and mental capabilities. The job is unquestionably difficult, but not impossible if the health tutors and parents can facilitate the supervision of a child’s smart phone activities. In addition, it is quite difficult to interest children with foods that are high in fiber, vitamins and nutrition value. The main challenge is whether it is possible for them to stick to a digital solution for their own betterment. Utilizing edutainment for N. Nigar (B) Department of Computer Science and Engineering, East Delta University, Chittagong, Bangladesh e-mail: [email protected] L. Chowdhury School of Science, Engineering and Technology, East Delta University, Chittagong, Bangladesh e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_12

137

138

N. Nigar and L. Chowdhury

spreading the knowledge of appropriate diet plan, nutrition charts, health habits, and serving sizes can be highly beneficial. If we keep checking our progress on a regular basis, it is possible to gain the usage pattern of current users. In our system, it is really important for us to collect data, analyze them and evaluate them for further outcomes in a collaborative process. As a result, it should be easy and automatic for the end users. Our study involves a three-party environment of children, parents, and health tutors. Child users will be assessed under the expertise of an authorized healthcare center. Parents will be able to receive progress in a timely manner through a constant notification generation module. We have created different user interfaces according to different types of profiles. Mobile app and web interface are joined together in order to maintain a smooth communication system. Components like wearables, modern games, and recommended nutrition knowledge lessons are used to take input from child users. Finally, after collecting the required data, we have compared two ensemble techniques called bagging and boosting to find out the better accuracy. In this way, the authentic health status of a participating child can be detected. We believe our healthcare management system offers a costeffective healthcare service for children. With the growing issue of childhood obesity and malnutrition, it can be of huge help for parents. While a lot of things can be gained by using digital technologies in the field of this study, the focus has been centered on the following objectives. • To predict the healthcare status of a participating child or future participants of the system. • To monitor a child’s mental, physical, and behavioral abilities through several steps and procedures. • To inspire children about the nutritional value of various food items and why they are important for health. This study is divided into four parts. Section 2 is about literature review. Section 3 discusses the model that is developed to make our children healthcare system. The result is elaborated in Sect. 4, before rounding off our study with conclusion and future work in Sect. 5.

2 Literature Review 2.1 Importance of mHealth for Children Preventing diseases through mobile phones and wireless technology is the primary goal of mHealth services. Health applications offer such services for tablet PCs and smartphones. Certainly, mHealth technologies are transforming the traditional ways of healthcare management and bringing greater outcomes. mHealth services may include patient monitoring, appointment reminders, surveillance, creating awareness

12 An Intelligent Children Healthcare System …

139

over diseases, consultation and so on. As the environment around us is changing on a constant basis, it is quite troublesome to create effective solutions for children [6]. Children use smart phones mostly for playing games, watching cartoons and animated movies. If parents and health instructors can supervise their activities, we can apply the practice towards better care and disease management. Introducing mHealth to children can make them aware of increasing diseases and teach the significance of proper nutrition. In [7], children who are d j (x); wher e i = j f or i, j = 1, 2, 3, . . . L

(5)

Then the decision boundary of separating two classes i and j can be specified as follows: di (x) − d j (x) = 0.

(6)

The simplest method is to design a decision boundary perpendicular to the line that connects the mean of the class. The classes are considered as the cluster centroids in our proposed method. For L number of classes, the decision function can be defined as follows: 1 di (x) = x T m i − m i m iT ; 2

f or i = 1, 2, 3, . . . L .

(7)

where m i is the mean vector of the ith class. Therefore, an unknown instance x is said to belong to the ith class if, upon substitution of x into all decision functions, di (x) yields the largest numerical value.

2.6 Output Segment After classification and noise detection, output processed in this part to a speckle-free clean image. The noise variance of ultrasound may differ from image to image. If the image is marked as clean it will skip the LDA analysis and go directly to output as shown in Fig. 5a. If the classification result finds the image as a noisy one it will be said noisy like Fig. 5b. The system starts to proceed next for LDA analysis and separate the noise signal from the image signal. Figure 5c has shown the speckle-free image after speckle reduction.

16 Computer-Aided Speckle Noise Analysis in Ultrasound Images …

(a)

191

(b)

Noisy Image

Clean Image

(c)

Speckle free Image after LDA analysis

Fig. 5 Output segment: a The image is classified as clean image b The image is classified as noisy image c Speckle reduction using LDA

3 Results and Evaluations The efficiency of computer-aided (CAD) medical video/image data analysis mostly depends on feature selection. Standard data availability is another important issue for this purpose. Thus, the performance of any CAD depends on the training dataset. To make the result more reliable, the proposed system utilizes more than 4000 standard ultrasound image sets besides its own datasets. Most of the data have been collected from MNI data sets (http://www.bic.mni.mcgill.ca/~laurence/data/data. html). Another important source of data set is (http://splab.cz/en/download/databaze/ ultrasound). The proposed system has been evaluated against standard dataset to assess its reliability. Also, the proposed system has been tested against human expert’s consulted dataset to assess its applicability in real life. Around 14,000 images are collected from the various training classifier, among which, one-third images are noisy and rest are clean. In this experiment, the whole dataset is split into training and test dataset. The features are extracted from training dataset and classified based on those features. Features from test dataset are extracted and passed through the classifier. The classification is performed through the fusion of CNN and wavelet features then compared with the existing without fusion method. The comparison is given in Table 1.

192

R. Mostafiz et al.

Table 1 Comparison between existing without fusion and proposed fusion Method

Accuracy (%)

Sensitivity (%)

Without fusion

90.71

91.11

Specificity (%) 80.54

Proposed system (with fusion)

98.54

98.19

98.25

For medical data classification, sensitivity (true positive rate) and specificity (true negative rate) are sometimes more reliable than accuracy (rate of successful detection). For this system, the following measures are calculated as shown in Fig. 6. From Fig. 6 and information above, it is observed that the proposed model of wavelet features and convolutional neural network features give much satisfactory outcome. A comparison among different speckle filtering methods [17, 20, 21] are showed in Table 2. To compute the diagnostic accuracy of proposed method performances have been compared with the diagnostic efficacy of human experts. Though there is no alternative to the human brain, several human factors lead to speckle miss-classification in

Fig. 6 Confusion matrix of our system

Table 2 Comparison of our method with other methods

True Positives

False Negatives

0.9819

0.0135

False Positives

True Negatives

0.0157

0.9825

Method

Ultrasound image SNR

MSE

EPF

Min

15.0981

0.5110

0.2791

Max

16.0204

0.4991

0.2540

Median

14.4053

0.4120

0.3109

Hard threshold

16.2246

0.4616

0.3824

Soft threshold

16.3009

0.4512

0.3549

Bayesian threshold

16.5682

0.4254

0.3871

Proposed system

16.9343

0.3933

0.4797

16 Computer-Aided Speckle Noise Analysis in Ultrasound Images …

193

Fig. 7 Probability density curve of the original and filtered image

open eyes. Computer-aided noise detection system can reduce the miss-classification rate especially in such cases. Without the help of computer-aided parameter, it is difficult to detect speckles. The Probability density curves of original image volume and filter image volume are shown in Fig. 7, for analysis. The change between the input image and the output image is almost minimum in probability density curve. Thus, it is clear that using the proposed method a little amount of information is lost and a very small amount of error may occur in the final speckle-free image. Therefore, the proposed system performs more reliable outcomes in terms of edge preservation and smoothness.

4 Conclusions and Future Works The proposed computer-aided system is capable of supporting the medical decision for speckle detection and reduction of the ultrasound image. Automatic speckle denoising will enhance the ability of the physicians to accurately detect the lesions which may go undetected and involve in malignant diseases. Proper features selection and classification are more important in this intelligent system. In this paper, we have combined the strength of wavelet features and the power of convolutional neural network features. Fusions of these two features are used to classify image in the proposed automatic system. The ultrasound image is the input of this system which produces the output as a clean image. The proposed system shows a greater accuracy from the ROC analysis and performs better noise reduction than the existing ones.

194

R. Mostafiz et al.

In the future, the proposed method can be used for 3D volume analysis of ultrasound image. Acknowledgements We are very grateful to Dr. Md. Farhan Matin, Associate Professor, Department of Radiology and Imaging, Uttara Adhunik Medical College & Hospital; for his valuable support, suggestions, and consultancy.

References 1. Vishwa A, Sharma S (2012) Speckle noise reduction in ultrasound images by wavelet thresholding. Int J Adv Res Comput Sci Softw Eng 2(2) 2. Sudha S, Suresh GR, Sukanesh R (2009) Speckle noise reduction in ultrasound images by wavelet thresholding based on weighted variance. Int J Comput Theory Eng 1(1):7 3. Cunningham RJ, Harding PJ, Loram ID (2017) The application of deep convolutional neural networks to ultrasound for modelling of dynamic states within human skeletal muscle. arXiv preprint arXiv:1706.09450 4. Wang P, Zhang H, Patel VM (2017) SAR image despeckling using a convolutional neural network. IEEE Signal Process Lett 24(12):1763–1767 5. Chierchia G, Cozzolino D, Poggi G, Verdoliva L (2017, July) SAR image despeckling through convolutional neural networks. In: 2017 IEEE international conference on geoscience and remote sensing symposium (IGARSS). IEEE, pp 5438–5441 6. Danilla C (2017) Convolutional neural networks for contextual denoising and classification of SAR images 7. Zhang K, Zuo W, Zhang L (2018) FFDNet: toward a fast and flexible solution for CNN based image denoising. IEEE Trans Image Process 8. Koziarski M, Cyganek B (2016, Sept) Deep neural image denoising. In: International conference on computer vision and graphics. Springer, Cham, pp 163–173 9. Foucher S, Beaulieu M, Dahmane M, Cavayas F (2017, July) Deep speckle noise filtering. In: 2017 IEEE International conference on geoscience and remote sensing symposium (IGARSS). IEEE, pp 5311–5314 10. Wang G, Wang G, Pan Z, Zhang Z (2017, Nov) Multiplicative noise removal using deep CNN denoiser prior. In: 2017 international symposium on intelligent signal processing and communication systems (ISPACS). IEEE, pp 1–6 11. Yao C, Cheng G (2016) Approximative Bayes optimality linear discriminant analysis for Chinese handwriting character recognition. Neurocomputing 207:346–353 12. Fan Z, Xu Y, Ni M, Fang X, Zhang D (2016) Individualized learning for improving kernel Fisher discriminant analysis. Pattern Recogn 58:100–109 13. Min HK, Hou Y, Park S, Song I (2016) A computationally efficient scheme for feature extraction with kernel discriminant analysis. Pattern Recogn 50:45–55 14. Mostafiz R, Rahman MM, Mithun Kumar PK, Islam MA (2017) Speckle noise reduction for 3D ultrasound images by optimum threshold parameter estimation of wavelet coefficients using Fisher discriminant analysis. Int J Imaging Robot™ 17(4):73–88 15. Mostafiz R, Rahman MM, Kumar PM, Islam MA (2018) Speckle noise reduction for 3D ultrasound images by optimum threshold parameter estimation of bi-dimensional empirical mode decomposition using Fisher discriminant analysis. Int J Signal Imaging Syst Eng 11(2):93–101 16. Azim G, Abo-Eleneen Z (2011) Thresholding based on Fisher linear discriminant. J Pattern Recognit Res 6(2):326–334 17. Ashique RH, Kayes MI (2013) Speckle noise reduction from medical ultrasound images—a comparative study. IOSR-JEEE 7(1)

16 Computer-Aided Speckle Noise Analysis in Ultrasound Images …

195

18. Karkanis SA, Iakovidis DK, Maroulis DE, Karras DA, Tzivras M (2003) Computer-aided tumor detection in endoscopic video using color wavelet features. IEEE Trans Inf Technol Biomed 7(3):141–152 19. Mia S, Rahman MM (2018) An efficient image segmentation method based on linear discriminant analysis and K-means algorithm with automatically splitting and merging clusters. Int J Imaging Robot 18(1):62–72 20. Zong X, Laine AF, Geiser EA (1998) Speckle reduction and contrast enhancement of echocardiograms via multiscale nonlinear processing. IEEE Trans Med Imaging 17(4):532–540 21. Benzarti F, Amiri H (2013) Speckle noise reduction in medical ultrasound images. arXiv preprint arXiv:1305.1344

Chapter 17

A Dynamic Bandwidth Allocation Algorithm for Gigabit Passive Optical Network for Reducing Packet Delay and Bit Error Rate Md. Hayder Ali and Mohammad Hanif Ali

1 Introduction The demand of data communication growth rate in telecommunication sector has been reducing the importance of outmoded cable line connection technologies like DSL and wire modem. These types of traditional technologies are not resourcefully sufficient to encounter the clients’ request for high-bandwidth demands. GPON (Gigabit Passive Optical Networks) is the recent innovation in the universe. It provides a total solution and support, including voice communication (TDM), faster internet service for Ethernet, etc. It maintains robust procedures for operation, direction, preservation, and the OAM&P abilities subscription for endwise provision administration. The GPON technology is not only providing significantly sophisticated competence as a carriage network, but also transports straightforwardness and wonderful scalability for upcoming development in subsidiary supplementary facilities. In order to increase the performance of GPON system, latency and bit error rate is the most important factor. For reducing the transportation duration (latency) and provision capability for multiple provisions for circulation forecast, an improved DBA is projected, it will reduce the packet delay and bit error rate (BER). The other portion of this paper is organized like this. The literature review is presented in Sect. 2. The dynamic bandwidth allocation algorithm procedure is presented in Segment 3. In part 4, P-DBA system is shortly designated. In part 5, Projected DBA algorithm is enlightened. Bit error rate (BER) calculation is described at Sect. 6. Md. Hayder Ali (B) · M. Hanif Ali Department of Computer Science and Engineering, Jahangirnagar University, Dhaka, Bangladesh e-mail: [email protected] M. Hanif Ali e-mail: [email protected]

© Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_17

197

198

Md. Hayder Ali and M. Hanif Ali

Simulation scenario is in Sect. 7 and performance analysis between C-DBA, P-DBA and Proposed DBA is stated in Sect. 8. Finally, Sect. 9 provides a conclusion to this paper.

2 Literature Review The OLT, in GPON system, is liable for ONT upbound traffic distribution, and it is a dominant concern to allocate the traffic more realistic [1]. Ozimkiewicz et al. [2] clarify that GPON contains two types of traffic-sharing approaches: static and dynamic. Qi-Yu et al. [3] explain a procedure which guarantees the higher urgency and equality of the procedure, but unable to assign the traffic compliantly conferring to the actual complaint. Consequently, the irrational condition might seem that portions of urgencies surplus the additional traffic. Smith et al. [4] has explain an AR-DBA, OLT uses modified AR model to predict all traffic flows, and ONUs need not to report their requests. In this proposed AR-DBA, it is hard to realize the effect of accurate prediction and reducing delay. Dixit et al. [5] stated that C-DBA is a procedure depending on T-CONT, and it is assigning a static traffic to T-CONT-1. It guarantees the requirement of high-priority, but easy to waste bandwidth in low load, leading to a poor performance of low-priority, so it cannot meet the fairness requirement. Technically, it is a better presentation to alter the extended variety dependence into a dumpy variety requirement with wavelet and expect the circulation, the difficulty of the procedure is amplified concurrently. Drakulic et al. [6] projected P-DBA. A provision could be done by it for multiple facilities, and its performance is very much enhanced compared to other DBAs. But it is unable to forecast circulation decoration and requested bandwidth distribution. De Lutiis et al. [7] stated a distribution procedure that generates no interruption modification for continuous bit rate circulation. It is on the basis of packet interruption modification of fixed bit rate circulation in extended-reach GPONs. Silva et a1. [8] proposed three diversities of procedure that are associated underneath equally symmetric and asymmetric circulation situations, they simulated for concerning normal packet postponement for numerous urgencies, interruption variance for fixed bit rate (FBR) circulation and bandwidth consumption. These algorithms are based on EPON and multipoint control protocol (MPCP). Colle et al. [9] offered a novel vibrant bandwidth distribution algorithm, for provision diversity that meets provision-level settlements of the operators. This algorithm can work for providing expectable normal packet suspension, extraordinary and average urgency traffic, decrease the interval difference and for protection of the packet loss rate under check. It is also based on EPON and it is stated for the maximum-priority and the lower priority circulation class, respectively, which progress the jitter for both class traffic.

17 A Dynamic Bandwidth Allocation Algorithm …

199

3 Dynamic Bandwidth Allocation In the GPON system network, the OLT is informed by the ONTs about the upside traffic distribution by communicating bandwidth mapping messages (BWMAP), it is constructed of several traffic distributions for the specific ONTs or the ONT Transmission Containers (T-CONT). Individually, traffic distribution is a gesture to an ONT to communicate in a distinct time period. The spirit of DBA is dynamically calculating the BWMAP to allocate the correct bandwidth for each ONT. As an example, Fig. 1 defines a situation where bandwidth is not used by ONTs 1 and 3 is assigned to other ONTs that request it. Each and every ONT has its individual slot for transmitting its packet. In Fig. 1 it is shown that as ONT-2 needs high priority traffic, ONT-3 leaves its third slot and ONT-1 leaves its fourth slot for ONT-2. By the same time, ONT-2 leaves its fourth slot for ONT-3. By this mechanism, ONT sets priority traffic to end users.

Fig. 1 Bandwidth allocation as per ONT/ONU’s requirement

200

Md. Hayder Ali and M. Hanif Ali

4 P-DBA System DBA is used to enhance the up-link utilization, for the purpose of improving network performance. In P-DBA procedure, ONUs/ONTs direct appeal frame to OLT for letting it know the magnitude of data reserved in Optical Network Terminal stream cushion. Conferring to the appeal data traffic, Optical Line Terminal appraises the traffic. Later on, during execution the analysis of the traffic distribution, Optical Line Terminal allocates cushion to ONUs/Optical Network Terminals to assign the upside link spell period [3] (Fig. 2). Procedure 1. Pseudocode for P-DBA 1. 2. 3. 4. 5. 6. 7. 8. 9. 10.

ONUi /ONTi directs REPORT communication to OLT at period t1 ONUi /ONTi desires OLT for GATE communication at period t2 ONUi /ONTi intervals for Allocating traffic from Optical Line Terminal Optical Line Terminal appraises the required traffic OLT allows the Frame to ONUi /ONTi seeing the demanded magnitude of the frame for an uplink side time period ONUi /ONTi directs traffic to OLT Recurrence of task 1–6 If step 5 flops, then Recurrence task 1–4 End

In this P-Dynamic Bandwidth Allocation model, there is no option for data storage; data and queue size calculation system are also absent. P-DBA cannot predict the priority traffic.

Fig. 2 P-DBA polling tackle plan

17 A Dynamic Bandwidth Allocation Algorithm …

201

5 Proposed DBA Model DBA system must provide multiple facilities and assurance QoS for urgencies primarily, then to guarantee the equality of all transmitted bandwidth, it must consider all necessary requirements into attention for transmission, and then improve the bandwidth consumption to decrease unnecessary time slot idle at the last. Projected DBA includes all data deposited in the queue and also maintains the incoming and anticipating time, it should meaningfully be declining anticipating period of the traffic and improving the actual period presentation and decrease the container damage frequency (Fig. 3). Procedure 2. Convert program for proposed Dynamic Bandwidth Allocation 1. Start 2. Optical Network Units/ONTs analyze the quantity of acknowledged data. 3. Optical Network Units/ONTs forecast the magnitude of traffic in upcoming period. 4. Optical Network Units/ONTs analyze the bandwidth for T-CONTs (T-CONT1 to T-CONT4). 5. Optical Network Units/Optical Network terminals increase the forecast of corT ) for nth sequence. rectness by scheming incoming in the upcoming period (Pi,n 6. Optical Line Terminal analyzes the mathematical calculation for the traffic acknowledged in upcoming period. 7. Recurrence task 2–6 for nth sequence. 8. After that 9. Analyze the variance for the amount of traffic acknowledged in the upcoming period at ONTi /ONUi in the nthe sequence. 10. Analyze the extent of bandwidth forecast incoming in the upcoming period at ONTi /ONUi in the nth cycle by calculating an allowance feature. 11. Upsurge the usual instruction to guarantee the forecast accurateness.

Fig. 3 Proposed DBA polling tackle plan

202

Md. Hayder Ali and M. Hanif Ali

12. Optical Line Terminal accepts traffic volume or assign the traffic volume for urgency bandwidth. 13. Optical Network Units/ONTs direct data conferring to Optical Line Terminal’s importance allocate traffic. 14. Finish The final variance among P-DBA model and the projected DBA scheme is that OLT allocates priority traffic and need based maximum traffic at grant frame. The proposed DBA model could calculate the data storage size and also it can predict the priority traffic by calculating the queue size, as it maintains the queue, it could measure the priority traffic. This way it minimizes the transmission delay. ONT/ONU direct appeal frame to OLT for urgency bandwidth and necessity-based extreme bandwidth by computing data size and examining bandwidth decoration.

6 Bit Error Rate (BER) Calculation for Proposed DBA Model Let B the bit rate signifies, then the possibility that n photo electron are acknowledged throughout a bit intermission 1/B is specified by  e−( p/ h fc B)

p h fc B

n

n!

.

Therefore, the opportunity of not getting any photoelectrons is e−( p/ h fc B) . Assuming equally likely ls and 0s, the bit error rate of this ideal receiver would be given as BE R =

1 − h fp B e c 2

Let M = P/hf c B. The constraint M signifies the regular amount of photoelectrons acknowledged throughout a 1 bit. Formerly, the bit error rate could be stated as BE R =

1 −M e 2

Appearance signifies the error rate of a perfect consumer and is stated the quantum limit. Here, k B is Boltzmann’s constant and has the value of 1.38 × 10−23 J/0 K and δ(τ ) is the ∞ Dirac delta function, defined as δ(τ ) = 0, τ =0 and −∞ δ(τ )dτ = 1. If Be is the bandwidth of the receiver, this current can be demonstrated as a Gaussian random process with mean I¯ and variance

17 A Dynamic Bandwidth Allocation Algorithm …

203

2 2 σ 2 = σshot + σther mal .

For a conveyed 1 bit, let the acknowledged optical power P = P1 and let the mean photocurrent be I¯ = I 1 . For the circumstance when 1 and 0 bits are likewise expected (which is the individual circumstance we reflect in this scenario), the edge photon current is assumed roughly by σ0 I1 + σ1 I0 σ0 + σ1  ∞ 1 2 Q(x) = √ e−y /2 dy. 2π −∞ Ith =

It now follows that  P[0|1] = Q

I1 − Ith σ1



 and P[1|0] = Q

Ith − I0 σ0



Using (3.8), the BER is given by 

I1 − I0 BE R = Q σ0 + σ1



Threshold setting yields a higher bit error rate given by      1 I1 − I0 I1 − I0 BE R = +Q Q 2 2σ1 2σ0

7 Simulation Design The simulation is done by using OptSIM (RSoft System Suit, version-2016.06) Model software. The imitation setup is shown in Fig. 4. At the central office, there is an OLT and after that, at some distance, there is a 1:32 splitter to split the optical signal. After that, there are 32 optical network terminals at the end user side. Considered 1.25 Gbps is for uplink traffic and 2.48 Gbps is for downlink traffic.

8 Performance Analysis The data has been captured from OptSIM simulator and analyzed by MATLAB. Delay measurement and bit error rate (BER) are calculated by MATLAB.

204

Md. Hayder Ali and M. Hanif Ali

Fig. 4 Simulation scenario in OptSIM simulator

Figure 5 stated that the number of users is more in the proposed DBA algorithm. Considering 2GBPS bandwidth allocation, C-DBA is allowing 17 users, P-DBA is allowing 20 users, while the proposed DBA is allowing 23 users. Bandwidth is shared with maximum users, which satisfy the properties of the GPON system. Figure 6 shows the number of errors versus total transmitted bit. It stated that proposed BDA has a smaller number of errors compared to C-DBA and P-DBA. Considering 107 bit transmission, C-DBA has 76 errors and P-DBA has 71 errors while the proposed DBA has only 65 errors. Figure 7 stated that delay per user is minimum in the proposed DBA. Considering 25 users in the system, C-DBA shows 3 ms/user delay, P-DBA shows 2.4 ms/user delay while proposed BDA shows only 1.7 ms/user. The per-user delay is less in the proposed DBA and it increases the system performance.

9 Conclusion DBA is a significant concern in GPON system. A better BDA could increase the system performance. Packet delay and bit error rate (BER) is also an important factor for system performance. In the proposed DBA algorithm, ONU/ONT predicts the

17 A Dynamic Bandwidth Allocation Algorithm …

205

Fig. 5 Bandwidth allocation versus number of users

Fig. 6 Bandwidth allocation versus number of users

data size and sends allocation request to OLT, OLT increases the prediction accuracy. It could calculate the queue size and store the data size. By calculating the queue size, it can predict the priority traffic. It transmits the priority traffic and this way it could minimize the transmission delay. The projected DBA algorithm has less delay and minimum bit error compared to other DBAs and it fulfils the total service of QoS necessities of the GPON system.

206

Md. Hayder Ali and M. Hanif Ali

Fig. 7 Bandwidth allocation versus number of users

References 1. ITU-T Recommendation G.984.1 (2003), 2 (2003), 3 (2004), 5 (2006) and 6 (2007) 2. Ozimkiewicz J, Dittmann L, Ruepp S, Wessing H, Smolorz S (2010) Dynamic bandwidth allocation in GPON networks. In: Proceedings of the 4th WSEAS international conference on circuits, systems, signal and telecommunications (CISST 2010), pp 182–187 3. Z. Qi-yu, W. Run-ze, and L. Bin:A Dynamic Bandwidth Allocation Scheme for GPON Based on Traffic Prediction. In: Proceedings of the 9th International Conference on Fuzzy Systems and Knowledge Discovery, (FSKD 2012), pp. 2043 – 2046 (2012) 4. Hinton K, Smith TG, Tucker RS, Tran AV (2009) Packet delay variance and bandwidth allocation algorithms for extended-reach GPON. IEEE Commun Mag 47(11):20–24 5. Dixit A, Das G, Lannoo B, Colle D, Pickavet M, Demeester P (2013) Dynamic bandwidth allocation with SLA awareness for QoS in ethernet passive optical networks. IEEE/OSA J Opt Commun Netw 5(3):240–253 6. Drakuliˇc S, Tornatore M, Verticale G (2012) Degradation attacks on passive optical networks. In: Proceedings of the 16th international conference on optical network design and modelling (ONDM 2012), pp 1–6 7. De Lutiis P, Costa L, D’Amico R (2010) Managing emerging nga security. J Telecommun Manag 2(4):302–309 8. Hajduczenia M, Silva HD, Inacio PM, Freire M, Monteiro P (2007) EPON security issues. IEEE Commun Surv 9(1):68–83 9. Onn H, Amir S (2008) The importance of dynamic bandwidth allocation in GPON networks. PMC-Sierra. 1(2):31–35

Chapter 18

Feature Selection and Biomedical Signal Classification Using Minimum Redundancy Maximum Relevance and Artificial Neural Network Md. Masud Rana and Kawsar Ahmed

Cancer is the unrestrained growth of irregular cells in the body and is a foremost death reason over the world. Recently, a number of studies are going on for cancer classification from gene expression data. In contrast to the traditional method, this paper proposes a minimum redundancy maximum relevance (mRMR)-artificial neural network (ANN) method to classify cancer. For this classification, the first step is to select prognostic genes. These tiny subset genes provide better classification accuracy. We proposed mRMR approach for picking more informative genes from microarray data sets. After the selection of genes, artificial neural network is used for training and cancer classification. The proposed mRMR- based ANN approach has been verified on a suite of benchmark data sets of various cancers. Numerical results show that proposed method outperforms compared with the existing methods.

1 Introduction Cancer is the main root cause of sickness among human deaths in many developed countries. Unfortunately, it is very hard to diagnose and classify cancers. There are many methods for these purposes. To begin with, the particle swarm optimization and probabilistic neural network are widely used for classify cancers [1–3]. Moreover, the support vector machine (SVM) and fuzzy neural network are used to classify the particular cancers in [4, 5]. Furthermore, the nonparallel plane proximal classifier is proposed in [6]. The K-nearest neighboring and belief propagation algorithms are adopted in [7–9]. Additionally, the extreme learning machine is used for cancer microarray gene expression data [10]. Recently, the artificial bee colony based Md. Masud Rana (B) · K. Ahmed Department of Electronics and Communication Engineering, Khulna University of Engineering & Technology, Khulna, Bangladesh e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_18

207

208

Md. Masud Rana and K. Ahmed

optimization algorithm is proposed for cancer classification [11, 12]. Indeed, minimum redundancy maximum relevance (mRMR) based artificial bee colony scheme is proposed for cancer classification [13]. Furthermore, the multistage neural network (MSNN) is presented in [14], where backpropagation with Kalman filtering algorithm is used for training the MSNN. Moreover, an oriented feature selection based support vector machine for cancer prediction using the gene expression data is presented in [15]. In addition, the convolutional neural networks is adopted for object classifications [16]. Also, the K-means clustering and discrete wavelet transform are used for image segmentation in [17]. Interestingly, the capsule networks are designed for brain tumor classification in [18]. Combining the genetic algorithm, SVM and mutual information, a method called computer-aided approach is proposed in [19]. This method provides better classification accuracy compared with the existing approaches. The idea is then extended in [20], where the genetic algorithm, artificial neural network and adaptive neuro-fuzzy inherent system are used to improve the classification accuracy. Intuitively, it can be seen that the MRMR-based artificial neural network (ANN) method to classify cancer is not available in the literature. This paper proposes a MRRR-ANN method to classify cancers. The main contributions of this paper are as follows: • Proposes a minimum redundancy maximum relevance based artificial neural network algorithm for classifying cancers. • Extensive simulation results show that the proposed method provides significant performance improvement compared with the existing methods. The rest of this paper is organized as follows. Section 2 describes the proposed method which follows simulation results. Finally, conclusions are made at the end of the paper.

2 Proposed method Figure 1 shows the flow diagram of the proposed method. It can be seen that the proposed algorithm entails three main steps: data normalization phase, gene selection phase, and classification phase. These steps are described as follows.

Fig. 1 Process of the proposed method

18 Feature Selection and Biomedical Signal Classification Using Minimum …

209

2.1 Data Normalization Process Data normalization is a process in which information attributes within a data model are prepared to increase the cohesion of entity types. It can be computed as follows:

xinor =

xi − xmin , xmax − xmin

(1)

where xi is the ith cancer data, x nor is the normalized value, and xmax and xmin are the maximum and minimum values.

2.2 Informative Gene Selection Using Minimum Redundancy Maximum Relevance The maximum-relevance selection criteria is obtained considering the strongest correlation between classification variables [1]. The minimum redundancy is computed considering far away from each variable. In other words, the minimum redundancy maximum relevance (mRMR) is obtained considering far way variables while maintaining the strongest correlation. Mathematically, the maximum relevance can be computed as follows [1]:

max [D(S; c); D] =

1  I (xinor ; c), |S| nor xi

(2)

∈S

where S is the feature set from xinor , c is the target class, and I (xinor ; c) is the mutual information. Sometimes, the feature is selected considering maximum relevance property which contains high redundancy. So, the minimum redundancy can be computed as follows: min [R(S), R] =

1 |S|2



I (xinor ; x nor j ).

(3)

xinor ,x nor j ∈S

Considering aforementioned (2) and (3) with operator φ(D, R), the minimum redundancy maximum relevance is determined as follows: m R M R = max φ(D, R), φ = D − R.

(4)

In summary, if there are many features, the number of informative feature set has been selected using mRMR approach which is used for training the ANN.

210

Md. Masud Rana and K. Ahmed

Fig. 2 mRMR-based ANN for cancer classification

2.3 Classification Using Artificial Neural Network Figure 2 shows the proposed mRMR-based ANN for cancer classification [21]. The selected feature vectors from the mRMR are the input of the ANN. From this diagram, it can be seen that using the set of weighting factors between hidden layer and output layer, the selected feature is refine. The process is an iterative way until the desired accuracy is obtained. Finally, the error output is computed between the desired and estimated signals. Using the weighting factors, the ANN can try to minimize the error so that signal can be classified perfectly. The performance of the proposed algorithm is demonstrated in the following section.

3 Simulation Results and Discussions In this simulation, we are going to use binary microarray cancer datasets as shown in Table 1 [22–24]. Binary class datasets are ALL/AML, Lung Cancer, Prostate Cancer, Ovarian Cancer, Colon1, Colon2, and Human Liver Cancer. The multi-class datasets are Small Round Blue Cell Cancer, Lymphoma, and Mixed Lineage Leukemia [22–24, 29]. Using mRMR scheme, a set of informative feature is selected. For neural network, the value of alpha and batch size is set to 1. Number of neurons at the input layer is set according to the number of selected features. Number of neurons at the output

18 Feature Selection and Biomedical Signal Classification Using Minimum … Table 1 Dataset description Name of disease No. of classes ALL/AML [22] Lung [22] Prostate [22] Ovarian [22] Colon1 [22] SRBCT [23] Liver [25] Lymphoma [26] Colon2[27] MLL[22] DLBCL [24] Breast [28]

2 2 2 2 2 4 2 3 2 3 14 2

No. of Genes

No. of Samples

12.533 12.533 12.600 15.154 2000 2308 1648 4026 7,457 12,582 16063 24481

181 181 136 253 62 83 156 62 36 72 190 97

211

Fig. 3 Variation of no. of genes and hidden neurons for prostate cancer dataset

layer is set according to the number of class of the problem set. Number of neurons at the hidden layer is varying for better classification accuracy. Sigmoid function is used as activation function. Network is trained over 500 iterations. It is observed the effect of different parameters of neural network for achieving better classification results. The number of hidden neurons and the number of genes are varied to observe their effect on the network. From Figs. 3, 4 and 5, it can be seen that the classification accuracy varies with number of genes and number of hidden neurons. For prostate cancer dataset, when number of genes are 400, top testing accuracy is achieved with 200 number of hidden neurons. For SRBCT cancer dataset

212

Md. Masud Rana and K. Ahmed

Fig. 4 Variation of no. of genes and hidden neurons for SRBCT cancer dataset

Fig. 5 Variation of no. of genes and hidden neurons for Colon2 cancer dataset

Fig. 6 Variation effect of no. of iteration

18 Feature Selection and Biomedical Signal Classification Using Minimum …

213

when number of genes are 500, top testing accuracy is achieved with 50 number of hidden neurons. Similarly for Colon2 cancer dataset, when number of genes are 100, top testing accuracy is achieved with 20 number of hidden neurons. From Fig. 6, it is seen that classification accuracy varies with number of iterations. For ovarian cancer dataset classification, accuracy is increased with the number of iteration. When the number of iteration is 900, top testing accuracy is achieved with the fixed number of genes and hidden neuron. For MLL cancer dataset, when number of iteration is 100 top testing accuracy is achieved. Similarly for lung cancer dataset, classification accuracy is increased with the number of iterations. When number of iteration is 300, top classification accuracy is achieved with the fixed number of genes and hidden neurons.

4 Conclusions and Future work We have proposed mRMR-ANN method for classify cancer with “gene microarray expression analysis”. Simulation results show that the proposed algorithm achieves superior performance improvement compared with the existing algorithms. It also requires less time to get accurate results. In future, we will apply our proposed scheme to another human disease for solving classification problem. Acknowledgements The first author is written the paper and the second author did simulations under supervision of first author.

References 1. Peng H, Long F, Ding C (2005) Feature selection based on mutual information criteria of max-dependency, max-relevance, and min-redundancy. IEEE Trans Pattern Anal Mach Intell 27(8):1226–1238 2. Sahu B, Mishra D (2012) A novel feature selection algorithm using particle swarm optimization for cancer microarray data. Proc Eng 38:27–31 3. Hu Y, Ashenayi K, Veltri R, O’Dowd G, Miller G, Hurst R, Bonner R (1994) A comparison of neural network and fuzzy c-means methods in bladder cancer cell classification. In: 1994 IEEE international conference on neural networks. IEEE world congress on computational intelligence, vol 6. IEEE, pp 3461–3466 4. Rajeswari P, Reena GS (2011) Human liver cancer classification using microarray gene expression data. Int J Comput Appl 34(6):25–37 5. Xu R, Cai X, Wunsch DC (2006) Gene expression data for DLBCL cancer survival prediction with a combination of machine learning technologies. In: Annual international conference of the engineering in medicine and biology society, pp 894–897 6. Ghorai S, Mukherjee A, Sengupta S, Dutta PK (2011) Cancer classification from gene expression data by NPPC ensemble. IEEE/ACM Trans Comput Biol Bioinf 8(3):659–671 7. Alhamidi MR, Wasito I (2017) Improved microarray images cancer classification using Knearest neighbor with canonical particle swarm optimization. In: International workshop on big data and information security, pp 37–42

214

Md. Masud Rana and K. Ahmed

8. Tiwari AK, Srivastava R (2015) Feature based classification of nuclear receptors and their subfamilies using fuzzy K-nearest neighbor. In: International conference on advances in computer engineering and applications, pp 24–28 9. Dash JK, Sahoo L (2012) Wavelet based features of circular scan lines for mammographic mass classification. In: International conference on recent advances in information technology, pp 58–61 10. Kumar CA, Ramakrishnan S (2014) Binary classification of cancer microarray gene expression data using extreme learning machines. In: International conference on computational intelligence and computing research, pp 1–4 11. Rajaguru H, Prabhakar SK (2017) Oral cancer classification from hybrid ABC-PSO and Bayesian LDA. In: 2017 2nd International conference on communication and electronics systems (ICCES), pp 230–233 12. Agrawal V, Chandra S (2015) Feature selection using artificial bee colony algorithm for medical image classification. In: International conference on contemporary computing, pp 171–176 13. Alshamlan H, Badr G, Alohali Y (2015) mRMR-ABC: a hybrid gene selection algorithm for cancer classification using microarray gene expression profiling. BioMed research international, vol 2015 14. Zheng B, Qian W, Clarke LP (1994) Multistage neural network for pattern recognition in mammogram screening. In: World congress on computational intelligence and neural networks, vol 6, pp 3437–3441 15. Shen Y, Wu C, Liu C, Wu Y, Xiong N (2018) Oriented feature selection svm applied to cancer prediction in precision medicine. In: IEEE Access, vol 6, pp 48 510–48 521 16. Bardou D, Zhang K, Ahmad SM (2018) Classification of breast cancer based on histology images using convolutional neural networks. In: IEEE Access, vol 6, pp 24 680–24 693 17. Karthiga N (2018) Automated diagnosis of breast cancer using wavelet based entropy features. In: Automated diagnosis of breast cancer using wavelet based entropy features 18. Afshar P, Mohammadi A, Plataniotis KN (2018) Brain tumor type classification via capsule networks. arXiv:1802.10200 19. Salama MS, Eltrass AS, Elkamchouchi HM (2018) An improved approach for computer-aided diagnosis of breast cancer in digital mammography. In: International symposium on medical measurements and applications, pp 1–5 20. Bilalovi´c O, Avdagi´c Z Robust breast cancer classification based on GA optimized ANN and ANFIS-voting structures 21. Ilkucar M, Isik AH, Cifci A (2014) Classification of breast cancer data with harmony search and back propagation based artificial neural network. In: Signal processing and communications applications conference, pp 762–765 22. Kent ridge bio-medical dataset. http://datam.i2r.astar.edu.sg/data-stet/krbd/. Accessed 20 Apr 2016 23. Dataset. http://bioinformatics.rutgers.edu/Static/Supplements/Comp-Cancer/datasets. Accessed 20 Apr 2016 24. Dataset. http://bioinformatics.rutgers.edu/Static/Supplements/Com-Cancer/datasets. Accessed 20 Apr 2016 25. Gene expression patterns in human liver cancers. http://genome-www.stanford.edu/hcc. Accessed 20 Apr 2016 26. “Distinct types of diffuse large [b-cell.” 27. Notterman DA, Alon U, Sierk AJ, Levine AJ (2001) Transcriptional gene expression profiles of colorectal adenoma, adenocarcinoma, and normal tissue examined by oligonucleotide arrays. Cancer Res 61(7):3124–3130 28. Dataset. http://csse.szu.edu.cn/staff/zhuzx/Datasets. Accessed 20 Apr 2016 29. Akhand MAH, Miah Asaduzzaman Md, Hussain Kabir M, Rahman, Hafizur MM (2016) Cancer classification from DNA microarray data using mRMR and artificial neural network. In: 2nd international conference on engineering, technologies, and social

Chapter 19

An Identity-Based Encryption Scheme for Data Security in Fog Computing Nishat Farjana, Shanto Roy, Md. Julkar Nayeen Mahi and Md Whaiduzzaman

1 Introduction In January 2014, Cisco introduced the world with a new computing paradigm named “Fog Computing" in order to bring cloud computing facilities to the edge of the network [6]. Fog Computing is also known as “edge computing" or ‘fogging"; which is a decentralized computing infrastructure in which several networkings, compute and storage are distributed in the most efficient place between the end devices and cloud computing data centers. People are adopting fog computing due to several characteristics; such as proximity to end users, fault tolerance, real-time interactions, and heterogeneity [7]. It reduces the load on the cloud and provides a faster paradigm with fewer network hops. The goal of fogging is to improve efficiency and reduce the amount of data to the cloud. Figure 1 represents the basic architecture of fog computing paradigm. With the increasing volume of data throughout the world, system hacking and data theft attacks have also become one of the major threats in the computing world. These attacks are difficult to address when the attacker is a malicious insider. Nowadays, social networking sites are a great platform for the attackers to steal the user’s personal and necessary information to conduct such type of attacks [3]. Fog computing is a new emerging technology and an interesting area for the malicious attackers and N. Farjana (B) · S. Roy · Md. J. N. Mahi · Md Whaiduzzaman Institute of Information Technology, Jahangirnagar University, Savar, Dhaka 1342, Bangladesh e-mail: [email protected] S. Roy e-mail: [email protected] Md. J. N. Mahi e-mail: [email protected] Md Whaiduzzaman e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_19

215

216

N. Farjana et al.

Fig. 1 Fog paradigm in between edge and cloud computing

hackers. So, our goal is to provide data security through HIBAF scheme which provides four-layered security in the system. Identity-Based Cryptography (IBC) is a popular method where any identifiable information can be used as a public key that provides a certificate-free approach [5, 21]. Our proposed HIBAF scheme for fog computing reduces the cost of generation of keys by reducing the overload of a root Private Key Generator (PKG) and replicating the root PKG to slave PKGs in each level. This scheme also overcomes the burden of using Public Key Directories (PKDs). Another advantage of HIBAF is that it reduces unnecessary delay and provides local processing. Unlike cloud applications, there are limited security certifications and measurements defined for fog computing. Traditional fog computing paradigm uses decoy technology [19] and Ciphertext- Policy-Attribute-Based Encryption (CP-ABE) [2], which return unnecessary delay and bandwidth consumption. A symmetric key-based encryption Advanced Encryption Scheme (AES) is applied in [20] where there is a risk of leakage of key. For cloud computing, a three-level HIBE hierarchy is implemented in [11]. To manage the higher amount of data, the firewalls in the cloud needs to be very big. Moreover, it is very difficult to process such high volume of data and latency and transportation cost is also very high in cloud. There are still some problems including system hacking and unauthorized data access in fog computing. Therefore, all these challenges motivated us to implement a four-level hierarchy of HIBE where fog lowers the focused loads on the cloud and provides faster failover and response during data transmission. We have proposed a HIBAF architecture for fog computing based on encryption and decryption methods with small key sizes. Our proposed scheme is comparatively more lightweight key management procedure than the existing ones. The primary objectives of our work are as follows: • To provide data security in the fog computing paradigm through IBE scheme and prevent the data theft attacks. • To discuss about related recent works done in this area in order to find out the limitations and research gaps.

19 An Identity-Based Encryption Scheme for Data Security in Fog Computing

217

• To implement different types and sizes of datasets to measure the efficiency of our proposed scheme. • To compare our results with other cryptography models to compare the complexity, performance, and security. The rest of the paper is organized as follows: Sect. 2 presents an in detailed literature survey and constructive feedback. Then, Sect. 3 illustrates the proposed methodology with necessary mathematical model, Sect. 4 discusses about security analysis, Sect. 5 analyses the performance, Sect. 6 discusses the advantages, limitations, applications, and future work of the proposed scheme. Finally, Sect. 7 concludes the paper.

2 Literature Review There are comparatively fewer works have been done in the security of fog computing. In [1], the authors have investigated and elaborated the fog computing motivations and demonstrated a Software-Defined Network (SDN)-based three-level architecture. In [18], the authors proposed a five-layer architecture of smart gateway-based communication in fog computing by integrating Internet of Things (IoT) with cloud computing. In [22], the authors proposed an automated software scripting model on Cloud Service Provider (CSP). In [12], the authors investigated and analyzed some security breaches and threats against cloud. In [23], the authors proposed an efficient Mobile Device-Based Cloudlet Resource Enhancement (MobiCoRE) model using the Birth-Death Markov Chain. In [13], the authors proposed a Secure Zone Routing Protocol (SZRP) in a MANET environment. Data security is one of the most confidential parts to provide in a fog or cloud-based environment. In [19], the researchers proposed an offensive decoy information technology “fog computing" to secure data in the cloud. In [10], the authors proposed an SDN-based security infrastructure to provide security in a combined fog-to-cloud environment. In [17], the authors proposed a secure data sharing scheme based on proxy re-encryption method. In [8], the authors proposed a combined elliptic curve cryptography and decoy technology to limit the data theft attacks in a fog platform. As far as we know, there have been fewer works that have applied IBC to cloud/grid computing. However, a symmetric key cryptography named Advanced Encryption System (AES) has been applied in fog computing [20]. Another cryptographical approach, Attribute-Based Encryption (ABE) has been implemented in fog computing in [2, 9] where the authors proposed a Ciphertext-Policy-Attribute-Based Encryption(CP-ABE) and digital signatures to establish a secure communication in fog. Except for cryptosystem, there is another security approach named tokenization discussed in [14]; but the approach needs a token vault management system which is quite costly in fog environments. However, IBC has been applied in cloud computing in [11, 15]. In [11], the authors proposed an Identity-Based Encryption (IBE) and an Identity-Based Signature (IBS) in a Hierarchical Cloud Computing Architecture (HACC). In [15], the authors proposed an identity-based method in cloud

218

N. Farjana et al.

computing. They considered the service URL as the ID as a public key of the user. Most of the recent works have used decoy technology [19] and CP-ABE [2, 9] approaches to provide security in fog communication. However, these approaches often return unnecessary memory utilization and delay to the system.

3 Methodology 3.1 Identity-Based Cryptography The concept of IBC was first introduced and proposed by Adi Samir in 1984 in [16]. In this paper, the authors proposed an asymmetric cryptography model in which the user’s unique identity is used as the public key. The more the organizations are moving toward to fog, the more the security in this area has been gaining importance. Therefore, our goal is to provide a secure communication through authorized users to secure their data through IBE. In IBE, we use the unique identity of the user as a public key. This unique identity can be the user’s name, mobile number, email address, fingerprint, etc. It can be anything that uniquely identifies the user. But this identity has to be converted into binary {0,1} string. Any string converted into binary can be used as the identity of the user. IBC scheme involves four steps. 1. Setup Phase: In this phase, a set of parameters are shared through a trusted Public Key Generation Server. This parameter set is public. And a master key is generated which is private and is only known to the Public Key Generation Center (PKG). Setup −→ ( params, M K )

(1)

2. Extraction Phase: In this phase, the PKG generates the secret key S K i of user i using the I Di of the user and the master secret key MK. Here, I D ∈ {0, 1}∗ . S K i ←− E xtract ( params, I Di , M K )

(2)

3. Encryption Phase: In this phase, the sender encrypts the file of data M using the unique I Di of the receiver i and generates a ciphertext C. C ←− Encr ypt (I Di , M)

(3)

4. Decryption Phase: In this phase, the receiver decrypts the ciphertext C using the secret key S K i generated by the PKG and gets the plaintext M. M ←− Decr ypt (C, S K i )

(4)

19 An Identity-Based Encryption Scheme for Data Security in Fog Computing

219

Fig. 2 Generic system architecture

In the following, Fig. 2 represents a general framework of an IBE-based encryption and decryption to present how this scheme actually works in the fog environment. Here, when a user i tries to send a message via fog server using Identity-Based Encryption and Decryption, the plaintext is encrypted with the ID of the user i. PKG server is implemented on fog. During decryption process, the secret key SK of user i is generated by the PKG using the I Di of user i.

3.2 Preliminaries In this section, we briefly review the concept of bilinear pairing and its several properties. Let, G 1 be a cyclic additive group of prime order p and G 2 be a cyclic multiplicative group of the same prime order p. And we call ê an admissible bilinear pairing ê: G 1 X G 2 −→ G 2 , which requires the following properties: – Bilinearity: For all (X, Y ) ∈ G 1 and all a, b ∈ Z∗p we have, ê(a X, bY ) = ê(X, Y )ab . – Non-degeneracy: There exists (X, Y ) ∈ G 1 such that ê(X, Y ) = 1. – Computability: For all (X, Y ) ∈ G 1 , there exists an efficient algorithm to compute ê(X, Y ) ∈ G 1 . The proof of our security assumption can be ensured using Diffie–Hellman message authentication algorithm [4].

220

N. Farjana et al.

3.3 Identity-Based Hierarchical Architecture for Fogging We have extended the work done in [11], where the authors used a three-level hierarchy for cloud computing. We have proposed here a four-level hierarchy where a new level is added as Fog layer. As shown in the Fig. 3, the HIBAF architecture is composed of four levels. Here, Level 0 is the top level which corresponds to Root PKG. Level 1 corresponds to sub-PKGs where each node corresponds to a data center, i.e., a cloud service provider. Level 2 are the sub-PKGs of Level 1 where each node corresponds to a data center, i.e., a fog service provider in the fog computing. The bottom level (Level-4) corresponds to the users in the fog computing. Each node in each level has a distinguished name or ID. For example, if the distinguished ID of root node is I D Root , node C2 in Level 1 is I DC2 , node F2 in Level-2 is I D F2 and node Un in Level 3 is I DUn , then we can define the identity of each node using the string concatenation through the root to the targeted node. Therefore, the calculation of the ID of each node in each level is as follows: In Level 0, I D N |0 = I D Root

(5)

I D N |1 = I D Root ||I DC2

(6)

In Level 1,

(taking C2 node from Level-1) In Level 2, I D N |2 = I D Root ||I DC2 ||I D F2 (taking F2 node from Level-2)

Fig. 3 Identity-based hierarchical architecture for fog computing

(7)

19 An Identity-Based Encryption Scheme for Data Security in Fog Computing

221

In Level 3, I D N |3 = I D Root ||I DC2 ||I D F2 ||I DUn

(8)

(taking Un node from Level-3) Here, “||” indicates string concatenation. Using the above rules, the identity of any node in any level can be calculated accordingly. The implementations of the above model require six different steps: Root (Level 0) PKG Setup, Mid-level (Level 1) Setup, Lower level (Level 2) Setup, User-level (Level 3) Setup, Encryption, and Decryption.

3.3.1

Root Level Setup

Root PKG works as follows. (i) It generates two groups G 1 ,G 2 of the same prime order p and an admissible pairing: ê: G 1 X G 2 −→ G 2 (ii) It chooses two cryptography hash functions H1 : {0, 1}∗ ∈ G 1 and H2 : G 2 ∈ {0, 1}k for some k. (iii) It selects a random S0 ∈ Z∗p and sets X 0 = H1 (I D Root ) and Y0 = S0 X 0 . The root PKG’s master secret is S0 which is only known to the Root PKG and the system parameters are (G 1 , G 2 ,ê,H1 , H2 , X 0 , Y0 ) which are public to the lower level PKGs.

3.3.2

Mid-Level Setup

Assume that, there are n nodes in the Level 1. For each node in the Level 1, the root PKG follows the following steps. Here, we have picked up an arbitrary node C2 among the n nodes for simplicity. (i) At first, compute the public key of node C2 using X C2 = H1 (I DC2 ), Where I DC2 = I D Root ||I DC2 (ii) Set the secret key of node C2 : SC2 = S0 X C2 (iii) It selects the secret element ρC2 ∈ Z∗p for node C2 . Here, ρC2 is only known to node C2 and the root PKG. (iv) Define the Y-value: Y I DC2 = ρC2 X 0 3.3.3

Lower Level Setup

Assume that there are n nodes in the Level 2. Here, F2 is a child node of C2 . For each child node (i.e. F2 ) in Level 2, its parent node (i.e. C2 ) works as follows: (i) At first, compute the public key of node F2 using X F2 = H1 (I D F2 ); Where I D F2 = I D Root ||I DC2 ||I D F2

222

N. Farjana et al.

(ii) Set the secret key of node F2 : S F2 = SC2 + ρC2 X F2 = S0 X C2 + ρC2 X F2 (iii) It selects the secret element ρ F2 ∈ Z∗p for node F2 . Here,ρ F2 is only known to node F2 and its parent node C2 . (iv) Define the Y-value: Y I D F2 = ρ F2 X 0 3.3.4

User-Level Setup

Assume there are n to j child nodes of node F2 which corresponds to each user in Level 3. For each child node, the parent node F2 works as follows. Here, we consider node Un as an arbitrary child node of node F2 . (i) At first, compute the public key of node Un using X Un = H1 (I DUn ); where I DUn = I D Root ||I DC2 ||I D F2 ||I DUn (ii) Set the secret key of node Un : SUn = S F2 + ρ F2 X Un = S0 X C2 + ρC2 X F2 + ρ F2 X Un (iii) It selects the secret element ρUn ∈ Z∗p for node Un . Here, ρUn is only known to node Un and its parent node F2 . (iv) Define the Y-value: Y I DUn = ρUn X 0 After successfully completing the above four steps for each and every node, all nodes in the Level 1, Level 2, and Level 3 will get their secret keys and secret elements which they keep secret and is only known to the individual nodes and their parent nodes. The public key and Y-value are publicized.

3.3.5

Encryption

Suppose that, U1 and U2 are two users in fog computing. Their parent node is F1 whose parent node is C1 . User U1 wants to send a message to user U2 . The identity of user U1 and U2 are as follows: I DU1 = I D Root ||I DC1 ||I D F1 ||I DU1 I DU2 = I D Root ||I DC1 ||I D F1 ||I DU2 In order to encrypt message m with I DU2 , user U1 follows the following steps: (i) Computes X 1 = H1 (I D Root ||I DC1 ) X 2 = H1 (I D Root ||I DC1 ||I D F1 ) X 3 = H1 (I D Root ||I DC1 ||I D F1 ||I DU2 ) (ii) Chooses a random r ∈ Z∗p (iii) Outputs the ciphertext C =< r X 0 , r X 1 , r X 3 , H2 (gr ) ⊕ m > Where, g = ê(Y0 , X 2 )

19 An Identity-Based Encryption Scheme for Data Security in Fog Computing

3.3.6

223

Decryption

After getting the ciphertext C =< P0 , P1 , P2 , V >, user U2 can decrypt the ciphertext C using its secret SU2 = S0 X 1 + ρ1 X 2 + ρ2 X 3 Here,ρ1 is the secret point of node F1 and C1 . And, ρ2 is the secret point of node U2 and F1 . (i) Computes 3 ê(Y I DU2 |i , Pi−1 ) d =ê(P0 , SU2 )/ i=2 where, Y I DU2 |2 = ρ1 X 0 , Y I DU2 |3 = ρ2 X 0 (ii) Outputs the message, m = H2 (d) ⊕ V

4 Security Analysis Our proposed HIBAF scheme provides four-layered security where each level’s security depends on the previous level’s security. Whenever an adversary tries to break into the system, he has to break this four-layered security. Moreover, our scheme provides unique secret keys for each and every node which in turn reserves the security of each node separately. It is quite impossible for an adversary to know the secret keys of each and every node and generate private keys. Thus, even if an adversary becomes successful to access the file, he would not be able to decrypt it without the private key. Moreover, every time the secret element of each and every node is generated randomly and chooses a prime number for its secret element which is only known to the root PKG and the particular node. Thus, we can claim that our proposed HIBAF scheme can prevent the data theft attacks through secure data transmissions.

5 Implementation and Result We have implemented our scheme in JAVA and evaluated our performance in terms of user load, memory utilization, response time and delay time. Then, we have compared our result with another cryptographic scheme named ABE (Attribute-Based Encryption). We have considered different sizes of datasets varying from 2 to 16 Mb to analyze the performances. In this case, we have used text, PDFs and documents type files. The red and green line in the figures show the computation of ABE and HIBAF, respectively. 1. User Load versus CPU Time: This metric shows how the cpu time varies with different sizes of datasets used for both ABE and HIBAF.

224

N. Farjana et al.

2. Memory Utilization versus File Size: This metric analyzed how much memory will be utilized for the selected sizes of datasets. It illustrates the comparison of memory utilization between ABE and HIBAF. 3. Encryption Time versus File Size: This metric calculates the varying encryption time for each selected datasets for both ABE and HIBAF. 4. Decryption Time versus File Size: This metric calculates the varying decryption time for each selected datasets for both ABE and HIBAF. 5. Delay Time versus File Size: This metric calculates the varying delays for each selected datasets for both ABE and HIBAF. Figure 4 presents that HIBAF provides less CPU time, more memory utilization, less encryption and decryption time, and less delay than ABE according to the larger sizes of datasets.

Fig. 4 Comparing cpu time, memory utilization, encryption time, decryption time, and delay time taken for varying sizes of datasets

19 An Identity-Based Encryption Scheme for Data Security in Fog Computing

225

6 Discussion The most important advantage of our proposed HIBAF scheme is that the transmission of private key and authentication is done locally.The root PKG delegates the key generation as well as identity authentication through the distribution of workloads to lower level PKGs which, in turn, reduces the computation cost of key generation. Another advantage of HIBAF scheme is that every level has individual secret keys, which limit the key escrow problem and removes the need for complicated certificate management. With the rapid increase of smart applications and the development of the world, the size of the network and data size are also increasing due to the future technologies. Managing such volume of data along with larger networks and providing security in every node of it seems to be a great deal in today’s world. HIBAF is an effective solution to this problem. It is applicable in any large network and IOT-based smart applications to provide security in every level of the management of data. From online marketing, fraud detection, Internet banking to research analysis everywhere HIBAF provides better utilization and performance to manage large networks in a simple and efficient way.

7 Conclusion Data security is one of the primary issues in fog computing as a number of IoT devices are connected via fogging. Our proposed HIBAF infrastructure consisting of four levels provides better security of data in fog computing paradigm. We presented the evaluation of the efficiency of our proposed scheme in terms of computation time, memory utilization, response time and delay according to the varying load of the data provided by the user. Finally, we compared our scheme with other cryptographic systems to evaluate the effectiveness of our scheme from other existing approaches. Through evaluation, we found our scheme overall 30% efficient and secure in fog computing environment. Although our proposed HIBAF scheme limits the key escrow problem, it does not remove it completely. Moreover, private key generation and transmission fully depend on the PKG which is a third-party provider. Thus, PKG must be secure and trusted. Furthermore, the private keys must be transmitted through secure channels.

References 1. Aazam M, Huh EN (2014) Fog computing and smart gateway based communication for cloud of things. In: 2014 international conference on future internet of things and cloud (FiCloud). IEEE, pp 464–470 2. Alrawais A, Alhothaily A, Hu C, Xing X, Cheng X (2017) An attribute-based encryption scheme to secure fog communications. IEEE Access 5:9131–9138

226

N. Farjana et al.

3. Bilge L, Strufe T, Balzarotti D, Kirda E (2009) All your contacts are belong to us: automated identity theft attacks on social networks. In: Proceedings of the 18th international conference on world wide web. ACM, pp 551–560 4. Blazy O, Kiltz E, Pan J (2014) (hierarchical) identity-based encryption from affine message authentication. In: International cryptology conference. Springer, Berlin, pp 408–425 5. Boneh D, Hamburg M (2008) Generalized identity based and broadcast encryption schemes. In: International conference on the theory and application of cryptology and information security. Springer, Berlin, pp 455–470 6. Bonomi F (2011) Connected vehicles, the internet of things, and fog computing. In: The eighth ACM international workshop on vehicular inter-networking (VANET), Las Vegas, USA. pp 13–15 7. Bonomi F, Milito R, Zhu J, Addepalli S (2012) Fog computing and its role in the internet of things. In: Proceedings of the first edition of the MCC workshop on mobile cloud computing. ACM, pp 13–16 8. Dong MT, Zhou X (2016) Fog computing: comprehensive approach for security data theft attack using elliptic curve cryptography and decoy technology. Open Access Libr J 3(09):1 9. Jiang Y, Susilo W, Mu Y, Guo F (2018) Ciphertext-policy attribute-based encryption against key-delegation abuse in fog computing. Futur Gener Comput Syst 78:720–729 10. Kahvazadeh S, Souza VB, Masip-Bruin X, Marn-Tordera E, Garcia J, Diaz R (2017) Securing combined fog-to-cloud system through SDN approach. In: Proceedings of the 4th workshop on crosscloud infrastructures & Platforms. ACM, p 2 11. Li H, Dai Y, Yang B (2011) Identity-based cryptography for cloud security. IACR Cryptol Eprint Arch 2011:169 12. Mozumder DP, Mahi, MJN, Whaiduzzaman M Cloud computing security breaches and threats analysis 13. Rahman MT, Mahi MJN (2014) Proposal for SZRP protocol with the establishment of the salted sha-256 bit hmac pbkdf2 advance security system in a manet. In: 2014 International conference on electrical engineering and information & communication technology (ICEEICT). IEEE. pp 1–5 14. Roy S, Shovon AR, Whaiduzzaman, M (2017) Combined approach of tokenization and mining to secure and optimize big data in cloud storage. In: Humanitarian technology conference (R10HTC), 2017 IEEE region 10. IEEE, pp 83–88 15. Schridde C, Dörnemann T, Juhnke E, Freisleben B, Smith M (2010) An identity-based security infrastructure for cloud environments. In: 2010 IEEE international conference on wireless communications, networking and information security (WCNIS). IEEE, pp 644–649 16. Shamir A (1984) Identity-based cryptosystems and signature schemes. In: Workshop on the theory and application of cryptographic techniques. Springer, Berlin, pp 47–53 17. Song YJ, Kim JM Secure data sharing based on proxy re-encryption in fog computing environment 18. Stojmenovic I, Wen S (2014) The fog computing paradigm: Scenarios and security issues. In: 2014 Federated conference on computer science and information systems (FedCSIS). IEEE, pp. 1–8 19. Stolfo, S.J., Salem, M.B., Keromytis, A.D.: Fog computing: Mitigating insider data theft attacks in the cloud. In: 2012 IEEE Symposium on Security and Privacy Workshops (SPW). IEEE, pp 125–128 20. Vishwanath A, Peruri R, He JS (2016) Security in fog computing through encryption. DigitalCommons@ Kennesaw State University 21. Waters B (2009) Dual system encryption: realizing fully secure IBE and HIBE under simple assumptions. In: advances in cryptology-CRYPTO 2009. Springer, Berlin, pp 619–636 22. Whaiduzzaman M, Gani A (2014) Measuring security for cloud service provider: a third party approach. In: 2013 International conference on electrical information and communication technology (EICT). IEEE, pp 1–6 23. Whaiduzzaman M, Naveed A, Gani A (2016) Mobicore: mobile device based cloudlet resource enhancement for optimal task response. IEEE transactions on services computing

Chapter 20

Modeling Photon Propagation Through Human Breast with Tumor in Diffuse Optical Tomography Shisir Mia, Md. Mijanur Rahman and Mohammad Motiur Rahman

1 Introduction Light imaging techniques in the near-infrared (NIR) wavelength benefit from low tissue absorption. NIR photons can propagate over several centimeters within biological tissue which aids highly to explore the inner tissue structure. The development of diffuse optical tomography (DOT) [1], which uses near-infrared light has led to a new imaging technique which is able to map chemical concentrations into optical properties of the tissue for breast tumor detection and other diagnoses as well. In diffuse optical tomography (DOT), biological tissue is illuminated by the nearinfrared light in the range from 700 to 900 nm. The reflected light is observed at the surface of the tissue. These observations are then used to assess the optical proprieties of the inner tissue. DOT is a very challenging imaging problem since NIR light is intensely scattered in the biological tissues. That means only a few amounts of light are transmitted through the tissue. Therefore, the accurate recovery of images using numerical modeling requires the efficient image reconstruction approach. Many researchers have put much efforts for developing photon propagation models and image reconstruction methods in diffuse optical tomography. Among them, Althobaiti et al. [2] used Radiative Transfer Equation with diffusion approximation [3, 4] to model the forward problem and measured the optical proprieties of the inner tissue using near-infrared fluorescence and spectral tomography (NIRFAST) S. Mia (B) · Md. M. Rahman Department of Computer Science & Engineering, Uttara University, Dhaka, Bangladesh e-mail: [email protected] Md. M. Rahman e-mail: [email protected] M. M. Rahman Department of Computer Science and Engineering, Mawlana Bhashani Science and Technology University, Santosh, Tangail 1902, Bangladesh e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_20

227

228

S. Mia et al.

[5, 6] and born approximation method [7, 8]. Peter [9] also used Radiative Transfer Equation with diffusion approximation to model the photon propagation in biological tissue and reconstructed the image using Levenberg–Marquardt method and Tikhonov Minimization approach. In this paper, we address the problem of mathematical modeling of near-infrared electromagnetic wave propagation through human breast and therefore finding clues for detecting anomaly like tumor in it. Our main contributions are as follows: (i) NIR photon propagation through human breast with anomaly like tumor in diffuse optical tomography has been mathematically modeled using Radiative Transfer Equation with diffusion approximation and (ii) signatures of photon density distributions at the location of tumor within the human breast has been explored through simulation of the aforesaid photon propagation model.

2 Photon Propagation Model There are mainly two possible approaches to model photon propagation in biological tissue [1]. First, the numerical Monte Carlo method can be used through the tracking of an adequate number of photons. The numerical Monte Carlo method has the ability to model complex geometries, as well as the complex heterogeneous media, but it takes historically extensive computation times. Second, the analytical models have the benefit of being computationally fast but undergo from the weakness of being limited to simple geometries. It can be possible to use analytically the Radiative Transfer Equation (RTE) in modeling photon propagation through biological tissue. However, the RTE is challenging to compute and therefore it is required to approximate the RTE to the diffusion equation. One approximation is that photon    propagation is controlled by scattering rather than absorption μa  μs . This is because the spectrum window of NIR light falls in the range from 700 nm to 900 nm. Another approximation is that as the propagation of light in biological tissue at NIR wavelengths over distances greater than a few scattering lengths becomes nearly isotropic, it can be predicted well by the photon diffusion. The diffusion equation for monochromic light is written as   ∂U ( r , t) + cμa U ( r , t) − c∇. D∇U ( r , t) = q( r , t) ∂t

(1)

where U is the photon density, q is photon source strength, c is the speed of  density   light inside the medium, and D = 1/3 μa + μs is the diffusion coefficient of the medium. This equation can be rewritten when frequency modulation (ω) is adopted as: −i ω U ( r ) + cμa U ( r ) − cD∇ 2 U ( r ) = Bδ( r) Which can be reformulated to Helmholtz wave equation [10] as

(2)

20 Modeling Photon Propagation Through Human Breast …

 2  δ( r) ∇ + k 2 U ( r ) = −B cD where k 2 = [11].

−cμa +i ω cD

229

(3)

is the wave propagation constant and δ( r ) is a delta function

3 Modeling Human Breast with Tumor To model the human breast, we employ a hemisphere and also use a small sphere of radius 0.1 m within the hemisphere as tumor. In this model, we consider from the domain expert that the human breast has absorption coefficient μa = 0.02 cm−1  and reduced scattering coefficient μs = 6 cm−1 and the breast tumor has absorption  coefficient μa = 0.002 cm−1 and reduced scattering coefficient μs = 6 cm−1 . The Dirichlet boundary condition is used at the surface or boundary of the human breast because we assume zero photon density at this space and also the Neumann boundary condition is used at the boundary of tumor for the continuous distribution of photon density. A simple geometric view of the model is given in Fig. 1.

4 Simulation To simulate the photon propagation model which is described by Radiative Transfer Equation (RTE) with diffusion approximation, we use the Helmholtz equation in

Fig. 1 Model of human breast with tumor

230

S. Mia et al.

COMSOL Multiphysics framework. In this framework, the Helmholtz equation is specified as ∇(−c∇u) + au = f

(4)

With parameters, u=U c=D=

1 3(μa + μs )

a = μa and, f = q where u is the photon density, c is the diffusion coefficient (isotropic), a is the absorption coefficient, and f is the photon density source term. In the Helmholtz equation, we have used the absorption coefficients of human breast and tumor, which are mentioned in the previous section and accordingly, the diffusion coefficients are computed from the absorption and reduced scattering coefficients for both objects, respectively. The four light sources of 700 nm wavelength are used instead of using photon density source f. To solve the Helmholtz equation for modeling, the photon propagation through biological tissue by applying finite element method (FEM) [12–14], we have exerted finer mesh for human breast and extra fine mesh for breast tumor. In the selection of the resolutions of mesh, we have taken into account the following numerical considerations. If a too coarse mesh is applied, the solution of diffusion equation is most likely incorrect due to discretization errors. On the other hand, working for the finest resolution of the mesh directly will render a huge number of mesh elements that most of the computers cannot store in memory. The type of study known as the stationary process is used to get the simulated results.

5 Results and Discussions This section presents and analyses the simulation results. Figure 2 represents the 3D meshing of the entire structure where, as mentioned, the mesh is further refined in the region of interest like the breast tumor because the mesh resolution defines the accuracy of the solution. Figure 3 represents the simulated 3D model of human breast with a tumor where we have used four light sources as discussed earlier. The sliced view in Fig. 3 shows the slice-wise distribution of the photon density for the human breast. In order to view photon density distribution across the important lines, we identified two lines in the model. The first line spans from the coordinates (0.5, 0, 0.83) m to (−0.5, 0, 0.83) m as shown in Fig. 4. The second line spans from (0.2, −0.3,

20 Modeling Photon Propagation Through Human Breast …

Fig. 2 3D mesh of the human breast and the tumor

Fig. 3 Simulated 3D model of a human breast with tumor

231

232

S. Mia et al.

Fig. 4 a Line in the 3D model of human breast with tumor; b Photon density of the corresponding line in the simulated 3D model

Fig. 5 a Line in the 3D model of human breast with tumor; b Photon density of the corresponding line in the simulated 3D model

0.85) m to (−0.2, 0.3, 0.85) m and is shown in Fig. 5. The photon densities across these lines are shown in Fig. 4b and Fig. 5b, respectively. A close analysis of the photon density profiles across these lines reveals that there would be a sharp fall in photon density at the periphery of the tumor and photon density would remain very low throughout the whole tumor. When we closely observe Fig. 4b, it becomes clear that there is significant photon density everywhere along the line except the line segment from 0.4 m to 0.6 m. This gives hints that there is some anomaly within this line segment. When correlated with Fig. 4a, we find that this line indeed falls within the tumor. Also, it is clearly seen in Fig. 5b that photon densities along the line first decreases rapidly from 0 to 0.29 m, then remain approximately flat from 0.29 m to 0.44 m. Finally, the densities rises from 0.44 to 0.75 m. When correlated with Fig. 5a, we find that the line segment from 0.29 to 0.44 m goes through tumor.

20 Modeling Photon Propagation Through Human Breast …

233

6 Conclusion and Future Works In this paper, photon propagation through human breast with anomaly like tumor is mathematically modeled and simulated in accordance with the principles of diffuse optical tomography. The modeling is done with the help of Radiative Transfer Equation (RTE) with diffusion approximation, which is finally transformed into Helmholtz equation. To develop a real-space model of the human breast with tumor, different types of optical properties are considered and the several boundary conditions are applied. The overall model is simulated using FEM in COMSOL Multiphysics framework. The results show the photon density distributions within the human breast with tumor in several ways and also reveal an effective way of estimating the position of tumor in human breast through analysis of the photon density distribution signature in and around the tumor. Our future research direction would consist of image reconstruction by solving the inverse problem in an efficient manner, and estimation of size and position of the anomaly from this reconstructed image.

References Hielscher AH, Bluestone AY, Abdoulaev GS, Klose AD, Lasker J, Stewart M, Netz U, Beuthan J (2002) Near-infrared diffuse optical tomography. Disease Markers 18(5–6):313–337 Althobaiti MM, Salehi HS, Zhu Q (2015) Assessment of diffuse optical tomography image reconstruction methods using photon transport model. In: Proceedings of COMSOL conference, Boston Arridge SR (1999) Optical tomography in medical imaging. Inverse Prob 15(2):R41 Tarvainen T, Vauhkonen M, Kolehmainen V, Kaipio JP, Arridge SR (2008) Utilizing the radiative transfer equation in optical tomography. Piers Online 4(6):655–660 Jermyn M, Ghadyani HR, Mastanduno MA, Turner WD, Davis SC, Dehghani H, Pogue BW (2013) Fast segmentation and high-quality three-dimensional volume mesh creation from medical images for diffuse optical tomography. J Biomed Optics 18(8):086007 Dehghani H, Eames ME, Yalavarthy PK, Davis SC, Srinivasan S, Carpenter CM, Pogue BW, Paulsen KD (2009) Near infrared optical tomography using NIRFAST: algorithm for numerical model and image reconstruction. Commun Numeric Methods Eng 25(6):711–732 Zhu Q, Xu C, Guo P, Aguirre A, Yuan B, Huang F, Castilo D, Gamelin J, Tannenbaum S, Kane M, Hegde P (2006) Optimal probing of optical contrast of breast lesions of different size located at different depths by US localization. Technol Cancer Res Treatm 5(4):365–380 Huang M, Zhu Q (2004) Dual-mesh optical tomography reconstruction method with a depth correction that uses a priori ultrasound information. Appl Opt 43(8):1654–1662 Peter S (2014) Comparative study on 3D modelling of breast cancer using Nir-Fdot in COMSOL. In: Proceedings of COMSOL conference, Bangalore Jensen FB, Kuperman WA, Porter MB, Schmidt H (2000) Computational ocean acoustics. Springer Science & Business Media Wang LV, Wu HI (2007) Biomedical optics: principles and imaging. Wiley, New York Bazrafkan S, Kazemi K (2014) Modeling time resolved light propagation inside a realistic human head model. J Biomed Phys Eng 4(2):49 Arridge SR, Schweiger M, Hiraoka M, Delpy DT (1993) A finite element approach for modeling photon transport in tissue. Med Phys 20(2):299–309 Jin JM (2002) The finite element method in electromagnetic. Wiley, New York

Chapter 21

A Network-Based Approach to Identify Molecular Signatures and Comorbidities of Thyroid Cancer Md. Ali Hossain, Tania Akter Asa, Fazlul Huq, Julian M. W. Quinn and Mohammad Ali Moni

1 Introduction Comorbidities, where at least one disease or disorder coexists in an individual with primary disease [13], may result from direct or indirect pathogenic relationships or common risk factors between the coinciding diseases and have been extensively studied in mental disorders, chronic obstructive pulmonary disease (COPD) and cancer [12]. Comorbidity is an important factor influencing treatment and survival in cancer and comorbidity disease associations have been identified at the molecular level by studies of protein protein interactions (PPIs), common gene dysregulation and metabolic pathways [14]. A pair of diseases can be said to be associated if they display common dysregulated genes or gene pathways and a number of studies have constructed networks for diseases comorbidities using such gene-disease associations [12]. For example, clinical studies have found that the treatments of cancer Md. Ali Hossain Department of Computer Science and Engineering, Manarat International University, Jahangirnagar University, Dhaka, Bangladesh e-mail: [email protected] T. A. Asa Department of Electrical and Electronics Engineering, Islamic University, Kushtia, Bangladesh e-mail: [email protected] F. Huq · M. A. Moni (B) Discipline of Pathology, School of Medical Sciences, Faculty of Medicine and Health, The University of Sydney, Sydney, Australia e-mail: [email protected] F. Huq e-mail: [email protected] J. M. W. Quinn · M. A. Moni Bone biology divisions, Garvan Institute of Medical Research, Sydney, Australia e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_21

235

236

Md. Ali Hossain et al.

play a role in increasing hypertension, cardiac, vascular disease risk among TC survivors and neurological disease incidence have also been shown to correlate with thyroid disorders [9]. However, a comorbidity may be crucial to patient treatment and survival, yet may not be detected the patient concerned. To determine whether this may be the case in some patients we performed a comorbidity study of TC patients using transcript data. The aim of our study is to describe the molecular signatures, and spectrum of comorbidity, to elucidate possible comorbidity impact on treatment and survival in patients with TC.

2 Materials and Methods In this research work, we developed and applied multi-step analysis method (see Fig. 1). For identifying the DEGs, DEGs regulatory patterns, gene expression dataset was statistically analyzed. The Enriched pathways, annotation terms (i.e., Gene Ontology terms), and biological processes were identified by considering DEGs and functional enrichment studies. From the previous study, it was found that thyroid cancer patient can be affected with other diseases. For this reason, comorbiditiy network of TC was developed by DEGs analysis and human disease-genes associations. After then, considering the common genes between TC and other diseases, Protein-Protein Interaction (PPI) network was constructed and the topological analy-

Fig. 1 The multi-stage analysis methodologies were used in this study. Gene expression dataset of TC were collected from the NCBI-GEO database. By using GEO2R, the dataset was statistically analyzed to identify DEGs. After then using the DEGs and the gene-disease associations data from OMIM classified TC DEGs into 21 diseases category and found 79 common genes to other diseases. To identify significantly enriched pathways among common DEGs, 4 types of functional enrichment analyses were then performed on common DEGs. So, we constructed PPI networks around the common DEGs, performed topological analyses to identify putative pathways hub proteins, identified possible transcriptions factor (TF) interactors and micro-RNA (miRNA), and to provide pathways enrichment,Gene Ontology annotation terms was used. microRNA and TF studies used miRTarbase and JASPAR databases, respectively. common DEGs were combined with those networks and betweenness centrality and higher degree was used to designate microRNAs and TFs. For pathway enrichment analyses, the target DEGs of reporter biomolecules (microRNAs and TFs) were subjected

21 A Network-Based Approach to Identify Molecular Signatures …

237

ses were conducted via Cytoscape to identify hub proteins. Then, to identify reporter biomolecules(i.e. TFs and miRNA), we integrated the intermediate analysis results with biomolecular networks. Dataset Employed and Statistical Methods Used From the NCBI Gene Expression Omnibus (GEO) (http://www.ncbi.nlm.nih.gov/ geo/)[2], we obtained the gene expression data of thyroid cancer (GSE82208) for our study. The data sets of the patients of different sex and age were analysised. For getting microarrays representing thyroid cancer, we performed several rounds of filtering, normalization and statistical analysis. To find out genes that was differentially expressed in patients, Students unpaired t-test was performed. A threshold value of fold change J˜ tp , the requested power is supplied by the PV and battery storage. The requested power is expressed as Qr = Jtp − J˜ tp , Qr > 0 ⎞ ⎛ ⎜ b pv ⎟ Qr = ⎝ Qr , Qr ⎠

  battery

(7)

(8)

PV

pv

Here, Qrb , Qr are the requested power from PV and battery, respectively, to flatten pv will supply the requested power based on its the load curve. If Q r  = 0, battery pv  b  = Q , Q constraints. While Q r r , grid will respond to the mismatch power,   b pv r Qr , Qr . Let us assume the lower and upper state-of-charge (SOC) i.e., Qr − max boundaries of battery are λmin b and λb , respectively. The maximum available power from a battery at tth time is ˜ rb = Q

   t b ˜b ∗ ϕcb ∗ η2 ∗ η3 , for λtb > λmin λb − λmin b b , Qr ∈ Qr

(9)

27 A Day-Ahead Power Demand Prediction for Distribution …

311

Here, λtb is the instantaneous SOC at tth time and ϕcb is the battery storage capacity. η2 , η3 are the efficiency of the converter used to connect the battery to the main bus. While Jtp < J˜ tp , the available power from the grid is calculated as Qg = J˜ tp − Jtp , Qg > 0, t ∈ tb , tp

(10)

This Qg is the maximum available power from the grid, as battery is connected to the ...  intermediate bus with the PV. Thus, the real power consumption of the battery Qg from Qg depends on the PV power generation. If the consumer has H number of 1 installed   PV modules, each having Q0 power output, the total PV power generation Qpv is expressed as  ...  ... Qg = Qgb − Qpv , Qg ∈ Qg (11)  

battery

Qpv = H ∗ Q01 ∗ η1

(12)

where η1 is the efficiency of the converter used to integrate the PV module with the DC bus. If the upper SOC boundary of battery is λmax b , the grid utilized power by the battery is ˜ gb = Q

   max b ˜b λb − λtb ∗ ϕcb ∗ η2 ∗ η3 , for λtb < λmax b , Qg ∈ Qg

(13)

Considering the scenarios of both Jtp < J˜ tp and Jtp > J˜ tp , the number of loads and sources connected to the AC bus is ⎡ Jtp ⎢

 ⎢ power consumption from grid t ˜t Loads ⎢ (14) ⎢ λmax − λt  ∗ ϕb  ∗ η ∗ η  for Jp < Jp 2 3 b c ⎣ b   battery

  H ∗ Q01 ∗ η1

  ⎢ ⎢ PV power generation    for Jtp > J˜ tp Sources ⎢  t b ∗ ϕ ∗ η ∗ η ⎣ λb − λmin 2 3 b

 c  ⎡

battery

(15)

312

K. Mahmud et al.

4 Case Studies In this section, the proposed control algorithm described in Sect. 3 is tested in a real Australian power distribution network, located in Nelson Bay, NSW. The local climate data is fed to the system model to get the real data of PV power generation. Based on that real data, ANN is modeled to predict the PV power generation. A similar approach is taken to predict the household power demand. Based on the = 95%, λmin predicted data, a battery storage (10 kWh capacity, λmax b b = 35%) and a PV are used to manage the power demand. In this case, two different houses (house 1 and 2) are used. The system tested in house 1 is shown in Fig. 8. The same model and approach are applied to house 2 (different load curve), as shown in Fig. 9. It is clear from both figures that day-ahead energy management can significantly reduce the peaks in the load curve.

27 A Day-Ahead Power Demand Prediction for Distribution …

313

4500

Power Demand (W)

Day-ahead power demand

4000

Day-ahead power demand after management

3500 3000 2500 2000 1500

24

12

0

36

48

Time (Hours)

Fig. 8 Day-ahead power demand management at house 1 4500

Power Demand (W)

Day-ahead power demand Day-ahead power demand after management

4000 3500 3000 2500 2000 1500

0

24

12

36

48

Time (Hours)

Fig. 9 Day-ahead power demand management at house 2 (different load curve) 4500

Power Demand (W)

Day-ahead power demand

4000

Day-ahead power demand after management

3500 3000 2500 2000 1500

0

12

24

36

48

56

Time (Hours)

Fig. 10 Performance of the day-ahead power demand management under uncertainty

The performance of the proposed system is tested under an uncertainty condition, as shown in Fig. 10. An uncertainty is added to the power demand to investigate the performance of the proposed algorithm. The figure shows that even in uncertainty the peak demand is reduced significantly.

314

K. Mahmud et al.

5 Conclusion The study aimed to manage the power demand of the consumer in a day ahead of the time. The investigation shows that ANN performs better to predict the variable load demand and PV power generations even under various uncertainties. However, for any short-time fluctuation, the prediction technique shows linear behavior which ultimately affects the demand management. The finding under various case studies suggests that the approach can substantially reduce the day-ahead peak power demand of consumers.

References 1. Merabet A, Ahmed KT, Ibrahim H, Beguenane R, Ghias AMYM (2017) Energy management and control system for laboratory scale microgrid based wind-PV-battery. IEEE Trans Sustain Energy 8:145–154 2. Kangning LIU, Qixin C, Chongqing K, Wei SU, ZHONG G (2018) Optimal operation strategy for distributed battery aggregator providing energy and ancillary services. J Mod Power Syst Clean Energy 1–11 3. Hesse HC, Schimpe M, Kucevic D, Jossen A (2017) Lithium-Ion battery storage for the grid—a review of stationary battery storage system design tailored for applications in modern power grids. Energies 10:2107 4. Yan R, Roediger S, Saha TK (2011) Impact of photovoltaic power fluctuations by moving clouds on network voltage: a case study of an urban network. In: 2011 21st Australasian Universities power engineering conference (AUPEC), IEEE, pp 1–6 5. Wang L, Bai F, Yan R, Saha TK (2018) Real-time coordinated voltage control of PV inverters and energy storage for weak networks with high PV penetration. IEEE Trans Power Syst 33:3383–3395 6. Otashu JI, Baldea M (2018) Grid-level battery operation of chemical processes and demandside participation in short-term electricity markets. Appl Energy 220:562–575 7. Yang Y, Ye Q, Tung LJ, Greenleaf M, Li H (2018) Integrated size and energy management design of battery storage to enhance grid integration of large-scale PV power plants. IEEE Trans Ind Electron 65:394–402 8. Li J, Wu Z, Zhou S, Fu H, Zhang X-P (2015) Aggregator service for PV and battery energy storage systems of residential building. CSEE J Power Energy Syst 1:3–11 9. Aktas A, Erhan K, Ozdemir S, Ozdemir E (2017) Experimental investigation of a new smart energy management algorithm for a hybrid energy storage system in smart grid applications. Electr Power Syst Res 144:185–196 10. Simões MG, Busarello TDC, Bubshait AS, Harirchi F, Pomilio JA, Blaabjerg F (2016) Interactive smart battery storage for a PV and wind hybrid energy management control based on conservative power theory. Int J Control 89:850–870 11. Tazvinga H, Zhu B, Xia X (2015) Optimal power flow management for distributed energy resources with batteries. Energy Convers Manag 102:104–110 12. Mahmud K, Hossain MJ, Town GE (2018) Peak-load reduction by coordinated response of photovoltaics, battery storage, and electric vehicles. IEEE Access 6:29353–29365 13. Reihani E, Sepasi S, Roose LR, Matsuura M (2016) Energy management at the distribution grid using a battery energy storage system (BESS). Int J Electr Power Energy Syst 77:337–344 14. Howlader HOR, Sediqi MM, Ibrahimi AM, Senjyu T (2018) Optimal thermal unit commitment for solving duck curve problem by introducing CSP, PSH and demand response. IEEE Access 6:4834–4844

27 A Day-Ahead Power Demand Prediction for Distribution …

315

15. Li C, Yu X, Yu W, Chen G, Wang J (2017) Efficient computation for sparse load shifting in demand side management. IEEE Trans Smart Grid 8:250–261 16. Shirazi E, Jadid S (2017) Cost reduction and peak shaving through domestic load shifting and DERs. Energy 124:146–159 17. Shakeri M, Shayestegan M, Abunima H, Reza SMS, Akhtaruzzaman M, Alamoud ARM, Sopian K, Amin N (2017) An intelligent system architecture in home energy management systems (HEMS) for efficient demand response in smart grid. Energy Build 138:154–164 18. Tabar VS, Jirdehi MA, Hemmati R (2017) Energy management in microgrid based on the multi objective stochastic programming incorporating portable renewable energy resource as demand response option. Energy 118:827–839 19. Mahmud K, Hossain MJ, Ravishankar J (2018) Peak-load management in commercial systems with electric vehicles. IEEE Syst J 1–11 20. Arcos-Aviles D, Pascual J, Guinjoan F, Marroyo L, Sanchis P, Marietta MP (2017) Low complexity energy management strategy for grid profile smoothing of a residential grid-connected microgrid using generation and demand forecasting. Appl Energy 205:69–84 21. Shams MH, Shahabi M, Khodayar ME (2018) Stochastic day-ahead scheduling of multiple energy carrier microgrids with demand response. Energy 155:326–338 22. Da Silva IN, Spatti DH, Flauzino RA, Liboni LHB, dos Reis Alves, SF (2017) Artificial neural networks. Springer International Publishing, Cham 23. Adhikari R, Agrawal RK (2013) An introductory study on time series modeling and forecasting. arXiv:1302.6613 24. Mahmud K, Morsalin S, Hossain MJ, Town GE (2017) Domestic peak-load management including vehicle-to-grid and battery storage unit using an artificial neural network. In: Proceedings of the IEEE international conference on industrial technology

Chapter 28

Simulation and Comparison of RPL, 6LoWPAN, and CoAP Protocols Using Cooja Simulator Arif Mahmud, Faria Hossain, Tasnim Ara Choity and Faija Juhin

1 Introduction Internet of Things (IoT) denotes the dynamic connectivity of physical devices with limited resources with the support of Internet infrastructure. It helps to develop interaction between these entities and also every Internet supported objects and networks [1]. IoT can prolong Internet communication to heterogeneous types of objects that are used within embedded technology to join with the surrounding through the active support of Internet technology [2]. Based on the necessities and regular application, IoT has already been included in the national five-year development plan outline of China [3]. The exploration and application of the Internet of things technology are certain to quicken the industrial promotion and revolution, at the same time, to securely promote the growth of the national economy and to constantly improve the comprehensive national power [4]. The amount of connected entities has already gone beyond the total amount of human being few years ago, and it is predicted to reach 50 billion by the end of 2025 [5]. As more and more devices are being introduced in our regular life, it has created immense pressure on Internet connectivity in terms of storage and routing. These devices are constantly communicating with each other and also with the server

A. Mahmud (B) · F. Hossain · T. A. Choity · F. Juhin Department of CSE, Daffodil International University, Dhaka, Bangladesh e-mail: [email protected] F. Hossain e-mail: [email protected] T. A. Choity e-mail: [email protected] F. Juhin e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_28

317

318

A. Mahmud et al.

through the Internet. Therefore, the necessities of application-specific communication protocols have risen up and will continue too in future days to come. This paper is mainly focused on a comparative analysis on the basis of the Internet of Things’ architecture. This analysis will help us to develop communication among three protocols, 6LoWPAN, RPL, and CoAP. In addition, it will also find out the best communication path between the IoT nodes. A large amount of the data that are consumed by IoT will be kept in the cloud. The real issue is to develop the ability of the people to understand the variations and their inferences more clearly, and to take solid actions accordingly [6]. Different simulators have been used by a number of researchers, such as NS-3, Tossim, and OPNET for different platforms like TinyOS, POSIX, lwIP, etc. However, cooja was selected to simulate contiki nodes on a large scale. This simulator is specially designed to simulate sensors that consume very low power and proven to be very accurate. This paper is concentrated on finding out the best protocol for communication between IOT nodes, and therefore, the study can be useful in personal and home application, health care, utilities and services, enterprise application, and industrial automation. Hence, the paper is schematized as follows: Section 2 illustrates the basic concept of three protocols, namely, RPL, CoAP, and 6LoWPAN; Sect. 3 defines the sample communication scenario using these protocols; Sect. 4 explains the results based on comparison among these protocols; and the paper is concluded in Sect. 5.

2 Background Communication protocols are similar to traffic directors that maintain organized communication with each other. The protocol has the main responsibilities to develop and maintain communication, and without protocols, the nodes of various networks might get connected but will not be capable to interact with each other. Therefore, if one machine wants to receive or send data to another machine, both of them should use a similar protocol for completing the task [7]. As a result, choosing an appropriate communication protocol is one of the vital concerns nowadays. A number of research works are going on these IoT nodes which consume limited resources. Development of communication protocols can be considered as the vital one. At present, three communication protocols have become popular, and therefore, we have chosen them to simulate using cooja simulator. The basic features of protocols are explained in the following text in brief: The 6LoWPAN concept was originated from the idea that the Internet Protocol could be applied to the smallest devices [8] and those low-power devices with limited processing capabilities. The standard has the liberty of choosing frequency band and the elasticity to complete over several communications stages, counting Ethernet, Wi-Fi, and 6LoWPAN protocol. This protocol plays a most imperative role in IoT wireless communication as it stands for IPv6. It is demonstrated by auxiliary addresses with diverse lengths, low bandwidth, star and mesh topologies, battery

28 Simulation and Comparison of RPL, 6LoWPAN, and CoAP Protocols …

319

supplied devices, low cost, large number of devices, unknown node positions, high unreliability, and long idle periods during the communications [9]. RPL is one of the distance vector routing protocols for network layer, and it is developed to function with little energy and lossy systems using IPv6. Besides, message confidentiality and reliable message delivery are cared by RPL. It has been prearranged in such a technique that link layer appliances can be castoff when available and appropriate; yet, in their absence, RPL can custom its own mechanisms [10]. RPL is popular for many reasons; for example, it can develop its own route very fast, share the routing information to other nodes, and have the capability to get adapted dynamically with the network topology [11]. CoAP is considered as one of the most recent application layer protocols founded by IETF to be utilized for intelligent objects. In addition, it is a trivial protocol for low resource consumed smart devices and can be embedded into buildings, vehicles. It is also considered as the substitution of HTTP protocol to be used in the application layer of IoT devices [12]. As a result, the data reception from the sensor has become more convenient. In addition, this protocol has the multicast feature, a little overhead though and very effective in M2 M interaction. CoAP nodes are divided into two types, namely, server and client from the architecture viewpoint. Server nodes are installed in sensors while client nodes are embedded into controller [13].

3 Communication Using Protocols IPv6 Low-Power Wireless Personal Area Network (6LoWPAN) is functional on adaptation layer that allows the consumption of IPv6 over low-power wireless and User Datagram Protocol (UDP) header compression (Fig. 1).

Fig. 1 Network architecture of 6LoWPAN [14]

320

A. Mahmud et al.

The 6LoWPAN stack is both highly scalable and reliable. The modular structure of the design makes possible for ARM to accommodate most of the requirements. The combination of 6LoWPAN stack and 6LoWPAN border router Access Point (AP) enables developers to use the extremely flexible and multipurpose mesh communication for most applications. As can be seen from Fig. 2, the server and the client nodes are communicating with each other using 6LoWPAN routing protocol. Here, the green node is a UDP server and yellow nodes are UDP client (Fig. 3). DIO carries information that allows a node to discover an RPL Instance, learn its configuration parameters, and select DODAG parents. A DIS solicits a DODAG Information Object from an RPL node. Destination Advertisement Object (DAO) propagates destination information upward along the DODAG. As shown in Fig. 4, the sender and the receiver are communicating with each other using RPL routing protocol. Here, green nodes are unicast sender, and yellow nodes are unicast receiver in RPL Nodes. As shown in Fig. 5, CoAP cooperative model is related to HTTP’s client/server model and CoAP employs a two-layer structure. The bottom layer is the message layer that has been designed to deal with UDP and asynchronous switching. The request/response layer concerns communication method and deals with the request–response message [12]. As shown in Fig. 6, the sender and the receiver communicate with each other using CoAP routing protocol. Here, green nodes are unicast sender, and yellow nodes are unicast receiver.

Fig. 2 6LoWPAN nodes are communicating on cooja platform

28 Simulation and Comparison of RPL, 6LoWPAN, and CoAP Protocols …

Fig. 3 Functional system of RPL [12]

Fig. 4 RPL nodes are communicating on cooja platform

321

322

A. Mahmud et al.

Fig. 5 Network architecture of CoAP [12]

Fig. 6 CoAP nodes are communicating on cooja platform

4 Implementation and Results The performance of 6LoWPAN, CoAP, and RPL protocols will be compared depending on different criteria like received packet, number of nodes, time, and number of IOT nodes on cooja platform. Cooja is utilized to simulate contiki nodes which consume low resources like energy and bandwidth to communicate with each other on Ubuntu operating system. At present, this simulator is very popular among the network researchers since accurate behavior of the system can be examined, and simulations of large networks are possible before embedded them into devices [15]. As can be seen in Fig. 7, the X-axis indicates the number of total packets, and Yaxis indicates the number of received packets with 17 nodes. The 6LoWPAN received more packets than RPL and CoAP. For example, 6LoWPAN received 425 and RPL received 220 packets out of 500. However, CoAP stopped receiving packets after a certain point. As in Fig. 8, the number of nodes has been increased to 150 from 17 and has been compared based on previous variables. Similar results have been observed, the performance of RPL improved though. It can be noticed that the 6LoWPAN received almost all the sent packets.

28 Simulation and Comparison of RPL, 6LoWPAN, and CoAP Protocols …

323

Fig. 7 Comparison of received packet for 17 nodes

Fig. 8 Comparison of received packet for 150 nodes

Transmission range has been increased from 10 to 50 m. It can be seen from Fig. 9 that the transmission rate of CoAP is very slow. But 6LoWPAN’s packet transmission rate is better than RPL’s packet transmission rate. Communication range has been varied for these three protocols in Fig. 10 for 10, 50, and 100 ms. The 100 meter range nodes received 800 packets, 50 m range device received 720, packets and 10 m range devices received less than 500 packets out of 1200.

324

A. Mahmud et al.

Fig. 9 Comparison of received packet for 50 m range

Fig. 10 Comparison of received packet for ranges 10, 50, and 100 m in 6LoWPAN protocol

Similarly, communication range has been varied for RPL protocol in Fig. 11. Communication range has been taken 10, 50, and 100 ms. As expected, 100 m range received more packets than others for a fixed amount of time. 100 meter range has received 1600 packets where 10 m and 50 m ranges have received 750 and 650 packets, respectively (Fig. 12). Not so much changes are observed for different ranges for CoAP protocols. However, the transfer rate of ranges 100 m and 50 m is slightly faster than 10 m range.

28 Simulation and Comparison of RPL, 6LoWPAN, and CoAP Protocols …

325

Fig. 11 Comparison of received packet for ranges 10, 50, and 100 m in RPL protocol

Fig. 12 Comparison of received packet for ranges 10, 50, and 100 m in CoAP protocol

5 Conclusion IoT is one of the key topics in the present field of networking research. Since small devices which are embedded into IoT are connected via Internet, the consumption of power, memory, energy, routing, bandwidth, etc., are the vital issues. In this paper, IoT protocol stack along with different protocols used at different layers has been studied for the efficient communication among IoT nodes through the Internet. The paper focused on investigating a solution of finding out the best communication among three protocols, 6LoWPAN, RPL, and CoAP. This result shows that CoAP packet transfer rate is fair but slow. But on the other hand, RPL packet loss rate is very high with fast communication. After the analysis, it can be stated that 6LoWPAN performs better than RPL and CoAP. These protocols

326

A. Mahmud et al.

will be merged in our next paper and, therefore, will be attempted to find out the changes in results in comparison to this paper. We are also hopeful to implement this protocol in real testbed and not so much deviations are expected though.

References 1. Ray PP (2018) A survey on Internet of things architectures. J King Saud Univ 291–319 2. Stroud F (2018) Iot-Internet of things. Lecture notes in Iot. Webopedia 3. Talari S, Shafie-Khah M, Siano P, Loia V, Tommasett A, Catalão JPS (2017) A review of smart cities based on the Internet of things concept. MDPI, Open Access J 4. Frangos J-M (2017) The Internet of things will power the fourth industrial revolution annual meeting of the new champions. World Economic Forum 5. Nedeltchev P (2015) The Internet of everything is the new economy. COSCO 6. Ranger S (2018) What is the IoT? Everything you need to know about the Internet of things right now, ZDNet. Cybersecurity in an IoT and mobile world 7. Panaseyko E (2018) Internet technologies and communication protocols. My Assignment Help 8. Mulligan G (2007) The 6LoWPAN architecture, semantic scholar. Allen Institute for Artificial Intelligence 9. Farej KZ, Abdul-Hameed AM (2015) Performance comparison among (Star, Tree and Mesh) topologies for large scale WSN based IEEE 802.15.4 standard. Int J Comput Appl 10. Vasseur JP, Agarwal N, Hui J, Shelby Z, Bertrand P, Chauvenet C (2011) RPL: The IP routing protocol designed for low power and lossy networks. Internet Protocol for Smart Objects (IPSO). Alliance 11. Khan MR (2012) Performance and route stability analysis of RPL protocol. Masters’ Degree Project Stockholm, Sweden 12. Chen X (2012) Constrained application protocol for Internet of things, research paper on computer science. Washington University 13. Teklemariam GK, Hoebeke J, Moerman I, Demeester P (2013) Facilitating the creation of IoT applications through conditional observations in CoAP. EURASIP J Wirel Commun Netw 14. Mulligan G (2007) The 6LoWPAN architecture. In: 4th Workshop on embedded networked sensors 15. Mehta S, Sultana N, Kwak KS (2010) Network and system simulation tools for next generation networks: a case study, book of modelling, simulation and identification

Chapter 29

Algorithms for String Comparison in DNA Sequences Dhiman Goswami , Nishat Sultana

and Warda Ruheen Bristi

1 Introduction In context of working in the diversified field of computational biology, string comparison is a key concept applied to determine different properties of DNA or genome. Here three proposed improvements of the existing three approaches is discussed and named as SSAHA for exact or partial match finding in sequences of DNA database, finding MUMmer using BWT and FM index and improvement in the tool Sibelia to find synteny block. SSAHA (Sequence Search and Alignment by Hashing Algorithm) [9], for performing efficient searching procedure on very large DNA database, sequences are preprocessed by dividing them consecutively in k-tuples of bases. Hash table method is used for the purpose of storing the positions of each k-tuple occurred. We search in the database of query sequences by retriving “hits” for each query sequence’s every k-tuple from the hash table and then find the sorted order of the results. Tuple length k has the impact on the speed of searching, usage of memory and algorithm’s sophistication. A dynamic programming approach is proposed to solve this problem which will give faster output for lower k-mers. D. Goswami (B) · N. Sultana · W. Ruheen Bristi Department of Computer Science and Engineering, Daffodil International University, Dhaka 1207, Bangladesh e-mail: [email protected] N. Sultana e-mail: [email protected] W. Ruheen Bristi e-mail: [email protected] D. Goswami · N. Sultana · W. Ruheen Bristi Department of Computer Science and Engineering, Bangladesh University of Engineering and Technology, Dhaka 1000, Bangladesh © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_29

327

328

D. Goswami et al.

MUMs are the Maximal Unique Matches between two genomes of different organisms/species. It shows the common sub-sequences between two genomes and the sub-sequences must be unique in both the genomes. That is each MUM can be only once in each genome, as it is unique. Many significant algorithms are there to align two genomes. For example, Needleman and Wunsch (Global alignment) [8] and Smith and Waterman (Local alignment) [11] algorithms. In this paper, we propose an approach using BWT matrix [1] and FM index [5] to find MUMs by comparing two genomes by ensuring the uniqueness of the MUMs. Synteny Block finding which is the undertaking of breaking down genomes into non-covering profoundly rationed portions has turned out to be essential in genome correlation. The motive is to explore synteny blocks.The genome sequences are closely related. Exploring blocks from various microbial genome sequences has some obstacles as they have less change in their characteristics. A general technique of building or generating graph from the existing sequence is using the de Bruijn graph. But as the real genome sequence has many hidden portions which can hide many characteristics. In this case, using the general version of de Bruijn graph could not capture those characteristics so in that case iterative de Bruijn graph is about to be implemented. As duplications are imprecise, de Bruijn graph structure needed an introductory stride of graph simplification to identify the accord of duplications, and next, fortifying the genome over the simplified graph to obtain the points of duplications.

2 Related Works 2.1 SSAHA Formal Definition To determine the complete or specific portion’s match of a query sequence Q in a subject sequences’ database D = {S1 , S2 , . . . , Sd }, where all the sequences in D is marked with a specific value i which is an integer used for the reference of index, the term k-tuple is used to represent a consecutive sequence of DNA nucleotides of length k. A DNA sequence S having length n includes (n–k+1) k-tuples [9] which overlaps in different positions. All the k-tuple’s offsets within S is the positional difference as regards to the first base of S. We use w j (S) to express the k-tuple of S having offset value j. Thus, the location within D of every instance of all k-tuple is demonstrated by (i, j). Four types of bases can be mapped with two binary digits which can be shown in the following way: f (A) = 002 , f (C) = 012 , f (G) = 102 , f (T ) = 112 . In this mapping procedure, 2k bit integer is needed to uniquely map any k-tuple w = b1 b2 . . . bk . So from the above binary representation weget the hash table position of a particular k-tuple by the following function: E(w) k 4i−1 f (bi ). = i=1 Hash Table Construction For example, we have database where we have three elements S1 , S2 , and S3 . Suppose, S1 = ACC GT T GT AGC T A, S2 = ACC T T GT T

A C G T

0 1 2 3

(1,0) (1,1) (1,3) (1,4)

(1,8) (1,2) (1,6) (1,5)

(1,12) (1,10) (1,9) (1,7)

Table 1 Hash table for k-mer value 1 W E(W) Positions (2,0) (2,1) (2,5) (1,11)

(2,10) (2,2) (2,12) (2,3)

(2,13) (2,8) (3,1) (2,4)

(3,10) (2,14) (3,4) (2,6)

(3,11) (2,15) (3,7) (2,7)

(3,12) (3,0) (3,14) (2,9) (3,6) (3,15) (2,11)

(3,2)

(3,3)

(3,5)

(3,8)

(3,13)

29 Algorithms for String Comparison in DNA Sequences 329

330

D. Goswami et al.

Table 2 Hash table query sequence for k = 1 t Wt(Q) Positions 0

A

1

A

2

A

(1,0) (1,8) (1,12) (3,12) (2,0) (2,10) (2,13) (3,10) (3,11) (1,0) (3,12) (1,8) (1,12) (2,0) (2,10) (2,13) (3,10) (3,11) (1,0) (1,8) (1,12) (2,0) (2,10) (2,13) (3,10) (3,11) (3,12)

H

M

(1,0,0) (1,8,8) (1,12,12) (3,11,12) (2,0,0) (2,10,10) (2,13,13) (3,10,10) (3,11,11) (1,−1,0) (3,12,12) (1,7,8) (1,11,12) (2,−1,0) (2,9,10) (2,12,13) (3,9,10) (3,10,11) (1,−2,0) (1,6,8) (1,10,12) (2,−2,0) (2,8,10) (2,11,13) (3,8,10) (3,9,11) (3,10,12)

(1,−2,0) (1,−1,0) (1,0,0) (1,6,8) (1,7,8) (1,8,8) (1,10,12) (1,11,12) (1,12,12) (2,−2,0) (2,−1,0) (2,0,0) (2,8,10) (2,9,10) (2,10,10) (2,11,13) (2,12,13) (2,13,13) (3,8,10) (3,9,10) (3,9,11) (3,10,10) (3,10,11) (3,10,12) (3,11,11) (3,11,12) (3,12,12)

C T AT G ACC and S3 = C GT T GT C GT C A A AT GG and we have a Target Sequence: A A A. So the hash table for the database for k-mer size equal to 1 (Table 1). Now we observe the hash table of the query sequence in the following table and there the solution can be found from the bold indices (Table 2). Now the approach is extended to k-mer value equal to 2 and the database is S1 = ACC GT T GT AGC T , S2 = ACC T T GT T C T AT G ACC and S3 = C GT T GT C GT C A A AT GG and Target Sequence is A A AT GG. Hash table for the database for this case is (Table 3). After that, we observe the hash table of the query sequence in the following table and there the solution can be found from the bold indices (Table 4).

29 Algorithms for String Comparison in DNA Sequences Table 3 Hash table for k-mer value 2

Table 4 Hash table query sequence for k = 2

331

W

E(W)

Positions

AA AC AG AT CA CC CG CT GA GC GG GT TA TC TG TT

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

(3,11) (1,1) (1,9) (2,11)

t

(2,1) (3,13)

(2,15) (1,3) (1,11) (2,13)

(3,1) (2,3) (1,8)

(3,15) (1,7)

(3,7)

(3,9) (2,5) (1,5)

(2,7)

(3,3)

Wt(Q)

Positions

H

M

0 1 2

AA AA AT

3 4

TG GG

(3,11) (3,11) (2,11) (3,13) (2,5) (3,15)

(3,11,11) (3,10,11) (2,9,11) (3,11,13) (2,2,5) (3,11,15)

(2,2,5) (2,9,11) (3,10,11) (3,11,11) (3,11,13) (3,11,15)

(3,7) (2,9) (1,12)

2.2 MUMs Finding An established system is available to align the whole genome [3]. Two sequences as genome A and genome B act as input to the system. At first MUMmer is found out by suffix tree structure. All the suffixes are found out from both the genomes A and B where genome A is “gaacctca” and genome B is“aaaaccgtt”, for example. Suffix tree is drawn from the suffixes of genome A. Later on suffixes of genome B is added to that tree. Figure 1 shows the suffix tree of genome A and genome B. By doing this the portion of suffixes which is common to both the genomes, MUMs, could easily be sorted out. In Fig. 1, red colored alphabets indicate the genome from which the string comes. After finding the MUMs, those are aligned by a variation of LIS algorithm. Lastly gaps among the MUMs are closed by four different processes by identifying many biological features like SNPs, large insertions, repeats, tandem repeats, etc. In this way, the whole genome is aligned (Table 5).

332

D. Goswami et al.

Fig. 1 Suffix tree on genome A and B Table 5 Suffices of genome A and genome B

Genome A

Genome B

gaacctca aacctca acctca cctca ctca tca ca a

aaaaccgtt aaaccgtt aaccgtt accgtt ccgtt cgtt gtt tt t

2.3 Synteny Block Finding Cycles in de Bruijn Graph Given String S = s1 . . . sn is roundabout which is a genome of length n. It is roundabout over the nucleotide letters in order A,G,C,T. A string having length k is denoted as k-mer. The de Bruijn graph diagram DB(S,k) speaks to each node in the graph that is defined as per the k-mer and associates both nodes by a coordinated arc on the off chance that they are compared to a couple of back to back k-mers in the genome sequence. The de Bruijn graph observed as a weighted multigraph where adjoining nodes can be associated by various arcs and with the assortment of an edge (a,b) defined as the occasions that the k-mers show up continuously in S. Cycles in de Bruijn diagrams are normally classified into two kinds [7]: swells (Fig. 2) and circles (Fig. 2). Instinctively, swells are formed by befuddles/indels between two homologous successions, and circles are caused by firmly found k-mer rehashes. To uncover rehashes in de Bruijn charts, these little cycles ought to be expelled. To maintain a strategic distance from the threading methodology or in other words, we receive a succession alteration way to deal

29 Algorithms for String Comparison in DNA Sequences

333

(a) (b)

(c) (d)

(e) (f)

Fig. 2 Figure containing genome sequence de Bruijn graph generated from the sequence and cycles in the de Bruijn graph. a Genome sequence S = TAGCCAATTGT...TAGCCAATTGT. b The two branches of the de Bruijn graph is drawn using different color (red and blue) using the sequence provided in a. c A sequence having two different streams of strings the dissimilarities are underlined. d Having two different branching De Bruijn graph of the sequence in c. Small differences between these two portions (shaded red and blue) are reflected by a bulge with two branches (the red and the blue parts of the bulge). e An arrangement with a firmly found rehashed k-mer TAG. f The de Bruijn diagram of the succession in e. The firmly found rehashed k-mer (TAG) shapes a circle in the chart

with expel swells in the de Bruijn diagrams. In the following subsection, it will be discussed that Sibelia does not have to expressly expel circles, but rather concentrates just on bulges. The purpose for this is only expanding the estimation of k can dispense with little circles in the de Bruijn diagram. Figure 2c demonstrates a circle or in other words a firmly found 3-mer rehash (TAG). The de Bruijn graph diagram having greater vertex measure k = 4 does not produce a loop. Sequence Modification algorithm Give C a chance to be a bulge with aggregate edge number smaller than C, framed by substrings P1 and P2 of S. To evacuate the bulge, the calculation changes S by substituting all events of P1 in S by P2 . Figure 3 shows the de Bruijn graph for S = TAGCCAATTGT...TAGCGAGTTGT, with two repeats with smaller difference. Minor differences between these two recurrent occasions frame a bulge having two branches (red and blue in the figure). By changing S into S0 = TAGCCAATTGT...TAGCGAGTTGT (i.e. selecting the red branch in lieu of the blue branch), the bulge is additionally simplified. Note that S0 currently has a correct rehash of multiplicity 2. Algorithm In Sibelia [7], Algorithm 1 is used to find synteny block. At first, the calculation builds the de Bruijn chart from a generally little estimation of k = k0 and performs diagram simplification with a little cycle length limit (C0 ). The calculation works on the de Bruijn diagram G 0 (S0 , k0 ) and simplifies all bulges utilizing the grouping modification approach depicted previously. Accordingly, we get a simplified de Bruijn chart G1 and the comparing modified genome S1 . S1 is a contorted adaptation of S with the end goal that its de Bruijn chart D B(S1 , k0 ) does not include short bulges of length tinier than C0 . We take note of that chart simplification ought

334

D. Goswami et al.

(a)

(b)

(c)

(d) Fig. 3 a De Bruijn diagram of a grouping with two estimated rehashes S = TAGCCAATTGT...TAGCGAGTTGT. The minor differences in the inaccurate rehashes shape a bulge having two branches: The red branch (AGC) → (GCC) → (CCA) → (CAA) → (AAT) → (ATT) → (TTG) and the blue branch (AGC) → (GCG) → (CGA) → (GAG) → (AGT) → (GTT) → (TTG). b We improve the diagram by changing the grouping from TAGCCAATTGT to TAGCGAGTTGT, therefore shaping a correct rehash. The modified arrangement compares to a non-fanning way on the de Bruijn diagram. c A firmly found rehashed k-mer (ATC) frames a circle in the diagram. d Enhancing the estimation of k can resolve the circle in c

to be connected utilizing the succession modification calculation; generally, S1 , the twisted adaptation of S0 , is not accessible for the development of the diagram utilizing a bigger estimation of k = k1 . The objective of the first cycle is to crumple swells caused by single point changes or tiny indels. Subsequently, we can build the estimation of k and develop another de Bruijn diagram G 1 = D B(S1 , k1 ), where k1 > k0 . The procedure proceeds until the point that we achieve an estimation of k that is sufficiently extensive to uncover huge scale synteny blocks (Algorithm 1). As a rule, the iterative procedure should proceed until the point when the genome is displayed as a solitary synteny block. This contention may seem irrational at first locate, as our objective was to break down genomes into synteny blocks. In any case, anyone can see here two perceptions shown in the above section that helps us to follow our means to find past synteny blocks. Result: St , Graph t Input: Genome Sequences G procedure I T E R AT I V E D E B RU I J N (G, ((k0 , C0 ), ..., (kt , Ct ))) S0 ← Concatenate(G) Asser t (k0 < k1 < ... < kt ) Asser t (C0 < C1 < ... < Ct ) i ←0 while i < t do Graph i ← Constr uct DeBr ui jnGraph(ki , Si ) Si ← Simpli f y Bulges Smaller T hanC(Graph i , Ci ) i ←i +1 end

Algorithm 1: Iterative de Bruijn Graph (Existing algorithm)

29 Algorithms for String Comparison in DNA Sequences

335

3 Methods 3.1 SSAHA Improvement by Dynamic Programming An improved dynamic programming approach for lower k-mer is proposed here. For this, we have to construct a two-dimensional array of size Quer y SequenceLength ∗ Subject SequenceLength. Improved Algorithm Recurrence relation of the algorithm is given below: Array[i][j]=Array[i-1][j-1]+1 ; row[i]==col[j] Array[i][j]=Array[i-1][j-1] ; row[i]!=col[j] Result: Array[Quer y SequenceLength][Subject SequenceLength] Input: Maximum V alue f r om Last Row or Column procedure H AS H T AB L E D P(Quer y Sequence , Subject Sequence) for i ← 0 to Subject SequenceLength − 1 do Array[0][i] ← 0 end for j ← 0 to Quer y SequenceLength − 1 do Array[ j][0] ← 0 end for i ← 0 to Subject SequenceLength − 1 do for j ← 0 to Quer y SequenceLength − 1 do if Quer y Sequence[i] == Subject Sequence[ j] then Array[i][ j] = Array[i − 1][ j − 1] + 1 else Array[i][ j] = Array[i − 1][ j − 1] end end end Algorithm 2: Hash Table Construction by Dynamic Programming DP table construction Suppose we have three sequences in a database D. They are ACCGTTGTAGCT, ACCTTGTTCTATGACC, and CGTTGTCGTCAAATGG and a query sequence is AAATGG. DP table for them are shown in the following three tables (Tables 6, 7 and 8). DP table analysis We have to get the highest value of last row and last column of each table. Highest Values from Table 1 is 2, from Table 2 is 3, and from Table 3 is 6. Now we have to get the highest value from the highest value of each table. Here we can see the value is 6 that means we have an alignment of Query sequence of length 6 to the 3rd subject sequence. Index finding from DP table Suppose, Query Length is m (column wise) and Subject Length is n (row wise). If the highest value is from ith index of last row then alignment value for that index will be (i-m). If highest value is from ith index of last column

336

D. Goswami et al.

Table 6 DP Table 01 A C 0 0 0 0 0 0 0

A A A T G G

0 1 1 1 0 0 0

0 0 1 1 1 0 0

Table 7 DP Table 02 A C C

C

G

T

T

G

T

A

G

C

T

0 0 0 1 1 1 0

0 0 0 0 1 2 2

0 0 0 0 1 1 2

0 0 0 0 1 1 1

0 0 0 0 0 2 2

0 0 0 0 1 0 2

0 1 1 1 0 1 0

0 0 1 1 1 1 2

0 0 0 1 1 1 1

0 0 0 0 2 1 1

T

T

G

T

T

C

T

A

T

G

A

A

C

0 0 0 1 1 1 0

0 0 0 0 2 1 1

0 0 0 0 1 2 1

0 0 0 0 0 2 3

0 0 0 0 1 0 2

0 0 0 0 1 1 0

0 0 0 0 0 1 1

0 0 0 0 1 0 1

0 1 1 1 0 1 0

0 0 1 1 2 0 1

0 0 0 1 1 3 1

0 1 1 1 1 1 3

0 0 1 1 1 1 1

0 0 0 1 1 1 1

Table 8 DP Table 03 C G T

T

G

T

C

G

T

C

A

A

A

T

G

G

0 0 0 0 1 1 0

0 0 0 0 0 2 2

0 0 0 0 1 0 2

0 0 0 0 0 1 0

0 0 0 0 0 1 2

0 0 0 0 1 0 1

0 0 0 0 0 1 0

0 1 1 1 0 0 1

0 1 2 2 1 0 0

0 1 2 3 2 1 0

0 0 1 2 4 2 1

0 0 0 1 2 5 3

0 0 0 0 1 3 6

A A A T G G

A A A T G G

0 0 0 0 0 0 0

0 0 0 0 0 0 0

0 1 1 1 0 0 0

0 0 0 0 0 0 0

0 0 1 1 1 0 0

0 0 0 0 0 1 1

0 0 0 0 1 0 1

then alignment value for that index will be (n-i). So from the above three DP tables the alignment can be found from the following table (Table 9). So from the subject sequences in the database, we get that the if we align the Query sequence to the 10th index of the third subject sequence we will get the maximum match of 6 characters.

29 Algorithms for String Comparison in DNA Sequences

337

Table 9 Result analysis Database Parameters Highest value ith index from column ith index from row Alignments from column (n-i) Alignments from row (i-m)

Table 1 2 4 4, 5, 7, 8, 10 8 −2,−1, 1, 2, 4

Table 2 3 6, 14 0, 8

Table 3 6 6 16 10 10

3.2 Improved MUMs Finding Using BWT Matrix and FM Index Existing algorithm of MUMs finding uses suffix tree of one genome to thread the suffices of other genomes to that tree. But using suffix tree an unique match cannot be always identified. For example, Fig. 4 shows the suffix tree of genome A and genome B which are caatacc and gtaccgtacc consecutively. The suffix tree shows that the string “tacc” is one of the MUMs as it is a maximal match between the genomes. But we can see that “tacc” is not unique in this set of genomes. Genome A contains the substring “tacc” twice, “gtaccgtacc”. So obviously “tacc” cannot be an MUM, as it is not unique (Table 10). In our approach, Burrows wheeler transform (BWT) matrix is used instead of that threading procedure. In case of BWT, the first and last column is needed to find whether the sequence exists in the other one or not. So without constructing suffix tree, BWT of one genome and can check the suffixes of other genome using FM

Fig. 4 Suffix tree on genome A and B

338 Table 10 Suffices of genome A and genome B

D. Goswami et al. Genome A

Genome B

caatacc aatacc atacc tacc acc cc c

gtaccgtacc taccgtacc accgtacc ccgtacc cgtacc gtacc tacc acc cc c

index to find maximal unique matches. Generally, MUMs of very short length must not be reported. So along with the genomes, a threshold value to set the minimum base pairs in the MUM alignment is provided as input parameter. As the MUMs can overlap, so the algorithm checks and compares the genomes by deleting only one character from last position each time. The algorithm is given below. Result: mum Storage Input: Genome A, Genome B, T hr eshold procedure MU M f inding(Genome A, GenomeB, thr eshold) S ← Genome B L ← BW T o f Genome A mum Storage ← ∅ Give each character in L a number, equal to the number o f occurr ence o f that character consecutively starting f rom zer o F ← Sor ted BW T o f Genome A while length(S) >= thr eshold do match ← ∅ match ← all matched substrings o f S f ound by modi f ied F M index with L and F if length(match) == 1 and length(matched Element) >= thr eshold then if mumStorage does not contain match then mum Storage ← match else delete match f r om mumStorage end end delete the last character o f string S end Algorithm 3: MUM finding by BWT and FM index

29 Algorithms for String Comparison in DNA Sequences

339

3.3 Improved Synteny Block Detection Method Synteny Block loop deletion of length ≤ bulge length In the existing method, it is considered that all the bulges are smaller than the threshold loop. But in the reallife genome sequence there can be many variations. There can be loop sizes that are smaller than the bulge or equal to the size of the bulge as the algorithm only concerns about the bulges smaller than the loop so there is a scope of improvement. Examples of genomes are given in Fig. 5 that contains the variation of the size of loop and bulge. Improved Algorithm In the improved algorithm, it is ensured that a bulge will be handled after the deletion of all the loops smaller or equal to the size of that particular bulge. The algorithm is given below. Result: St , Graph t Input: Genome Sequences G procedure I T E R AT I V E D E B RU I J N (G, ((k0 , C0 ), ..., (kt , Ct ))) S0 ← Concatenate(G) Asser t (k0 < k1 < ... < kt ) Asser t (C0 < C1 < ... < Ct ) i ←0 while i < t do Graph i ← Constr uct DeBr ui jnGraph(ki , Si ) LoopLength ← Find LoopLengthU sing D F S(Graph i ) BulgeLength ← Find BulgeLength(Graph i ) while LoopLength ≤ BulgeLength do i ←i +1 Graph i ← Constr uct DeBr ui jnGraph(ki , Si ) LoopLength ← Find LoopLengthU sing D F S(Graph i ) BulgeLength ← Find BulgeLength(Graph i ) end Si ← Simpli f y Bulges Smaller T hanC(Graph i , Ci ) i ←i +1 end Algorithm 4: Iterative de Bruijn Graph (Improved Algorithm)

4 Results 4.1 Statistical Analysis of Hash Table Verus DP For the k-mers upto length 2, the dynamic programming approach triumphs over the hash table method. But from k-mer of length 3 and above the hash table method

340

D. Goswami et al.

Fig. 5 Loop created in the de Bruijn graph is (i) smaller and (ii) equal to bulge size Table 11 Statistical analysis of dynamic programming

k-mer length

Query length

DP

1 2 3 1 2 3

1000 1000 1000 500 500 500

93.18% faster 39.28% faster 75.21% slower 92.80% faster 34.37% faster 77.3% slower

is more improved in terms of speed though runtime complexity of both algorithms is the same. For a database of 1000 sequences each having length 1000, the comparison is given below (Table 11).

4.2 MUM Finding: Suffix Tree Versus BWT and FM Index Existing approach of MUM finding fails to detect whether the MUMmer is unique or not in any of the two genomes that are to be aligned. This proposed approach will maintain the uniqueness of the MUMmers. If any repetitive maximal match is found greater than or equal the threshold value then the match is discarded. Statistically for smaller length genome, the proposed approach gives comparatively faster result. But with the increase of genome length, the speed slows down. Though the time complexity for both the methods is same (Table 12).

29 Algorithms for String Comparison in DNA Sequences

341

Table 12 Statistical analysis of MUM finding by BWT and FM index

Length of genome A Length of genome B Improvement

Table 13 Tabular analysis of synteny block generation with genome sequence

Sequence

Loop and bulge size

i ii

Loop = 0 Bulge = 5 Can handle Can handle Loop = 3 Bulge = 3 Can’t Can handle handle Loop = 3 Bulge = 5 Can’t Can handle handle

500 1000 1500 2000

iii

500 1000 1500 2000

16.21% faster 14.02% faster 11.47% faster 10.63% faster

Sibelia

Proposed method

4.3 Synteny Block: Detection Versus Avoidance of Loop ≤ Bulge There is a potential scope of improvement in that case if the bulge larger or equal in accordance to the loop. In that cases loop has to be detected than the determination of the length of the loop has to be done and the length of the bulge and the size comparison of the loop and bulge will decide the computation that whether the size of the K-mer has to be increased or the Sequence Modification Algorithm has to be used.We have proposed to detect the loop using Depth-First Search as it has linear time complexity. We find the loops in lexicographical order. If discovered cases are not resolved than there can be branching paths in the synteny block which will not lead to the proper comparison of microbial genomes. Analysis over three genomes (i) AT C GGT T A AC T...AT C G AT C A AC T, (ii) A AT C AT C T C A A...A AT C GT C T C A A and (iii) A AT C AT C T C A A...A AT C GT GT C A A to generate synteny block is shown below in Table 13.

5 Time and Space Complexity Analysis 5.1 Dynamic Programming Approach for SSAHA In the dynamic programming approach, the main task is to construct the table which takes O(mn) where m and n is the length of query sequence and subject sequence, respectively. Moreover, finding out the index of alignment can be done in linear time by finding the highest value from last row and column. So overall time complexity

342

D. Goswami et al.

will be O(mn). To calculate a specific index of the array we just need the upper diagonal value of that index. So the space complexity will be minimum of O(m) and O(n).

5.2 MUM Finding by BWT and FM Index Construction of Suffix tree needs linear time and space complexity by using pointers efficiently [2, 6, 12, 13]. On the other hand building BWT and FM index and working with those requires the same linear time and space complexity.

5.3 Synteny Block : Detection of Loop and Bulge In the actual version of Sibelia, the time complexity is mainly dependent on the construction of de Bruijn graph and finding the length of bulges. Linear time algorithms are available to construct de Bruijn graph [4, 10]. The complexity to find bulges smaller than that of the threshold loop can be also done in linear time. In this improved algorithm it is suggested that the Depth-First Search (DFS) algorithm to find out the size of the loop which can be done in linear time with the complexity of O(V+E), where V is the number of vertex and E is the number of edges. So overall time complexity will be linear.

6 Discussion In the diversified field of string comparison, three different issues are described and better approaches are shown here. At first, a dynamic approach is proposed against hash table algorithm of DNA database which shows tremendous improvement of runtime (though same in complexity) for smaller k-mer. Then an improved way of ensuring uniqueness of MUMs is described using BWT matrix and FM indexing instead of existing suffix tree method where repetitive MUMs are found unexpectedly. Last of all, a major discrepancy of avoiding the loops not greater than bulge size to find synteny block by Sibelia is determined and improved algorithm is proposed. Thus, the aim of improving some existing string algorithms in the field of computational biology is expected to be served with flying colors.

References 1. Burrows M, Wheeler DJ (1994) A block sorting lossless data compression algorithm 2. Chen MT, Seiferas J (1985) Combinatorial algorithms on words. Springer, New York, pp 97– 107

29 Algorithms for String Comparison in DNA Sequences

343

3. Delcher AL, Kasif S, Fleischmann RD, Peterson J, White O, Salzberg SL (1999) Alignment of whole genome 4. Dinh H, Kundeti KV, Rajasekaran S, Thapar V, Vaughn M Efficient parallel and out of core algorithms for constructing large bi-directed de Bruijn graphs. Department of computer science & engineering, University of Connecticut, 371 Fairfield way, U-2155, Storrs, CT, 06269, USA 5. Ferragina P, Manzini G (2000) Opportunistic data structures with applications. Proc IEEE FOCS 390–398 6. McCreight EM (1976) J ACM 23:262–272 7. Minkin I, Patel A, Kolmogorov M, Vyahhi N, Pham S (2013) Sibelia: a scalable and comprehensive synteny block generation tool for closely related microbial genomes. Department of computer science and engineering, UCSD, La Jolla, CA, USA. St. Petersburg Academic University, St. Petersburg, Russia 8. Needleman SB, Wunsch CD (1970) J Mol Biol 48:443–453 9. Ning Z, Cox AJ, Mullikin JC (2001) SSAHA: a fast search method for large DNA databases. Informatics division, The Sanger centre, Wellcome trust genome campus, Hinxton, Cambridge CB10 1SA, UK 10. Rahman MS, Rahman MM, Sharker R (2017) HaVec: an efficient de Bruijn graph construction algorithm for genome assembly. Depatment of CSE BUET, ECE Building, West Palashi, Dhaka1205, Bangladesh 11. Smith TF, Waterman MS (1981) J Mol Biol 147:195197 12. Ukkonen E (1995) Algorithmica 14:249–260 13. Weiner P (1973) Linear pattern matching algorithms. In: Proceedings of the 14th IEEE symposium on switching and automata theory, p 111

Chapter 30

A New Approach for Efficient Face Detection Using BPV Algorithm Based on Mathematical Modeling Tangina Sultana, M. Delowar Hossain, Niamul Hasan Zead, Nur Alam Sarker and Jannatul Fardoush

1 Introduction Face detection can be defined as a computer technology that is used in a variety of applications that identifies human faces in digital images. It can also be regarded as a specific case of object-class detection. The term “object-class detection” refers the task to find the locations and sizes of all objects in an image that belong to a given class. The applications of face detection are face recognition, facial expression recognition, face tracking, facial feature extraction, gender classification, identification system, biometric system, human computer interaction system and so on. In this paper feature-based approaches are used which are dealing with face detection of individual facial features and their geometrical relationships. These features generally include the eyes, nose, and mouth. The main objective of this paper is to find and describe an efficient algorithm on the basis of mathematical model for recognizing a face on an input image. Every person varies from each other and also there are many factors such as face proportions, deformations, expressions, captured angle etc. It is reachable to use an algorithm that is defined by decisive rules but another requirement is to develop a lightweight algorithm, that would be enough robust to provide very high detection rate which is known as true-positive rate as well as very low false positive rates. There T. Sultana (B) · N. H. Zead · N. A. Sarker · J. Fardoush Department of Electronics and Communication Engineering, Hajee Mohammad Danesh Science and Technology University, Dinajpur, Bangladesh e-mail: [email protected] N. H. Zead e-mail: [email protected] M. D. Hossain Department of Computer Science and Engineering, Hajee Mohammad Danesh Science and Technology University, Dinajpur, Bangladesh e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_30

345

346

T. Sultana et al.

are several methods that can be used for face detection, such as Viola-Jones method, Local Binary pattern (LBP), CamShift, Adaboost etc. Many research activities has been done to find out the best face detection algorithm. Previous researches on face recognition falls into two main categories: feature-based and holistic. It is found on literature survey that face recognition is done by measuring the distance between the eyes and shapes of lines which are connecting the facial features. On the other hand, holistic methods are well known approach of Eigen-faces. Face detection is the basic element of face recognition as well as face tracking system and also has been implemented in several applications in many areas such as image capture, video coding, video conferencing and crowd control in a public place [1]. Among various face detection algorithms Viola Jones algorithm produces performance comparable to previous systems. This algorithm is also implemented on a conventional desktop with face detection of 15 frames per second [2]. Another algorithm that was used to detect faces is the Haar cascade method. This algorithm is used both face and vehicle detection. It is also found that Haar cascade classifiers were used in object detection in ultrasound images [3] as well as the detector was trained by AdaBoost algorithm. However, Local Binary Pattern (LBP) algorithm was proposed for texture analysis and applications of image processing and computer vision [4]. But the main problem of these paper is that there is no comparison shown among the only face detection algorithms in any research paper thus we can’t know which one is the most efficient algorithm to detect the face. That’s why in this paper, it is tried to analysis and compare all the face detection algorithms based on the mathematical model so that we can give a concrete suggestion about which face detection algorithm performs the best result to detect the face. Our main contributions of this paper are summarized as follows: (a) Find out several face detection algorithms which are widely used in the field of face detection and tracking system. (b) Performance analysis of these face detection algorithms based on mathematical modeling. (c) Compare the algorithms on the basis of mathematical model and try to find out the most efficient algorithm among them. (d) Propose a new algorithm along with its mathematical modeling which could be more efficient in practical system. The rest of the paper is organized as follows. Section 2 describe the mathematical analysis of the different face detection algorithms. Our proposed method is illustrated in Sect. 3. In Sect. 4, performance analysis is given. Finally, conclusions are presented in Sect. 5.

2 Mathematical Analysis of the Face Detection Algorithms 2.1 Camshift Algorithm Camshift algorithm is an object tracking algorithm which is a modified version of Mean-Shift tracking method. Mean-Shift method is a herculean nonparametric technique for finding the mode in a probability distribution. Mean-Shift algorithm is

30 A New Approach for Efficient Face Detection …

347

modified in Camshift algorithm in such a way so that it can deal with dynamically changing color probability distribution which is taken from a video frame [5, 6]. The methodology of Camshift algorithm can be summarized with following steps. 1. 2. 3. 4.

It chooses the initial region of interest which contains the object we want to track. It makes a color histogram of that region as the object model. A probability distribution of the frame is created using the color histogram. This algorithm is based on the probability distribution image which is used to find the center mass of the search window using mean-shift method. 5. It centers the search window to the point taken from step 4 and iterate step 4 until convergence. 6. It processes the next frame with the search window position from the step 5. It is easy to understand that the background information is important for target tracking. Let âu is the discrete representation (histogram) of the background in the feature space. We can define the background histogram, which is a discrete unweighted representation of a significant region outside the target region [7]: m    Background : aˆ = aˆ u 1...m , aˆ u = 1

(1)

u=1

We use the standard derivation of the weight where â* is the smallest non-zero entry selected from the background model.   ∗  aˆ Feature weights : au = min aˆ u u=1...m

(2)

The new target model representation is then defined by. q = N au

n



    2 k

xi∗

δ b xi∗ − u

(3)

i=1

With the normalization constant N expressed as 1



   ∗ 2 m

x ∗

i=1 k u=1 au δ b x i − u i

N= n

(4)

The new target representation is 





y − xi

2



δ[b(xi ) − u] Tu (y) = Nh au k



h i=1 nh 

(5)

348

T. Sultana et al.

where now N h is given by 1 



 ∗ 2 m

y−xi

i=1 k u=1 au δ b x i − u h

Nh =  nh

(6)

There are some limitations of Camshift algorithm that it cannot track the face properly with similar color background and also tracking fails in varying illuminations. It also faces problem to track multi hue object.

2.2 AdaBoost Algorithm AdaBoost stands for Adaptive Boosting which is a machine learning meta- algorithm. This algorithm can be used in synchronism with many other types of learning algorithms to improve performance. Furthermore it is not affected by multi hue and illumination variance as well as there are two methods of image vector based approaches. These two methods are well known as Eigen faces method and Fisher face method [8, 9]. It is found that the face motion consists of two types of motions. The first one is rigid motion that is formed by the rotatory and trans-rotatory motion of the head or face. This motion just comes after the second one which is called by non-rigid motion. Both the motions are dependent on one another. The non-rigid motion results from the local motion of the face generally relates the contraction of the facial muscles or facial expression. These two motions are mixed together to form the 2-D motion of faces Given training data (x1 , y1 )… (xm , ym ) Yi e{−1, +1}, xi e X is the object or instance, yi is the classification. For t = 1, … ,t we have to crate distribution Dt on {1,…m}. Let us consider a weak classifier with smallest error et on Dt.  et = PrDt ht(xi) = yi and Ht : X→{−1, +1} Then we will get the output single classifier Hfinal(x) 1 m

(7)

Dt (i) c(x) Zt

(8)

e−at : yi = h t (xi ) eat : yi = h t (xi )

(9)

Dt (i) = Given that Dt+1 =  c(x) =

30 A New Approach for Efficient Face Detection …

349

So, Dt+1 =

Dt (i) −at yi h t (xi ) e Zt

(10)

where, Zt is the normalization constant. at =

1 1− ∈t ln >0 2 ∈t

(11)

Hence Hfinal(x) = sign

 

 at h t (x) .

t

But AdaBoost algorithm is sensitive to noisy data and outliers. Moreover, It is less susceptible to the over floating problem than most learning algorithm.

2.3 Local Binary Patterns (LBP) Algorithm A texture analysis base algorithm known as LBP texture analysis operator introduced by Ojala et al. [5]. This algorithm is defined as a grayscale invariant texture measure which can be derived from a general definition of texture in a local neighbor-hood. Local Binary Patterns (LBP) [10, 11] is a simple as well as very efficient texture operator which entitles the pixels of an image by ambitting the neighborhood of each pixel with the value of the center pixel and considers the result as a binary number. Furthermore, the LBP has an adjacent feature that it is not sensitive to noisy data. The original LBP operator forms entitles for the image pixels by ambitting the 3 × 3 neighborhood of each pixel with the center value and considering the result as a binary number. The computation process is very simple i.e. comparing the center pixel (Vc ) of a 3 × 3 pixel with the gray scale of its pixels adjacent to it (Vp ). If the result is greater than or equal to the center pixel count as 1 and if less than (Vc ) it is written as 0. Mathematically the LBP coding is calculated as following. L B P(P, R) =

p−1  

s V p − Vc 2 p p=0



s V p − Vc =



1; V p − Vc ≥ 0 0; V p − Vc < 0

(12)

 (13)

where, P = pixel adjacencies with the center pixel and R = radius. LBP computation process is shown in Fig. 1.

350

T. Sultana et al.

Fig. 1 LBP computation process

This computation starts from p0 = 135 and continues through clockwise direction. Now we get the binary equivalent decimal values using Eq. (12) and the LBP value is 20 + 23 + 24 + 25 = 57 It is a texture value and it is possible that it can be either a face area or a nonface area which effects the efficiency of this algorithm. So, the key problem of this algorithm is that it is a texture base algorithm and hence it needs high quality image and requires long training time.

2.4 Viola-Jones Algorithm Viola-Jones algorithm is a method that is relatively fast, accurate and efficient in the field of face detection system. This algorithm is based on Haar-like features and cascade AdaBoost classifier. It is the first face detection algorithm that is capable of providing real time performance. Many image processing applications which require faces as their input are building using Viola Jones algorithm and it is the most commonly used face detection algorithm. This algorithm mainly consists of three basic steps. In the first step, a Haar classifier is trained using thousands of

30 A New Approach for Efficient Face Detection …

351

Fig. 2 Line and edge haar features Fig. 3 The real values are detected on an image

0.1

0.2

0.6

0.8

0.2

0.3

0.8

0.6

0.2

0.1

0.6

0.8

0.2

0.1

0.8

0.9

positive and negative images which acts as templates to detect faces. Haar classifier training is followed by cascade classifier training which allows for efficient face detection during runtime. Finally, a window is swept across image and its scaled version to detect faces at different locations and sizes. Haar features are the basic element of Viola Jones algorithm. The concept of Haar feature can be described more easily by mathematical model as follows. In the above figure the line and edge Haar Features of an image is shown. The white parts are represented by 0 and the black parts are represented by 1 (Fig. 2). The process of using integral image can be described with mathematical model of an image. Let consider a real image as follows (Fig. 3). Now the comparison will be taken place between the average dark pixels and the average light pixels as below. 0.6 + 0.8 + 0.8 + 0.6 + 0.6 + 0.8 + 0.8 + 0.9 8 0.1 + 0.2 + 0.2 + 0.3 + 0.2 + 0.1 + 0.2 + 0.1 − 8

δ=

(14)

or, δ = 0.74 − 0.18 = 0.58

(15)

Here, δ is the pixel difference between dark and white pixels. The closer the value to 1, the more likely we have found the Haar Features. But in a real image there are thousands of relevant Haar features frames. For example in a 24 × 24 image frame there are 224 = 16,777,216 pixel values. This is very time consuming and the time complexity of this process is O(N2 ). Viola Jones Algorithm uses Integral Image approach to achieve O(1) running time. Computation process of integral image can be defined as mathematical model (Fig. 4).

352

T. Sultana et al.

0.1 0.2 0.2 0.2

0.2 0.3 0.1 0.1

0.6 0.8 0.6 0.8

0.1 0.3 0.5 0.7

0.8 0.6 0.8 0.9

0.3 0.8 1.1 1.4

0.9 2.2 3.1 4.2

1.7 3.6 5.3 7.3

Fig. 4 Input Image and Integrated image Fig. 5 Computation of pixel value

0.1

0.3

0.9

1.7

0.3 0.5 0.7

0.8 1.1 1.4

2.2 3.1 4.2

3.6 5.3 7.3

A pixel value of the integrated image is the sum of pixels on top and to the left pixel values of the Original image (Fig. 5). If we want to calculate the sum of the pixels surrounded by Blue rectangle then we have to calculate only the four rectangles and the formula is Calculated value = (7.3 + 0.3) − (1.7 + 1.4) = 4.5. Since both rectangle 1.4 and 1.7 include rectangle 0.3, the sum of 0.3 has to be added to the calculation. In this way we can achieve O (1) running time for handling Haar Features. So clearly it reduces time complexity. The next stage is to classify the frame windows to detect a face image with the help of cascade classifier. It can be sown by a block diagram (Fig. 6). Cascade classifier is composed of stages each containing a strong classifier. So all features are grouped into several stages which includes certain number of features. When a sub window is found, it immediately discarded as not a face if it fails in any of the stage. So the algorithm should concentrate on discarding non-faces quickly and spend more on time on probable face region. But one of the major problem of Viola Jones is that an integral image integrates both face and non-face pixel values. These redundant values of the integral image sometimes cause complexity. So, in order to overcome the most of the limitations of the above mentioned algorithms, a new method is proposed in this paper which is discussed in the next section.

Input

Stage 1 Is input a face?

May be

Stage 2 Is input a face?

Definitely no

Definitely no

Discard image

Discard image

Fig. 6 Cascading process of Cascade classifier

May be

30 A New Approach for Efficient Face Detection …

353

3 Proposed Method In Viola Jones algorithm an integral image reduces the computation time of the dark pixels and light pixels. It is a great assist of integral image to reduce the run time complexity of Haar features but it has some limitations too. The major problem is that an integral image integrates both face and non-face pixel values. Thus a question about the accuracy of the detection may be raised. Viola Jones algorithm is the breakthrough in the face detection systems but the redundant value of the integral image sometimes causes complexity. In our proposed model Local Binary Pattern (LBP) and integral image value are compared to determine a value (test value) which is equal to the subtraction value between integral image value of a 3 × 3 particular rectangle region and LBP value of that particular rectangle region of the original image. This new value is compared with the Break Point Value (BPV; we have named after this new term). If the test value is less than or equal to the BPV then the selected 3 × 3 rectangle is considered as a relevant feature frame and if it is not, it is considered as a non-relevant frame. Thus it can both reduce the redundancy of the non-face area pixels and can reduce the number of non-relevant features. So the detection can take place with higher accuracy. The proposed method will be easily understood with the help of mathematical model. For the mathematical model let us consider a 5 × 5 original image and the pixel values are as follows (Fig. 7). Now let us consider a 3 × 3 particular rectangle region from both original and integral images and calculate the LBP and BPV values. The BPV value of a 3 × 3 particular rectangle region can be calculated by BPV3∗3 = Max (meanR1 , meanR2 , meanR3 )

(16)

And mean, R=

1 2 2 2 1

2 3 1 1 2

6 8 6 8 6 (a)

8 6 8 9 8

P l r1 + P l r2 + P l r3 3

4 5 5 6 9

Original image

Fig. 7 A 5 × 5 original and an integral image

1 3 5 7 8

(17)

3 8 11 14 17

9 22 31 42 51

17 36 53 73 90

(b) integral image

21 41 58 79 105

354

T. Sultana et al.

1

2

6

8

4

2

3

8

6

5

2

1

6

8

5

2 1

1 2

8 6

9 8

6 9

Threshold=6

0

0

0

0 0

* 0

1 0

LBP=24=16

Fig. 8 Calculation of LBP of a 3 × 3 rectangle

P-lr = Pixel value of the row (left to right). The LBP value can be determined by LBP =

P−1  

S V p − Vc = 2 p P=0



s V p − Vc =



1; V p − Vc > 0 0; V p − Vc ≤ 0

(18)  (19)

And the test value = LBP ∼ integral value

(20)

For computing LBP at first we have to create a binary model of 3 × 3 represented as a blue rectangle in Fig. 8. The computation is done by comparing each original pixel value of that rectangle with the threshold value (here threshold = 6) and according to Eq. (19) if the compared value is less than or equal to zero then it is equivalent to binary zero (0) otherwise it is equivalent to binary one (1). After creating the binary model, we have to arrange a uniform binary pattern in clockwise direction and calculate the LBP value using Eq. (18). Thus we get 00010000 = 0 × 20 + 0 × 21 + 0 × 22 + 0 × 23 + 1 × 24 + 0 × 25 + 0 × 26 + 0 × 27 = 16. A pixel value of the integral image of that blue rectangle (Fig. 9) is the sum of pixel values on top and to the left pixel values of the original image. Thus the integrated image value of the 3 × 3 matrix is 1 + 2 + 6 + 2+ 3 + 8 + 2 + 1 + 6 = 31. The test value is the compared value between LBP and integrated value i.e. |16 − 31| = 15.

Fig. 9 Computation of integral image value of the same region as original image

1 3 5 7 8

3 8 11 14 17

9 22 31 42 51

17 36 53 73 90

21 41 58 79 105

30 A New Approach for Efficient Face Detection …

355

Now we have to compute the BPV to reduce the redundant integral image values and it is done by taking mean value from each row of the selected 3 × 3 matrix. Thus the values for non-face region can be easily rejected before applying boosting method which helps to find out more strong relevant frames. In our consideration, say, the first column belongs to a non-face region and we want to reject this portion to get more relevant frames. Let us find out the BPV using Eqs. (16) and (17) 1+3+9 = 4.33 3 3 + 8 + 22 = 11 meanRlr2 = 3 5 + 11 + 31 = 15.66 meanRlr3 = 3 meanRlr1 =

Now we can see that all mean values reject the first column values very easily and according to Eq. (16) our value is BPV = 15.66. We get our test value 2) problem, the consequent confusion matrix will be of dimension n × n (n > 2). It can be noted that this matrix has n rows, n columns, and n × n entries in total. From this matrix, the frequencies of FPs, FNs, TPs, and TNs need to be computed, which is not directly possible. FP, FN, TP, and TN for class i are computed as per the approach described in [15]: TPi = aii . FPi =

aji .

(11)

aij . j=1,j=i n n  

(12)

FNi = TNi =

n 

(10)

j=1,j=i n 

ajk .

(13)

j=1,j=i k=1,k=i

In continuation of this process, the ultimate confusion matrix of dimension 2 × 2 has the average values of the n confusion matrices for all classes. Accuracy, sensitivity, specificity, precision, false positive rate (FPR) and false negative rate (FNR)—these six performance metrics of a classifier are calculated by using this confusion matrix. After our classifier is trained, by using test data set, performance is computed in index of these metrics. Taking the ultimate confusion matrix into account, accuracy, sensitivity, specificity, precision, FPR, and FNR are calculated in percentage as: TP + TN × 100%. TP + FN + FP + TN TP Sensitivity = × 100%. TP + FN TN Specificity = × 100%.. FP + TN TP Precision = × 100%. TP + FP

Accuracy =

(14) (15) (16) (17)

36 A Comparative Study of Classifiers in the Context of Papaya …

FP × 100%. FP + TN FN FNR = × 100%. FN + TP FPR =

425

(18) (19)

3.2 Description of Diseases and Features Disease study is a much significant portion of our approach since it is not only beneficial for well understanding the defects due to diseases but also giving signs to appropriate features. We have worked with five (5) diseases in this work. The attacks of these diseases often happen all over Bangladesh. They are black spot, powdery mildew, brown spot, phytophthora blight, and anthracnose. Depending on the classifier comparison and disease analysis, a feature set, which is composed of statistical and gray-level co-occurrence matrix (GLCM) features, is chosen in order to recognize the diseases. Since we are in a quest for the comparative analysis of the classifiers in this work, we have selected some distinguishing statistical and GLCM features like mean, standard deviation, variance, kurtosis, skewness, contrast, correlation, energy, entropy, and homogeneity. In this work, five (5) statistical and five (5) GLCM features have been constructed. The reader can read [2] for more information about the diseases and features.

4 Experimental Evaluation and Implementation We perform an in-depth experiment following our approach as described in the previous section (Sect. 3). The start point of our experiment is capturing a papaya image. One hundred and twenty-nine (129) color images of defective and defect-free papayas are collected. There are wide variations in sizes of the captured images. Different-sized images are used in our research by taking the matter into account that various people send images of defective and defect-free papayas to our proposed expert system after capturing them with their own devices. First, an image of size 300 × 300 pixels is converted from the captured image. This size has been selected by considering the handheld device size. Then the images contrast stretching is performed by using the technique of color intensity mapping. Then the image, which has been color-mapped, is divided into more than one region by using k-means clustering segmentation technique. We select this segmentation algorithm, because it outperforms most other off-the-shelf segmentation algorithms as per the claim in [16]. Thus, defective portions are differentiated from defect-free portions. The stepwise changed images are shown for a papaya disease, namely, phytophthora blight in Fig. 3.

426

Md. T. Habib et al.

Fig. 3 a Diseases, b captured image, c image after resizing, d image after contrast enhancement, and e disease-attacked region after segmentation

We divide the whole sample data set into two portions, i.e., training portion and testing portion. We use the holdout method [9] for choosing the data proportion targeted for training and testing separately. We divide the sample data set at the proportion of about two-thirds (84 color images) for training and about one-third (45 color images) for testing. In order for getting rid of the model overfitting problem, a validation set is used. In compliance with this approach, we split the original training data set into two smaller subsets. One subset is for training, whereas the other one is for validation. Two-thirds (56 images) of the training set is fixed for classification model training while the remaining one-third (28 images) is applied for error estimation. The holdout method is repeated five (5) times in order for finding the finally trained classifier. After each execution of the holdout method, the performance of the classification model is calculated with the test set. The fivetimes-found results are averaged in order for constructing multiclass confusion matrix and consequent six binary confusion matrices (since we have six classes). Then the calculation of performance evaluation metrics takes place. We calculate six metrics, namely, accuracy, sensitivity, specificity, precision, FPR, and FNR, in total. We use nine classifiers, namely, C4.5, kNN, logistic regression, naïve Bayes, RIPPER, random forest, SVMs, BPN, and CPN, in total. We set values of all parameters of all these classifiers through the training process. The point by point determinations of these five classifiers are given in Table 1. The results obtained after experimentally evaluating all these classifiers in index of the performance metrics as given in Eqs. (14)–(19) are given in Table 2. We observe from Table 2 that SVMs beats the other eight classifiers in terms of all measurements utilized, where kNN’s execution is the poorest. In case of accuracy, good accuracy is also achieved by CPN, random forest, and BPN. In case of other five metrics, CPN, BPN, and random forest outperform all other classifiers. C4.5 stands in the middle by defeating four classifiers. Taking all six metrics into account, we can assert that SVMs performs the best among all nine classifiers in the context of papaya disease recognition.

36 A Comparative Study of Classifiers in the Context of Papaya …

427

Table 1 Detailed specifications of the nine classifiers used Classifier

Specifications

C4.5

Splitting criterion: Gain ratio, where Gain ratio =

inf o Split Info

=

−

k 

inf o P(vi ) lg P(vi )

i=1

kNN

Distance metric: Euclidean distance k=1

Logistic regression

Model: Multinomial

Naïve Bayes

Gaussian Mean (μ) distriVariance (σ 2 ) bution: Probability density function (f ), where f =

RIPPER

− √ 1 e 2πσ 2

(x−μ)2 2σ 2

 Metric used in rule building: FOIL sinformation gain = p1 lg pt11 − lg pt00 Amount of data used for pruning = 3 Metric used in pruning:

Random forest

p−n p+n

Size of each bag (percentage of size of training data) = 100% Trees maximum depth = unlimited Number of attributes that are randomly chosen = 0

SVMs

Kernel: Linear C =250007

BPN

Network architecture: Fully connected feedforward (10-14-6) Activation function: Sigmoid function, sc (x) =

1 1+e−cx

Learning rate, γ = 0.3 Momentum rate, α = 0.2 CPN

Network architecture: Fully connected feedforward (10-17-6) Layer 1–2:

Learning: Kohonen unsupervised Activation function:f (x) =

1, if x is the closest according to Euclidean distance 0, all other cases

Layer 2–3:

Learning: Grossberg supervised Activation function: f (x) = x

428

Md. T. Habib et al.

Table 2 Comparison of experimentally evaluated classifiers in terms of six standard performance metrics. The best classifiers performance is highlighted Classifier

Accuracy (%)

Sensitivity (%)

Specificity (%)

Precision (%)

FPR (%)

FNR (%)

SVMs

95.2

85.6

97.12

85.6

2.88

14.4

C4.5

86.67

60.0

92.0

60.0

8.0

40.0

Naïve Bayes

77.78

33.33

86.67

33.33

13.33

66.67

Logistic regression

75.91

27.73

85.55

27.73

14.45

72.27

kNN

71.11

13.33

82.67

13.33

17.33

86.67

Random forest

90.36

71.07

94.21

71.07

5.79

28.93

BPN

89.58

68.75

93.75

68.75

6.25

31.25

CPN

91.66

75.0

95.0

75.0

5.0

25.0

RIPPER

78.83

36.50

87.30

36.50

12.70

63.50

5 Conclusion For comparing machine learning classification techniques, papaya disease recognition based on machine vision is an important context, because this context plays an important role in building our proposed agro-medical expert system. In this paper, a thorough comparison of nine prominent classifiers’ performances in the context of papaya disease recognition has been presented. SVMs have been found as the best classifier, whereas kNNs performances have been the poorest. We have found such results that they are beneficial for building an agro-medical expert system. We have performed our experiment with not so large data set. We can make our experiment rigorous by increasing the size of the data set. Moreover, there remain future works for performing the same experiment with a much larger data set of some other local fruits’ images.

References 1. Bangladesh – agriculture, https://www.nationsencyclopedia.com/economies/Asia-and-thePacific/Bangladesh-AGRICULTURE.html. Accessed 04 Oct 2018 2. Habib MT, Majumder A, Jakaria AZM, Akter M, Uddin MdS, Ahmed F (2018) Machine vision based papaya disease recognition. J King Saud Univ – Comput Inf Sci https://doi.org/10.1016/ j.jksuci.2018.06.006 3. Entezari-Maleki R, Rezaei A, Minaei B (2009) Comparison of classification methods based on the type of attributes and sample size. JCIT 4:94–102. https://doi.org/10.4156/jcit.vol4.issue3. 14

36 A Comparative Study of Classifiers in the Context of Papaya …

429

4. Robles-Granda PD, Belik I (2010) A comparison of machine learning classifiers applied to financial data sets 5. Li C, Wang J, Wang L, Hu L, Gong P (2014) Comparison of classification algorithms and training sample sizes in urban land classification with landsat thematic mapper imagery. Remote Sens 6:964–983 6. Vaithiyanathan V, Rajeswari K, Tajane K, Pitale R (2013) Comparison of different classification techniques using different data sets. Int J Adv Eng Technol (IJAET) 6(2):769–779 7. Noi PT, Kappas M (2017) Comparison of random forest, k-nearest neighbor, and support vector machine classifiers for land cover classification using sentinel-2 imagery. 22 Dec 2017 8. Rahman R, Afroz F (2013) Comparison of various classification techniques using different data mining tools for diabetes diagnosis. J Softw Eng Appl 6(3):85–97. https://doi.org/10.4236/jsea. 2013.63013 9. Tan PN, Steinbach M, Kumar V (2006) Introduction to data mining. Addison-Wesley, Reading 10. Han J, Kamber M, Pei J (2012) Data mining concepts and techniques. Elsevier, Amsterdam 11. Logistic regression, https://www.medcalc.org/manual/logistic_regression.php. Accessed 04 Oct 2018 12. Logistic regression, https://en.wikipedia.org/wiki/Logistic_regression. Accessed 04 Oct 2018 13. Rojas R (1996) Neural networks: a systematic introduction. Springer, Berlin 14. Anderson D, McNeill G (1992) Artificial neural networks technology. Contract report, for Rome laboratory, contract no. F30602-89-C-0082 15. Confusion matrix, https://en.wikipedia.org/wiki/Confusion_matrix. Accessed 04 Oct 2018 16. Rozario LJ, Rahman T, Uddin MS (2016) Segmentation of the region of defects in fruits and vegetables. Int J Comput Sci Inf Secur 14(5):399–406

Chapter 37

A Hierarchical Learning Model for Claim Validation Amar Debnath, Redoan Rahman, Md. Mofijul Islam and Md. Abdur Razzaque

1 Introduction In recent years, the massive growth of different social media platforms allowed the rapid dissemination of information to the mass people. Moreover, almost all of the well-known news media utilized these social platforms to spread their contents. Hence, people increasingly depend on these social media platforms to seek out and consume news. For example, in 2016, about 62% of U.S. adults depend on social media to get news content.1 On the other hand, propagandists exploit these platforms which lead to the rapid dissemination of the fake or distorted news. For instance, during United States Election of 2016, there were 156 fake news that was identified later, among which 115 were pro-Trump and 41 of them were pro-Clinton [1]. Thus, fake news intentionally persuades consumers to accept biased or false beliefs and subsequently, it is very much crucial to stop the spreading of distorted contents. To mitigate the negative effects of fake news, it is critical to develop methods for automatically detecting fake news on social media platforms. However, it is a strenuous task to detect the distorted news, as the propagandists carefully craft the news to present it as validated news. Even one study reveals that human evaluators 1 www.journalism.org/2016/05/26/news-use-across-social-media-platforms-2016/.

A. Debnath · R. Rahman · Md. Mofijul Islam (B) · Md. Abdur Razzaque Green Networking Research Group, Department of Computer Science and Engineering, University of Dhaka, Dhaka, Bangladesh e-mail: [email protected] A. Debnath e-mail: [email protected] R. Rahman e-mail: [email protected] Md. Abdur Razzaque e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_37

431

432

A. Debnath et al.

can identify the fake news with only 50–63% accuracy [14]. Moreover, brief contents in the social media platforms make this task more arduous. Therefore, designing an automated fake news detection approach on social media has recently become an emerging research area. Recently, a couple of content verification systems have been introduced to address the fake news problem. For example, expert-based fact-checking platforms have been introduced (PolitiFact,2 Snopes3 ). Furthermore, crowd-sourced expertise has been utilized to detect the check worthy statements, for instance, Fiskkit4 allows the users to annotate news contents. Even though these approaches are able to verify the news articles, the main drawbacks of these systems are the scalability. Hence, in recent years, Natural Language Processing practitioners have been trying to solve the deception detection issue with the help of various statistical machine learning approaches [15]. Nevertheless, the main obstacle to develop and evaluate these learning approaches is the lack of comprehensive benchmark dataset. Additionally, fake news contain various linguistic cues, which enforce a diverse dataset to train and evaluate the state-of-the-art learning models. Recently, a couple of datasets have been introduced in the literature to aid the claim verification problem [3, 16, 17]. In this work, we proposed TLCV: Transfer Learning based Claim Validation model, where we incorporate composite Convolutional Neural Network with deep contextualized word vector representation model, ELMo [12]. Our proposed approach TLCV outperforms the state-of-the-art-works in identifying the check worthy statements. The major contributions of this work are summarized below: – Design a transfer learning-based claim validation learning model. – Develop a composite Convolutional Neural Network to achieve the consistent state-of-the-art performance. – Employ the deep contextual word representation to capture diverse linguistic cues to improve the fake news detection approach. The rest of the works are designed as follows: in Sect. 2, we have discussed the related works. Thereafter, the proposed claim verification model is presented in Sect. 3. Subsequently, in Sect. 4, the detailed experimental analysis and performance evaluation of the proposed models are discussed. Finally, Sect. 5 has concluded the work with future plans.

2 Related Works In recent years, deception detection problem has been studied in various domains from different angles. In [8], computational linguistics-based deceptive opinion mining approaches have been proposed and a crowd-sourced dataset has been developed. 2 https://www.politifact.com. 3 https://www.snopes.com/. 4 http://fiskkit.com/.

37 A Hierarchical Learning Model for Claim Validation

433

However, the detection of fake news is challenging, due to the diverse linguistic cues. A few works have been addressed for the validation of news contents, because of the lack of benchmark datasets to aid this fact-checking problem. In [3], the author proposes a fact-checking dataset, nonetheless, it contains only 221 training examples and thus, this dataset is not enough to design a comprehensive claim validation learning model. In [17], the author established a novel dataset, named LIAR, which consists of 12.8K human-labeled short statements from Politifact platform and each of these statements is categorized into six different labels—pants-fire, false, barely-true, half-true, mostly-true and true. LIAR is one of the most comprehensive datasets, which exclusively addressed the claim validation problem. Moreover, LIAR contains textual statements along with the speakers’ profile metadata—party affiliations, state, job and credit history. Recently, another dataset, FEVER: Fact Extraction and Verification [16], ushered a new way to confront the deception detection problem. The characteristic of FEVER dataset is that it includes the evidences of the claim statements, which are extracted from a knowledge base (Wikipedia). A number of computational approaches have been proposed to validate the news contents by utilizing the publicly available datasets. In [4], the author analyzed user reviews from Amazon and concluded a few assumptions that fraud reviewers tend to stay with their review style and make a lot of similar comments with various word combinations. Linguistic-based approach has been proposed to detect deceptive text on a crowd-sourced dataset in [11]. Nevertheless, the lexical features are not enough for designing the fraud detection learning model. In [6], the authors showed that speaker profiles are also useful for classifying fake news. Additionally, in the LIAR [17] dataset, the author utilized a hybrid Convolutional Neural Network to develop a computational claim verification learning model. Moreover, in [7], the author proposed a hybrid model with attention mechanism, which utilized Long Short-term Memory learning model as well as attention blocks on speaker and topic attributes. This approach outperforms the baseline approach of LIAR dataset [17]. Unlike the state-of-the-art-approach, in this work, we propose a computational approach, TLCV, to validate a claim statement, where we not only just utilize the textual features and author profile but also linguistic cues retrieved using the composite CNN model. Moreover, we also utilize the transfer learning approach to extract the deep contextual information from the claim statements.

3 Proposed Computational Claim Validation Approaches In this work, we have designed a computational approach, TLCV, to validate the check worthy statement by leveraging the deep contextual word representation and the ruling comments extracted from the Politifact platform. In addition, we also incorporated the speaker information from LIAR dataset to train our claim validation learning model.

434

A. Debnath et al.

3.1 TLCV Architecture In TLCV, we employed a composite Convolutional Neural Network (CNN) model to train the claim validation model. The proposed TLCV model can be divided into three modules: – Retrieve Evidence. – Vectorized Representation of Composite Features. – Composite CNN Training Model. Retrieve Evidence The textual information of the claim statement is not sufficient to check the validity. For this reason, this module is designed to retrieve evidence and augment this information in training the fact-checking model. A couple of stateof-the-art approaches have been proposed to retrieve related information of a given statement. For instance, we can utilize the TF–IDF similarity [13] to extract related information from the large knowledge base. Furthermore, Sent2vec [9] with compositional n-gram features can be utilized to retrieve contextual information. TF–IDF [13] performs fairly well but it focuses on mainly syntactic features. On the other hand, Sent2vec [9] uses n-gram features to represent a sentence into a vector to capture the semantic information. This vectorized format of the sentence can be utilized to retrieve related information. In this work, later approach, Sent2vec, has been incorporated to extract sentence as the compositional n-gram features can help to capture the appropriate textual semantic cues. The extracted sentence vector representation has been utilized to calculate the cosine similarity between news claim statements and the sentences of the ruling comments, which are extracted from the Politifact, for determining the effective evidence. In this work, we directed our attention to the evidence retrieved from Politifact.com. Vectorized Representation of Composite Features This module of TLCV is responsible to represent the textual and other meta-features into a vectorized format. As the TLCV has been designed by incorporating the composite CNN model, hence the representation of the textual features into the vectorized format was needed to feed into the CNN model. In this work, the pretrained ELMo [12] model was utilized to represent the textual features into the vectorized format, as ELMo can capture deep contextual word representation. The ELMo model consists of 13 million parameters and it is trained on the 1 billion Word Benchmark [2], which can represent a sentence in 256- dimensional vectors. In the TLCV, the following three groups of features have been incorporated. – The claim sentences in the LIAR dataset are padded to be 100 long word sequence. Then, the sentences were fed into the ELMo model and the words of the sentences are converted to 256-dimensional vectors making each sentence size to be (100, 256). – Ruling comments are scrapped from Politifact and sent2vec [9] method has been employed to fetch the appropriate evidence. Similar to claim sentences represen-

37 A Hierarchical Learning Model for Claim Validation

435

Fig. 1 Training model of TLCV

tation, the evidence sentences are processed and converted to (100, 256) sized matrix. – In TLCV model, the speaker information (speaker’s name, job and party affiliations) have also been incorporated. We concatenated these three pieces of information into one sentence. After that these sentences were converted to 256-sized vector representations using the ELMo model, similar to the previous approach. Hence, each speaker information is represented by (3, 256) sized matrix. Composite CNN Training Model Convolutional Neural Networks have proven to be efficient when it comes to sentence classification task [5]. Hence, to design the underlying architecture of TLCV, the composite CNN model is utilized, where each of the feature groups is fed into a separate CNN filter set. The complete composite training model architecture is depicted in Fig. 1. In the first layer of TLCV training model, three input matrices were constructed. Feature embeddings of claim statements, evidence sentences, and author profile information were turned into a vectorized format using ELMo model. Three sets of convolutional filters were initialized and applied into these features embedding. The three kernel sizes of the first two convolutional filter sets are {2, 4, 8} and the kernel sizes for the third convolutional filter set are {1, 2, 3}. The resulting feature embeddings from the first layer are concatenated and a composite feature embedding is created. After that, another filter set of size {1, 2, 5}, has been applied to this composite feature embedding. The output of the filter is flattened to be passed to a dense neural network of hidden neuron size 256. Subsequently, the output of the first dense layer is again passed to another dense layer consisting of six neurons with softmax activation for producing the final output. Relu activation

436

A. Debnath et al.

and valid padding are used for all the convolutional filters. Standard dropout with discrete probability has been utilized to prevent over-fitting. This architecture gave a consistent result and proved efficient in extracting input features.

3.2 Extended Models of TLCV The proposed model TLCV has been extended and two computational models are designed. However, the underlying training module architecture of these two extended models is similar to the previous TLCV model. The purpose of these two extended models is to serve as a baseline model to evaluate the performance of the proposed TLCV model. CVMwoSF: Claim Validation Model without Supplementary Features In this extended learning model, GloVe [10] word vector representation model has been employed to create the feature embeddings. Moreover, in CVMwoSF model the supporting features such as retrieved evidence and speaker profile was not incorporated. CVMwSF: Claim Validation Model with Supplementary Features In CVMwSF mode, similar to the TLCV, retrieved evidence and speaker profile was incorporated as supporting features. However, in this model, GloVe word vector representation is utilized for feature embeddings instead of ELMo.

4 Experimental Analysis For evaluating the performance of our proposed claim verification model, TLCV, the LIAR dataset [17] was utilized, as it not only consists of the claim statement but also the speakers’ profile information. Each of the claim statements of LIAR dataset is categorized into six labels—pants-fire, false, barely-true, half-true, mostly-true, and true. Besides the speaker profile information and claim statements, the LIAR dataset was augmented with ruling comments which serve as an evidence of the claim statements. A python scraper was written which helps to extract the ruling comments from the Politifact platform. The detailed information of LIAR dataset is presented in Table 1. We conduct the performance evaluation of TLCV model in Google Colaboratory Platform.5 This platform provides Tesla K80 GPU, having 2496 CUDA cores, 12 GB GDDR5 VRAM. The CPU has a single-core hyper-threaded Xeon Processors with 2.3 GHz clock speed. Additionally, it contains 13 GB RAM and 33 GB available disk space.

5 https://colab.research.google.com.

37 A Hierarchical Learning Model for Claim Validation

437

Table 1 LIAR dataset

(a) Top 3 speakers Top-3 Speaker Barack Obama 94 Donald Trump 54 Hillary Clinton 48 Top-3 Party Affiliations Democrats 4150 Republicans 5687 None (e.g., FB posts) 2185 (b) Count of each label Label Pants-true False Barely-true Half-true Mostly-true True Training Set 839 1995 1654 2114 1962 1676 Validation Set 116 263 237 248 251 169 Test Set 92 249 212 265 241 208

4.1 Performance Result In this section, the detailed performance evaluation of our proposed claim validation model TLCV is presented. The performance of TLCV was compared with two baseline approaches, which are proposed in LIAR dataset [17]. The main difference between these approaches is that one approach utilizes only the claim statements and another model incorporates meta-features of fact, speaker profile, and history, in addition to the textual claims. In addition to these two state-of-the-art baseline approaches, the performance of TLCV model was compared with the two extensions of TLCV: CVMwSF and CVMwoSF. The detailed performance evaluations of these claim verification models are presented in Table 2. It can be easily identified that CVMwoSF claim verification model slightly outperforms the baseline approaches, which are proposed in LIAR work [17]. The only reason is that CVMwoSF utilizes the composite CNN training model, where each of the features groups is trained using a separate CNN model to get the feature embeddings. However, in the CVMwSF model, when the speaker profile was incorporated with CVMwoSF, the performance improves with accuracy 28.75%. Because the author of fake news has a distinctive profile. Nonetheless, both CVMwoSF and CVMwSF models incorporate GloVe model to get the vectorized representation of the textual features. On the other hand, fake news has diverse linguistic features, hence the pretrained GloVe or word2vec model cannot fully capture the semantic textual features. For this reason, in TLCV model, we leverage the pretrained ELMo model to retrieve

438

A. Debnath et al.

Table 2 Performance comparison of different claim verification models Model Accuracy Precision Recall F1-score LIAR baseline [17] LIAR baseline with speaker profile [17] CVMwoSF CVMwSF TLCV

MCC

27.0 27.4

– –

– –

– –

– –

27.52 28.75 29.38

28 29 30

27 28 29

27 28 28

0.1238 0.1294 0.1327

the deep contextual word representation to extract the composite feature embeddings. As ELMO model is able to capture the deep contextual word representation, TLCV approach outperforms the state-of-the-art models with accuracy of 29.38%. This performance improvement indicates that the claim verification model can be improved by incorporating the language model and transfer learning approach. Additionally, in another work [7], the author used a speaker profile and credit history along with the claim statement to detect authenticity. Although in [7], the author achieved a better result by incorporating the speaker credit history and bidirectional long short-term memory model (Bi-LSTM), however, in most general cases, speaker profile containing their truth history may not be available, which makes the pro-

Fig. 2 Confusion matrix of TLCV approach on test data

37 A Hierarchical Learning Model for Claim Validation

439

Fig. 3 Receiver operating characteristics (ROC) curve of TLCV

posed claim verification system impractical. For this reason, we did not consider this approach to evaluate the performance of TLCV approach. The categorical performance of the TLCV approach is depicted in Fig. 2. It can be easily identified from the ROC graph of TLCV in Fig. 3 that the categorical performance of our proposed model is quite consistent. The reason behind this consistent performance is that TLCV model leverages the deep contextual vector representation of ELMo.

5 Conclusion In this work, we proposed TLCV, a transfer learning-based claim verification approach, where we utilized the composite Convolutional Neural Network to detect the authenticity of political news. Furthermore, to improve our proposed claim verification model, we incorporated evidence retrieval module, which helps to extract evidence of the claim statement by utilizing the sent2vec approach. In the experimental analysis, our proposed model achieved 29.38% accuracy which outperforms the state-of-the-art baseline approach by 8.81%. One of the main reasons behind this performance improvement is that we leveraged ELMo model, to extract the deep contextual feature embedding of the textual information. Moreover, separate CNN models have been applied to develop composite feature group representation. The experimental evaluations of TLCV indicated that there is a lot of scope available to

440

A. Debnath et al.

improve the claim verification model by leveraging the transfer and deep learning approaches. In future, we have a plan to extend the evidence retrieval model of TLCV to extract evidence from a large and diverse knowledge base. Additionally, transfer learning approach can be employed to fine tune the pretrained model in the target domain.

References 1. Allcott H, Gentzkow M (2017) Social media and fake news in the 2016 election. Working paper 23089, National Bureau of Economic Research. https://doi.org/10.3386/w23089 2. Chelba C, Mikolov T, Schuster M, Ge Q, Brants T, Koehn P, Robinson T (2014) One billion word benchmark for measuring progress in statistical language modeling. In: INTERSPEECH. ISCA, pp 2635–2639 3. Ferreira W, Vlachos A (2016) Emergent: a novel data-set for stance classification. In: Proceedings of the 2016 conference of the North American chapter of the association for computational linguistics: human language technologies, pp 1163–1168 4. Kaghazgaran P, Caverlee J, Alfifi M (2017) Behavioral analysis of review fraud: linking malicious crowdsourcing to Amazon and beyond. In: ICWSM, pp 560–563 5. Kim Y (2014) Convolutional neural networks for sentence classification. In: EMNLP. ACL, pp 1746–1751 6. Long Y, Lu Q, Xiao Y, Li M, Huang CR (2016) Domain-specific user preference prediction based on multiple user activities. In: 2016 IEEE international conference on big data (Big Data). IEEE, pp 3913–3921 7. Long Y, Lu Q, Xiang R, Li M, Huang CR (2017) Fake news detection through multi-perspective speaker profiles. In: Proceedings of the eighth international joint conference on natural language processing (volume 2: short papers), vol 2, pp 252–256 8. Ott M, Choi Y, Cardie C, Hancock JT (2011) Finding deceptive opinion spam by any stretch of the imagination. In: Proceedings of the 49th annual meeting of the association for computational linguistics: human language technologies-volume 1. Association for Computational Linguistics, pp 309–319 9. Pagliardini M, Gupta P, Jaggi M (2018) Unsupervised learning of sentence embeddings using compositional n-gram features. In: Proceedings of the 2018 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long papers). Association for Computational Linguistics, pp 528–540. https://doi.org/10. 18653/v1/N18-1049 10. Pennington J, Socher R, Manning C (2014) Glove: global vectors for word representation. In: Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pp 1532–1543 11. Pérez-Rosas V, Mihalcea R (2015) Experiments in open domain deception detection. In: Proceedings of the 2015 conference on empirical methods in natural language processing, pp 1120–1125 12. Peters ME, Neumann M, Iyyer M, Gardner M, Clark C, Lee K, Zettlemoyer L (2018) Deep contextualized word representations. In: Proceedings of NAACL 13. Ramos J et al (2003) Using TF-IDF to determine word relevance in document queries. In: Proceedings of the first instructional conference on machine learning, vol 242, pp 133–142 14. Rubin VL (2017) Deception detection and rumor debunking for social media. The SAGE handbook of social media research methods, pp 342–363 15. Shu K, Sliva A, Wang S, Tang J, Liu H (2017) Fake news detection on social media: a data mining perspective. ACM SIGKDD Explor Newsl 19(1):22–36

37 A Hierarchical Learning Model for Claim Validation

441

16. Thorne J, Vlachos A, Christodoulopoulos C, Mittal A (2018) FEVER: a large-scale dataset for fact extraction and verification. In: NAACL-HLT 17. Wang WY (2017) Liar, liar pants on fire: a new benchmark dataset for fake news detection. In: Proceedings of the 55th annual meeting of the association for computational linguistics (volume 2: short papers), vol 2, pp 422–426

Chapter 38

D-CARE: A Non-invasive Glucose Measuring Technique for Monitoring Diabetes Patients Md. Mahbub Alam, Swapnil Saha, Proshib Saha, Fernaz Narin Nur, Nazmun Nessa Moon, Asif Karim and Sami Azam

1 Introduction Diabetes is one kind of disease of the pancreas, which is responsible for producing a hormone called insulin. Diabetes happens when the pancreas gland is no longer able to produce insulin or the body cannot properly utilize the insulin delivered by the pancreas, resulting in the high level of blood glucose in the bloodstream. However, the human body needs some glucose as it provides the necessary energy for the body to work. The prevalence of diabetes was about 8% in 2011, which will probably rise to 10% by 2030. The alarming fact is, 80% of diabetic people are from lowincome or middle-income countries. In Bangladesh about 10%, in China 9% and in Md. M. Alam · S. Saha · P. Saha · F. N. Nur (B) · N. N. Moon Department of Computer Science and Engineering, Daffodil International University, Dhaka, Bangladesh e-mail: [email protected] Md. M. Alam e-mail: [email protected] S. Saha e-mail: [email protected] P. Saha e-mail: [email protected] N. N. Moon e-mail: [email protected] A. Karim · S. Azam College of Engineering, IT and Environment, Charles Darwin University, Casuarina, NT, Australia e-mail: [email protected] S. Azam e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_38

443

444

Md. M. Alam et al.

India 8% of the population are affected by the disease [1]. Deaths occurring from diabetes mellitus in Bangladesh has reached to 40,142, which is nearly 5.09% of total deaths according to WHO [2]. As the statistics show, diabetes is a key reason for deaths in Bangladesh in recent years.Statistics shows that the prevalence of diabetes in Bangladesh has increased from 5% (2001–2005) to 9% (2006–2010). Such rate will rise up to 13% by 2030 according to the International Diabetes Federation [1, 3]. One worrying stat is that the diabetes-affected infants and aged people are less cautious about their condition than the adults and middle-aged people. Supposedly, the middle-aged people are more cautious about their condition, and it is seen that they are capable of understanding their medical conditions better than the aged people and the younger ones. Non-invasive method for blood glucose monitoring is becoming more of interest to researchers as it has become a must for constant monitoring of glucose level in blood. In this paper, a design has been presented to monitor blood glucose level using Near-Infrared (NIR) light source, which is of 940 nm wavelength [4–7]. The intensity of light passed through the finger is used to calculate the blood glucose level. Arm, finger and earlobe; these three different probes were used to calculate the blood glucose using 940 nm NIR LED. The data obtained from the sensor through the microcontroller subsequently is sent to a web server and also to an observer (any relative/someone taking care of the patient), in case, the level of glucose in bloodstream becomes too low or too high. That way, a person can always be aware of any impending danger if such condition does arrive. This system can be used to continuously monitor the blood glucose level and notify the patient accordingly to take necessary measures for an initial treatment; e.g. if the patient’s blood level is critically low, the system will advise the patient to consume some kind of desert or any sweet stuff. However to determine the exact level of insulin dose required, the sugar level in bloodstream needs to be determined by an invasive method.

2 Related Works Through an in-depth literature review, some research initiatives based on the abovementioned method have been identified, where the main objective has been to measure the glucose level and the proper insulin dose. The widely used method to calculate glucose level in the bloodstream is an invasive technique that is painful, high-priced and may cause the outbreak of infectious diseases. Besides, the invasive technique results in damaged finger tissues in case of frequent application. As an alternative, the non-invasive approach may be used, which helps frequent check-ups and relieves ache and discomfort caused by common finger pricks [4, 8–15]. Daarani and Kavithamani [4] have proposed a non-invasive method of glucose level measurement with a NIR sensor. The method displays the measured value in LCD display and also sends out to the Android application, as well as stores data via Bluetooth. Buda and Addi [8] have proposed a system, where they have developed a non-invasive

38 D-CARE: A Non-Invasive Glucose Measuring Technique …

445

blood glucose monitoring device to observe glucose concentration in blood. The device shows the glucose level and the proper insulin dose, resembling the body mass index (BMI) of the user. The work projected a 4–16% accuracy in glucose detection. Rahmat et al. [9] have proposed IoT-based non-invasive glucose monitoring technique where the researchers have put fingertip into Near-Infrared (NIR) LED to calculate blood glucose level and the concentration measure of glucose in the blood. They displayed glucose reading on the LCD display and showed it to the patient’s family and the doctor via SMS and Android application for monitoring patients remotely. The system demonstrated results with an approximate accuracy of 88.89%. Based on seven assessments, the work had average 7.20% error of glucose reading in comparison to their proposed method with invasive method. Saleh et al. [10] have designed a non-invasive system to measure the blood glucose level. They showed 17% accuracy by implementing a notch filter. Yadav et al. [11] have introduced the glucose sensor based on the principle of NIR LED. In their work, the 940 nm spectrum continuous wave has been used to explore different concentration of glucose for experimentation. The authors had experimented on the human forearm and overlooked the reflectance spectra of blood. Bobade and Patil [12] have described a non-invasive blood glucose level detection method for diabetic and non-diabetic peoples. Near-Infrared (NIR) sensor for the measurement of blood glucose has been used in their proposed method. The measured glucose level is further transfused to the smart Android app for exploration and storage of the data. Narkhede et al. [13] have introduced the method for non-invasive glucose estimation using the nearinfrared-based optical technique. The structured method comprises of LED emitting signals with the wavelength of 940 nm. Hotmartua et al. [14] have developed nearinfrared sensor-based non-invasive earlobe testing method for detecting the blood glucose. In their research work, two approaches had been devised to generate a formula for the estimation of blood glucose concentration with a maximum error rate of 30%. Lawand et al. [15] have designed a compact framework for non-invasive blood glucose measurement. They tested and found the valid results by using statistical techniques. The outcome showed that there is a correlation between voltage intensity level due to the pulsatile nature of blood and blood glucose level. Menon et al. [16] have put forward a non-invasive voltage intensive blood glucose monitoring. Near-infrared sensor has again been used for the proposed method and after obtaining result, it is communicated with a smartphone through Bluetooth.

3 Methodology and System Architecture 3.1 Block Diagram of Proposed Work From Fig. 1 it is seen that Glucose level is measured by Photodiode and Near Infrared (NIR) Sensor. NIR with 800–1700 nm wavelength is suitable for measuring continuous glucose level. Once NIR sensor transmits continuous wave, a photodiode with

446

Md. M. Alam et al.

Transmitter (Near Infrared)

Fingertip

Receiver (Photo-diode)

Noise Filter and Amplication

Micro-controller (Arduino UNO)

GSM Module

Server

LCD Monitor

Emergency SMS

Demonstrate Real time reading in SmartPhone

Fig. 1 Proposed block diagram of non-invasive blood glucose monitoring system

1550 nm wavelength receives these waves. After noise filtration and amplification, these wave signals are converted into a suitable voltage value and the microcontroller converts these voltage value to an equivalent glucose value. This glucose value is displayed to the patient with a LCD monitor. The Microcontroller also sends this value to a server using a GSM module (SIM808). Observer can observe this glucose level obtained from the patient by using a smartphone. In critical situation, emergency SMS is also sent to an observer (Fig. 1).

3.2 Architecture of System The proposed system works on a diabetes patient, and simultaneously aids the patient’s carer or observer. Patient’s fingertip is placed between the glucose measuring sensors that measure glucose value through the aid of microcontroller. The patient can see his glucose reading in the LCD display and observer can monitor the patient’s glucose condition through a smartphone. The observer also gets notified with SMS when the patient’s glucose level reaches at a worrying level (Fig. 2).

38 D-CARE: A Non-Invasive Glucose Measuring Technique …

447

Displaying blood glucose level to the patient Measuring Blood Glucose by Non-

Arduino UNO

Sending SMS to

Invasive method

Observer in critical condition

GSM and Wi-Fi Module Real-time glucose reading on SmartPhone

Fig. 2 Architecture of the proposed system

3.3 Flowchart of the Proposed System Figure 3 illustrates the flowchart of the proposed system. First, NIR sensor is powered up. Now to measure blood glucose, the patient’s finger has to be placed between the NIR sensor and photodiode sensor. NIR sensor generates optical wave to the photodiode and attenuated light wave, which is then measured and converted into a signal by the photodiode sensor. In the subsequent steps, noise frequency of the signal from NIR is reduced through noise filtering procedure and the signal is amplified through amplification procedure in order to expand the weak signal. After the conversion of signal into electrical current value, Arduino converts this current value into a relative glucose value. The obtained glucose value is then compared using some predefined conditions, and according to the conditions, the derived glucose value is displayed in an LCD monitor with categorized level such as dangerously low, low, normal, high and very high. When blood glucose value is less than or equal 50 mg/DL, then this will be categorized as dangerously low. An SMS will be sent to the observer to check on the patient whether he or she has sought medical attention. If blood glucose is between 51 and 70 mg/DL, then it will be categorized as low. An SMS will also be sent to the observer at this time to provide some sugar\dessert to the patient as soon as possible. If the blood glucose is between 71 and 179 mg/DL, then it is considered as normal. If the value ranges between 180 and 237 mg/DL, then it is taken as high. And if the blood glucose value is more than 237 mg/DL, then it will be categorized as very high level, and this time also an SMS will be sent to the observer to advise the patient to visit a healthcare profession as soon as possible. All glucose readings are stored in a database before a monitoring cycle completes.

448

Md. M. Alam et al. Start

Yes

Display blood glucose possibly low

Yes

Display blood glucose NORMAL

if 51 6.93 mmol/l (126 mg/dl); (4) After 2 h plasma glucose load was >11: (a) the mean of ACR was ≥ 301 μg/mg; (b) the BUN was ≥ 21 mg/dl; and (c) the plasma creating was ≥ 1.7 mg/dl. The residua of the tolerated persons were separated as the T2D except for DN group.

44 Type 2 Diabetics Treatment and Medication Detection …

523

3.2 Dependent and Independent Variable Selection All data that have been collected, separated, and stored for each T2D tolerated people. There are eight dependent and independent elements. It is mandatory to distinct the independent and depended variable from these collected data. In order to recognize the prediction, we need to find out the relation between dependent and independent elements. It is also essential to check multicollinearity and redundancy among the subjected elements. In our dataset, the “Fasting”, “2 h after glucose load”, “BMI”, “Duration”, “Age”, “Sex”, “Blood pressure” are considered independent ingredients and “Medications” is considered as a dependent component.

3.3 Features Selection Features selection implementation needs to identify for T2D classification pattern. The entered data set holds eight different elements; the “Medications” column is taken as objective variables. The elements which are significant will be considered from those elements that their p-values are not more than 0.05. Accordingly, p values of the all seven independent elements, the approved ingredients are the four ingredients that are “Fasting”, “2 h after glucose load”, “BMI”, and “Duration”. In order to move to the next experiment, this task has established the connection between CDF(P) costs and P costs. In narration, the cost of this operation must be as near as probable to one. In conformity with the P-CDF(P), connection with ingredients gathered by this task is 42.85% among elements, which is equal to three elements that are not momentous. However, the resting 57.14% which is same to four ingredients are momentous. The relations with momentous free ingredients with the target element are more massive than 40%, which is shown in Fig. 4.

Fig. 4 Best selected features achievement

524

Md. Kowsher et al.

3.4 Data Normalization The way, for instance, Z-Score and Min-Max are necessary in order to keep the data in a rule. The graphical transformation is entangled on raw data via Min-Max. Min X and Max X are minima and maxima cost for the prominent X respectively. This way explicates the ingredients of prominent X into the limit and the consequent formula: v =

v − MinX Max A − MinX

Z-Score keeps the deliberation as a necessary formalization approach just as the real minimal and maximal amount of particular is not known. That can be explicated under the mathematically: v =

v−X σX

The main benefit of that way is to keep focus in the row data from a unique extent in order that the patterns can be drawn graphically input and output interaction normally. The principle complication that must be neglected in an existing way is about 0’s redundant inside an exact quality.

3.5 Data Validation The formalization of the entered data which is relevant to approve the data as well as consecutive in case they are developed or not. The task accomplished the fivefold form in order to obtain the achievement of each fold of data so that the pair get the formalized tables which have been encouraged to the prospective classifier with a districted distribution of training and testing data sets. For the pair tables, the same stride is mimicked. The approach inaugurated to get 5% of all data as a testing chunk of data and the training chunk is kept as rest detections.

3.6 Algorithm Selection The work prospective here used seven classification algorithms to predict the medications of T2D in humans. The applying classifiers are Logistic regression, LDA, Naive Bayes, Decision Tree, Random forest Classifier, SVM, and K-NN. The data set for T2D was gathered and applied on each classifier to predict or target the T2D medications and the achievement of the classification algorithms is evaluated based on the accuracy of achievement.

44 Type 2 Diabetics Treatment and Medication Detection …

3.6.1

525

Logistic Regression

Logistic regression is a statistical and mathematical procedure in order to explore a data set and it has one or more than one free variant that holds a result. The result is deliberated with a diploid variant. The main aim of logistic regression is to catch the best probable proper technique to derive the correlation between the diploid manner of a section and a set of free variants. Logistic regression provokes the coefficients of a pattern to detect a logit formalization of the chance of the existence of the distinctive of interest: logit( p) = b0 + b1 X 1 + b2 X 2 + b3 X 3 + · · · + bk X k Here, p is the chance of the distinctive of interest. The logit formalization can be selected as the logged odds. odds =

pr obabilit y o f presence o f characteristic p = 1− p pr obabilit y o f absence o f characteristic

And 

p logit( p) = ln 1− p



After performing the normalization procedure, the whole data set has been run via logistic regression method.

3.6.2

Linear Discriminant Analysis (LDA)

LDA is the logical reasoning of Fisher’s linear classes; LDA is an algorithm which is used in statistics, mathematics, acceptance pattern, and machine learning to obtain a linear correlation among the elements. The outcome of correlations can be used as a linear classifier. LDA builds detections by measuring the chance that a new list of entered data belongs to every class. The class that holds the best chance is the resultant class and detection is built. Bayes Theorem is imparted to measure the chance of the innovations. In sort, Bayes’ theorem may be imparted to measure the chance of the possible result class and entered the input via the chance of every class, and the chance of the entered data is connected to every class.

3.6.3

K-Nearest Neighbor Classification (K-NN)

The K-NN is not a parametric model that is necessary to separate data and predictive Regression [10]. In every pair case, the entered data are formed of the K best nearest training instance in the elements place [10]. K-NN is one kind of instance establish-

526

Md. Kowsher et al.

ing learning. In K-NN classification, the result contains a group of members. The separation of a group is selected by the multiplicity election by the acquaintance of data. If K = 1, then the class has an along nearest neighbor. If k = 2, then the class has bi or double closet and so on. In general, the weighting decorations, the signal neighbor is asserted to an amount of 1/x where x is the length to the neighbor to the point which is separated. The minimum length between any two acquaintance is always a solid line and the length is called Euclidean distance. In elements place, the extraction is taken place on raw data before asserting K-NN procedure. Mathematically, the term of Euclidean distance is

Here, D(a, b) means Euclidean length between b and a. In our system, to run the data with the help KNN, we have taken k = 5 with neighbors.

3.6.4

Decision Tree

The decision tree is a procedure which uses the structure of a tree and map the system probable consequences according to result of event, ability to classify from their costs, and service. In this procedure, it showcase a system that only holds on conditional control [11]. It is a procedure in order to show a system that only holds on conditional control affirmations. Decisions trees are generally held in several fields such as operations research, decision analysis, to assist separate a strategy most likely to reach an aim. Here is a simple graph of how decision tree works (Fig. 5).

3.6.5

Naive Bayes Classifier

The Naive Bayes Classifier system is established on probably is known as Bayesian system and is specially set up when the numbers of the inputs are too big. In spite of its easiest, Naive Bayes is the most usable algorithm in the field of mathematics or statistics. A Naive Bayesian model can be developed easily without complexity and has parametric measurement which creates inherently utility for large datasets.

44 Type 2 Diabetics Treatment and Medication Detection …

527

Fig. 5 How Decision Tree works

Here, P(c|x) is the predator chance of a class or group, P(c) is the system chance of class, and P(x|c) is the likelihood.

3.6.6

Support Vector Machine (SVM)

Support Vector Machine (SVM) is known as well-established system to separate data among themselves. It is a state-of-the-art innovations and separation classifier inherently build by an isolated hyper line or plane. In SVM all given labeled training data (supervised learning), the algorithm results an optimal hyperplane which categorizes new examples. In two-dimensional space, this hyper line or plane is a line to separate a plane in two groups or a part wherein each group lay in either side or another [12]. SVMs can have accuracy rate to execute a nonlinear classification holding what is said the kernel trick, similarly, SVM is used for mapping their large data into high-dimensional element spaces.

3.6.7

Random Forest Classifier

Random Forest Classifier is a classification instrument. Random Forest Classifier which is initiated for calculating a complex’s quantitative or unconditional biological work risen on a quantitative narration of the compound’s molecular structure [13]. It was being also roofed in our research as one kind of the classifiers for making the T2D medications prophecy structure. The outcome of the random forest algorithm is a store of decision trees risen on the indiscriminately choice of shape. Figure 6 shows an ordinary graph of how Random Forest Classifier work.

528

Md. Kowsher et al.

Fig. 6 How Random forest Classifier works

3.7 Comparison of Classification Algorithms The execution of the investigated methods is evaluated applying the entire correction rate or exactness (ACC) and confusion matrix, where ACC assembles to the fraction of the entire exact classified to the dataset. The confusion matrix created with a true specific rate, true weak rate, false specific rate, and false weak rate. The true specific rate is separated as a coefficient of true specific classified of a specific point and reverse the false weak points, while the true weak rate is the coefficient of true weak points and reverse the false specific points. Hence, we get the measurement of correction of each algorithm applying confusion matrix actions, as well as compare the result among all the classification algorithms (Fig. 7). In the raised seven classification algorithms, Random Forest Classifier gives 93.8% accuracy which showed the satisfying achievement. So, our selection classification algorithm is Random Forest Classifier. In Table 1, we have shown the accuracy

Fig. 7 Comparison of classification algorithms

44 Type 2 Diabetics Treatment and Medication Detection … Table 1 Comparison of classification algorithms

529

S. No

Technique

Accuracy Standard deviation

1

Logistic regression

0.811905 0.112632

2

LDA

0.923810 0.083027

3

K-NN

0.923571 0.071178

4

Decision tree

0.923095 0.058010

5

Naive Bayes

0.913810 0.066750

6

SVM

0.875714 0.052788

7

Random forest classifier

0.938095 0.060422

of all proposed classification algorithms with a standard deviation and also presented the graphical view compared among all techniques.

3.8 Experimental Results This work is performed in Anaconda circulation and Python 3.6 tools. There are several Python libraries, which provide solid implementations of a range of machine learning classification innovations. One of the best known is scikit-learn, a package that provides dynamic forms of a large number of common algorithms. To build our model, we used scikit-learn which provides essential tools for cleaning data, prepossessing data, and running classification algorithms. The systems such as matrix manipulations, plotting of functions and data, implementation of algorithms, the developing of user interfaces, and interfacing with programs have been written in Python using module Matplotlib and others. To read the provided data, we have imparted Pandas The experimental correlation Logistic regression, LDA, Naive Bayes, Decision Tree, Random forest Classifier, SVM, and KNN classifier are done based on the achievement parts of classification exactness and precision.

3.9 Final Result By conducting several types of classifications procedures, we are going for obtaining the ultimate conclusion which is established on the fateful exploration. After construction the machine learning action we have to conduct to the pattern and select the applicable dataset which is handy in the columns and check likeness of all selected classification algorithms in conformity with the best validity. Thus, the ultimate result of diabetes as shown in Fig. 8.

530

Md. Kowsher et al.

Fig. 8 Achievement evaluation of prospective system

4 Future Work Lots of possible evolutionary ideas which can be used for the sake of development the best issue of the prospective classifiers. In this paper, Logistic regression, LDA, Naive Bayes, Decision Tree, Random forest Classifier, SVM, and KNN classifier are applied to detect medications of T2D. We can also decide and correlate the achievement of the held classifiers with other existing classifiers, and then elect the best classifier according to their achievement of efficiency rate. T2D early treatment helps the tolerated persons suffering from the disease and also aid in order to elude diabetes problem from being worse. A more classifier such as Neural Network may be developed and their achievement can be decided and correlated to get a better result of the objective function in future work.

5 Conclusion In Medical Sector, we have already known various types of data mining and machine learning algorithms. In this disposable, a new decision support system is implemented for the prediction of T2D medications. Although the classifiers worked efficiently in the forecasting of other fever also. In this disposable, T2D is predicted holding seven several classifiers and a comparative study of their achievement is pretty done. From the investigation, we found that out of seven classifiers Random Forest Classifier

44 Type 2 Diabetics Treatment and Medication Detection …

531

performed better than the others. Thus, the rate of medications prediction of T2D medications is enhanced.

References 1. Rashid TA, Abdullah SM, Abdullah RM (2016) An intelligent approach for diabetes classification, prediction and description. In: Advances in intelligent systems and computing vol 424, pp 323–335. https://doi.org/10.1007/978-3-319-28031-8 2. Nai-Arun N, Moungmai R (2015) Comparison of classifiers for the risk of diabetes prediction. Procedia Comput Sci 69:132–142. https://doi.org/10.1016/j.procs.2015.10.014 3. Al-Rubeaan K et al (2014) Diabetic nephropathy and its risk factors in a society with a type 2 diabetes epidemic: a Saudi National Diabetes Registry-based study. PloS One 9(2):e88956 4. Priyam A, Gupta R, Rathee A, Srivastava S (2013) Comparative analysis of decision tree classification algorithms. Int J Curr Eng Technol 3:334–337. ISSN 2277 – 4106 5. Orabi KM, Kamal YM, Rabah TM (2016) Early predictive system for diabetes mellitus disease. In Industrial conference on data mining. Springer, pp 420–427 6. Nai-Arun N, Sittidech P (2014) Ensemble learning model for diabetes classification. Adv Mater Res 931–932:1427–1431 7. de Luis DA et al (2019) Role of the variant in adiponectin gene rs266729 on weight loss and cardiovascular risk factors after a hypocaloric diet with the Mediterranean pattern. Nutrition 60: 1–5 8. Pradhan M, Bamnote GR (2014) Design of classifier for detection of diabetes mellitus using genetic programming. In: Advances in intelligent systems and computing, vol 1, pp 763–770 9. Sharief AA, Sheta A (2014) Developing a mathematical model to detect diabetes using multigene genetic programming. Int J Adv Res Artif. Intell. (IJARAI) 3:54–59 10. NirmalaDevi MS, Appavu alias Balamurugan, Swathi UV (2013) An amalgam KNN to predict diabetes mellitus. In: 2013 IEEE International conference on emerging trends in computing, communication and nanotechnology (ICECCN) 11. Al Jarullah AA (2011) Decision tree discovery for the diagnosis of type II diabetes. In: 2011 International conference on innovations in information technology, IEEE 12. Santhanam T, Padmavathi MS (2015) Application of K-means and genetic algorithms for dimension reduction by integrating SVM for diabetes diagnosis. Procedia Comput Sci 47:76–83 13. Casanova R et al (2014) Application of random forests methods to diabetic retinopathy classification analyses. PLoS One 9(6):e98587

Chapter 45

Initial Point Prediction Based Parametric Active Contour Model for Left Ventricle Segmentation of CMRI Images Md. Al Noman, A. B. M. Aowlad Hossain and Md. Asadur Rahman

1 Introduction Cardiovascular diseases (CVDs) are the prominent reason for death globally, especially coronary heart disease [1]. The timely diagnosis of the CVDs draws a vital role in the recovery. There are several available modalities to diagnosis CVDs such as electrocardiogram (ECG), echocardiogram (Echo), computed tomography coronary angiography (CT Angiography), and cardiac magnetic resonance imaging (CMRI) [2]. ECG and Echo can provide functional information of cardiac activities but cannot provide the internal cardiac structure. Although cardiac CT angiography provides the internal structure and functional information, this modality is associated with hazardous radiation. Therefore, CMRI is more effective and safe modality for collecting both functional and structural information of the heart. CMRI is used as an important diagnostic imaging method by the physicians for assessing cardiac activities, accurate information about morphology, muscle perfusion, tissue sustainability, etc. All these concerning activities evaluation need the assessment of the condition of the left ventricle (LV) [3]. The structural positions of the endocardium, epicardium, LV, and right ventricle (RV) have been shown in a CMRI image in Fig. 1. Md. Al Noman (B) · Md. Asadur Rahman Department of Biomedical Engineering, Khulna University of Engineering & Technology (KUET), Khulna 9203, Bangladesh e-mail: [email protected] Md. Asadur Rahman e-mail: [email protected] A. B. M. Aowlad Hossain Department of Electronics and Communication Engineering, Khulna University of Engineering & Technology (KUET), Khulna 9203, Bangladesh e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_45

533

534

Md. Al Noman et al.

Fig. 1 Typical view of a human heart on CMRI image [4] Left Ventricle Endocardium Papillary Muscles Epicardium Right Ventricle

However, LV segmentation from CMRI can be performed manually, manual LV segmentation is time-consuming, effort extreme, and subjective. Into this bargain, the automatic segmentation of LV with good enough accuracy turns into the actual important challenge for the researchers [5]. Medical image segmentation is a vital problem and therefore several types of researches have contributed in this arena like image-driven based segmentation [6–9], atlas-guided segmentation procedures, and deformable method based segmentation [10–12]. Image-driven based segmentation commonly focuses on the region growing, thresholding and pixel clustering techniques. Otsu thresholding and Region growing techniques very simple and their capabilities are not promising for LV cavity segmentation from CMRI images. On the other hand, the threshold-based techniques bank on local pixel information which is effective only if the intensity levels of the objects drop squarely outside the range of levels in the background. Considering the CMRI images, the aforementioned pitfalls limit the utility of these methods on the segmentation problem of the LV boundary. Although edge-based segmentation procedure is a robust method, this method shows limitations in two specific conditions: (i) edge presence in locations where there is no border and (ii) no edge presence where a real border exists. Therefore, the edge-based deformable model can be a possible solution for LV boundary segmenting problems. The main intention of the deformable model is to spot on the edge of an object on the biomedical image such as parametric active contour model (PACM). Deformable methods based techniques can be categorized into two groups: geodesic active contour models (GACM) (i.e., level set technique) [13] and parametric active contour models (PACM) (i.e., snake models) [14, 15]. Nonetheless, the GACM technique is basically slower than the PACM technique for their computational complexity. The PACM is an initial deformable curve in two- or threedimensional images that interchange under the impact of internal and external forces and changes shape to fit the spot-on boundaries. The amounts of elasticity and stiffness are configured by internal forces, while external forces are liable for driving the delineated curve towards the expected image features like edges, spline, and lines.

45 Initial Point Prediction Based Parametric Active Contour Model …

535

The PACM method needs initial contour (manually) to start its segmentation procedure which is a major drawback of this technique to achieve an automatic edge detection or segmentation. Since in CMRI images, the LV is situated in a grossly unambiguous position, it can be predicted the surrounded contour of LV with the high-level hypothesis. Therefore, there is a scope to hypothesize the initial contour points of LV to achieve automatic LV area segmentation. This work proposes an artificial neural network (ANN) based regression model to forecast the initial contour points of the LV cavity in a CMRI image to automate the segmentation of the LV cavity. In addition, this research work also finds the finely fitted value of the different parameters of the PACM model, especially for the CMRI images. From the final outcomes, it has been found that the proposed method outperforms the manual PACM through its automatic segmentation procedure. In addition, the proposed method required less time than the manual segmentation process. The segmenting deviation by the proposed method compared to the manual PACM is negligible.

2 Theory Here, the mathematical background of the existing PACM algorithm has been broadly described with its limitations, exclusively the initial point selection. A predictive regression model based on ANN has been proposed to predict the initial points to contour the starting the active contour operation to achieve an automatic segmentation of the LV cavity from the CMRI images. Here, at first, the basic theory of the PACM method has been discussed and then, the limitations and proposed solutions have been described in the next subsection.

2.1 Background of PACM PACM (snake) is stated as a parametric curve [16] as the relation given in (1). u(φ) = {x(φ)y(φ)}; φ ∈ [0, 1]

(1)

The parameter φ is known as snaxels that control the snake points. These parameters (snaxels) are combined together to produce a deformable shape or an active contour as given in Fig. 2. The relation given in (1) denotes energy function of PACM, which is a continuous vector that contains (x, y) coordinates of points, φ on the parametric deformable curve and mathematically it can be represented in (2). 1 E snake = 0

  α E elastic (u(φ)) + β E bending (u(φ)) + γ E image (u(φ)) dφ

(2)

536

Md. Al Noman et al.

Fig. 2 Parametric deformable curve u(φ) with arbitrary points

The internal energy is described by the first two terms of the snake total energy function. The physical constraints (stiffness and rigidity) of the snake are captured by these energies [17]. The total internal energy is the sum of two forces like elastic and bending forces and their derivative energy function is stated in (3).        du(φ)2 2  du(φ) 2 1  + β(φ) ·   E internal = (3) α(φ) ·   dφ 2  2 dφ  where α is the weighting parameter that controls the tension and β weighting parameter specifies the rigidity in the contour. The first-order derivative function discourages the elasticity of the contour and is liable for shrinking the contour curve, while the second-order derivative treats like a thin plate [18] which tries to be a smooth curvature during the deformation process. The third term is named external energy that represents how the delineation curve will match with objects of the image. Indeed image-driven function strains the snake to move the desired object boundaries. The external energy function can be expressed as E exter nal = −γ (φ) · |∇(I (φ))|2

(4)

where ∇ is a gradient operator and G σ (φ) states as a two-dimensional Gaussian kernel with standard deviation σ parameter, I (φ) indicates the image intensity and ∗ specifies the convolution process. The contour becomes unable to move when the internal (E elastic + E bending ) and external (E image ) forces are at the same condition. Eventually, three different energy functions are necessary to attract the snake to the line, edge, and termination those can be equated with their weighted product as E image = Wline Eline + Wedge E edge + Wter m E ter m

(5)

45 Initial Point Prediction Based Parametric Active Contour Model …

537

Here, Eline = I (x, y) represents the image intensity that depends on the sign of wline which indicates the snake to be attracted toward either the dark line or light line and E edge = −|∇ I (x, y)|2 represent the image gradient which is controlled by the edge-based term, Wedge . Plus, the termination energy,  2  potential    3/2 ∂θ ∂ C ∂C E ter m = ∂n ⊥ = ∂n 2 / ∂n = C yy C x2 − 2C x y C x C y + C x x C y2 / C x2 + C y2 ⊥

depends on the gradient angle, θ = tan−1 (C y /C x ) where n = (cos θ, sin θ ) and n ⊥ = (− sin θ, cos θ ) are the unit vector along and perpendicular of the gradient angle. The C(x, y) are the pixel values of the slightly smoothed version of the original image. Since the adjusting weights (Wline , Wedge , and Wter m ) can create a wide range of snake behavior [17], these parameters would adjust for finely fitted in the LV cavity segmentation in CMRI images.

2.2 Existing Problems and Further Proposal Existing PACM demands the initial points to find the edges within the range which is a limitation to achieve automatic detection of LV boundary [18]. Manual initial point selection is a time-consuming procedure for CMRI images because a single patient may have more than 20 images. The associated technician becomes bored with time and so, precision diagnostics would be hampered. Additionally, being a parametric method, this PACM needs image quality based quantified parameters to model the target identification [16]. Therefore, following limitations are necessary to improve its functionality to achieve full automatic detection of LV boundary: (i) existing PACM method cannot initiate the initial points to start operation, (ii) being a parametric method, PACM needs the values of the parameter those are actually authentic for the CMRI. The alluded limitations have been solved by our proposed method. This proposal scopes to improve the functional properties considering the existing limitations so that the modified method can perform full automatic detection of LV boundary from a CMRI image. Initial points of the predicted contour can be modeled by the ANN. According to the proposed ANN-based model with some known selected fair contour for the CMRI images can be used to model a predictive ANN network. Generally, an artificial neural network can be used to fit a curve with its spatial domain. In a particular ANN network, a nonlinear neuron is represented as an activation function [19] given in (6). f out = f (λ1 , λ2 , . . . , λn ; π1 , π2 , . . . , π p ) ˆ πˆ ) = f (λ;

(6)

In (6), λˆ and πˆ represents the input and the weight vectors of the neurons, respectively. Here, f (·) is the activation function with nonlinear characteristics. In the

538

Md. Al Noman et al.

consequence, the response variable of the ANN is defined as a composition of a number of nonlinear functions as given in (7). fr = ψ1 ∗ ψ2 ∗ . . . ψ N ( f 1 (λˆ ; πˆ ), f 2 (λˆ ; πˆ ), . . . , f p (λˆ ; πˆ ))

(7)

where ψ1 is the nonlinear function for i = 1, 2, …, N (N = Number of hidden layers), f p (λˆ ; πˆ ) are the activation functions according to the value of p (p is the number of neurons in the hidden layer), and the symbol (∗) indicates the composite operations in (7). According to the conventional feedforward ANN model having a single output neuron with h hidden layers [20], can be defined as f s = Φ(b0 +

h

i=1

bi Ψ (ai +

n

πi j λ j,s )) = F(πs ; )

(8)

j=1

The output value of the feedforward network depends on the activation functions (·) and Ψ (·). Here, λ j,s are the input variables with respect to the spatial positions. As a result, the feedforward network becomes a function of πs and , where  = (b0 , b1 , . . . , bh , a1 , a2 , . . . , ah , π11 , π12 , . . . , πhn ). Therefore,  are the parameters vector of the proposed network which is established on the minimization of the sum n  ( fr − fˆs )2 [21]. of the squared difference of s=1

In the proposed ANN-based regression analysis, the structure of the network can be presented as given in Fig. 3. In this ANN-based regression model, the number of inputs and number of outputs is the same. Here, we have used a set of n numbers of x positions of the CMRI images. The corresponding y values were considered as the outputs of the ANN structure. In this work, we trained the ANN model with 40 initial points (40 set of x’s and 40 set of y’s where every set contains 50 points of x or y discrete points). The training was conducted up to 98% regression fitting and meanwhile, the weights of the ANN were updated by the Matlab-based pattern recognition toolbox. To perform the regression analysis, we have used three hidden layers in our proposed ANN. Furthermore, for the initial point gathering from the training images, manually 4–5 points were selected to prepare a gross outlier around the LV position. Then, 10 point spline interpolation was used to predict the internal points between two manually selected points so that a precise initial boundary supposed to be formed around the LV cavity. After the training process, we found an ANN-based model to predict the initial contour points of y based on a single set of x positions of a CMRI image. Consequently, this set of the values of x and y can be used as the initial contour of the LV cavity of all CMRI images regarding the specimen.

45 Initial Point Prediction Based Parametric Active Contour Model …

539

Fig. 3 The proposed structure of the ANN-based regression model

3 Materials and Methods The CMRI image on which the proposed methodology has been applied was described here with their basic information. In addition, the methods of the proposal associated with automatic segmentation of the left ventricle have been discussed here.

3.1 Materials In this research work, we solely worked on the CMRI images. The CMRI images were collected to the dataset of several patients from [4], where there were 20 slices of CMRI images per patient regarding one cardiac cycle. The slice spacing is 7–13 mm per slice. Each image matrix is of size 256 × 256 pixels. The pixel spacing is 1.33–1.64 mm. The raw images were of .bitmap format those were converted to a .tiff format for the analysis. All images were grayscale images that were preprocessed by the Gaussian filtering before applying the proposed method. A sample image was given in Fig. 1. All image processing algorithms were accomplished in Matlab R2016a environment on a computer with Intel Core i7-4790 CPU-3.60 GHz, RAM8.00 GB, and Windows 8.1 operating system.

540

Md. Al Noman et al.

Input CMRI images of a patient

Get the gross x values of initial contour of LV cavity

Proposed ANN based regression model

Standard initial contour y points

Fig. 4 Flow diagram of initial contour prediction utilizing the proposed regression modeling

3.2 Methods Since the morphological structure of a heart, especially LV cavity remains in a certain area on a specific slice of CMRI image. Therefore, the initial contour of the LV cavity should be slice-dependent with slight uncertainty. With the help of an artificial neural network, the initial contour positions have been modeled as the description given in the previous section so that the system can be automatic. The procedure of the predictive modeling by ANN has been summarized by the block diagram given in Fig. 4. After the initial contour point prediction, the following control parameters of the active contour model were customized for the specific application in CMRI images: (a) Alpha parameter, α which controls the elasticity of the snake that is image dependent, (b) Beta parameter, β that specifies the rigidity in the contour by combining with the second derivative term, (c) Gamma parameter, γ that quantifies the step size, (d) Kappa parameter, κ which acts as the scaling factor for the energy term, (e) W (Eline ) is a weighting factor for intensity-based potential term, (f) W (E edge ) is another weighing factor for edge-based potential term, (g) W (E ter m ) is termination potential weighting factor. These parameters have been customized to evaluate the appropriate LV area from CMRI images by means of check and trial method. The estimated customize values of the aforementioned parameters have been tabulated in the next section. Additionally, optimum iteration has also been figured out to achieve the target contouring so that computational burden can be reduced. Without proper stopping criterion and the optimal iterations, the system may not be computationally efficient in the sense of time consumption. Considering the optimal values of these parameters, all the slices of a testing data set were taken as input to the proposed algorithm. The predicted initial contour points of the specific slice were taken from the database which was created previously according to the method described by the flow diagram given in Fig. 4. This set of initial contour points acted as the initial points of an active contour model and with the increment of the iteration the snake was reformed by energy minimization procedure. After completing the total iterations, the system stores the segmented contour for further analysis. The procedure of the proposed methodology can be summarized by the block diagram given in Fig. 5.

45 Initial Point Prediction Based Parametric Active Contour Model … Fig. 5 Block diagram of the proposed methodology to perform automatic segmentation of the LV cavities from the CMRI images

541

Load all CMRI image of a patient N=Number of slices i=i+1; (initially i=0)

Database of the contour points

Get the initial contour points

Initiate the procedure to delineate the ROI with the help of snake model (PACM) Evaluate the final deformed snake around the ROI (LV cavity) and store it for further usage

Is i>N? Yes

No End of the session

Storage of the individual edge points of LV cavity of each slice

4 Results and Discussions Considering some standard inhomogeneity images as control the values of the parameters of the PACM were found to be fairly fitted for the CMRI images. By the check and trial method of the Matlab toolbox [22], the finely fitted values of the parameters were found. The corresponding values are given in Table 1. Then, these customized parameters helped to find the optimal iterations to achieve the best segmentation from the images. With the values of finely fitted parameters, the CMRI images needed around 200 iterations to fit the snake with the LV cavity. The corresponding results of control images and the CMRI images with different iterations and fitted corresponding snakes have given in Fig. 6. To complete contouring the edges of LV 200 iterations were needed. With the previous problem of initial point selection of 20 slices images of a single patient is a time-consuming issue. If at least 10 s is

542

Md. Al Noman et al.

Table 1 The nomenclatures of the parameters with applied customized values

Serial no.

Nomenclature

Customized values

1

Alpha parameter α

0.40

2

Beta parameter β

0.20

3

Gamma parameter γ

1.00

4

Kappa parameter κ

0.50

5

W (Eline )

0.30

6

W (E edge )

0.40

7

W (E ter n )

0.70

8

Iteration

200

Fig. 6 The snake approaches to the expected edges of LV after 20, 50, 70, and 100 iterations, those are given in 2nd, 3rd, 4th, and 5th figure from the left

O Predicted Intial Contour Points

Y axis of image

200

150

100

50

50

All: R=0.98394

(b) Output ~= 0.97*Target + 3.8

(a) 250

100

150

X axis of image

200

Data Fit Y=T

130 120 110 100 90 80 80

90

100

110

120

130

Target

Fig. 7 a The predicted initial contour points by the proposed ANN-based predictive model for a typical slice and b the regression curve of the model after training and validations

required for selecting the initial contour for a single image, it takes 200 s extra time for segmenting all 20 slices of a single patient. Such a procedure will bore the diagnostic personnel. As a result, this work proposes to predict the initial contour point by an ANN-based predictive model. A graphical presentation of the initial contour prediction has been given in Fig. 7. With these initial contour points, a set of 20 slice images have been segmented and given the corresponding results in Fig. 8.

45 Initial Point Prediction Based Parametric Active Contour Model …

543

Fig. 8 Total 20 slices of CMRI images and their segmented area by the proposed method

There is a slight deviation between the segmented area of LV by the manually selected initial points and the predictive initial points. The relative absolute deviation (RAD) was calculated as % RAD =

Area by Manual Process − Area by the Proposed Method × 100 Area by Manual Process

(9)

The results of RAD are given in Table 2. Here, the areas are calculated by the numbers of pixel existed inside the contour. From Table 2, we get that the deviation is very small (Mean ± standard deviation = 2.64 ± 2.01) which can be convincing enough to overlook. In addition, we have carefully observed the time elapsed during the automatic image segmentation by initial point prediction and manual procedures through 10 different trials. From the results, we got that the proposed method needs average 132.93 s and manual process needs 332.86 s. The manual process needs 199 s more in average due to initial point selection.

544

Md. Al Noman et al.

Table 2 The segmented area (number of pixels and sub-pixels inside the contour) detected by the manual and proposed method and the RAD between the two methods Serial no.

Manual process

Proposed method

RAD (%)

1

1379.7871

1436.028

4.076056

2

1447.1379

1382.86

4.441726

3

1360.9816

1422.116

4.491934

4

1320.1795

1298.499

1.642239

5

1234.5939

1308.016

5.947065

6

1365.0116

1395.574

2.238985

7

1337.5978

1326.497

0.829906

8

1333.6392

1297.888

2.680725

9

1276.8520

1169.435

8.412643

10

1121.2000

1102.086

1.704781

11

1124.8274

1116.922

0.70281

12

1364.7478

1330.436

2.51415

13

1697.7474

1717.956

1.190318

14

1642.9252

1674.74

1.936473

15

1644.3952

1614.4463

1.821271

16

1568.4770

1593.3624

1.586596

17

1578.5828

1584.1509

0.352728

18

1637.2780

1626.308

0.670014

19

1506.8533

1560.8492

3.583355

20

1582.8833

1548.8136

2.152382 Mean ± std = 2.64 ± 2.01

5 Conclusions By the proposed method, the limitation of initial contour point selection has been removed. Due to this approach, the time consumption of the LV segmentation has been reduced. In addition, this automatic segmentation approach limits the subjective inefficacy during the manual selection of the initial contours. Therefore, we hope that this proposal has been added a new dimension to the parametric active contour model in the field of image segmentation. The proposal was solely designed for the LV cavity segmentation from CMRI images considering its morphological structure. Therefore, this mechanism may not be applicable to the other images.

45 Initial Point Prediction Based Parametric Active Contour Model …

545

References 1. Nichols M, Townsend N, Scarborough P, Rayner M (2013) Cardiovascular disease in Europe: epidemiological update. Eur Heart J 34:3028–3034 2. Ali ER, Mohamad AM (2017) Diagnostic accuracy of cardiovascular magnetic resonance imaging for assessment of right ventricular morphology and function in pulmonary artery hypertension. Egypt J Chest Dis Tuberc 66:477–486 3. Hadhoud MMA, Eladawy MI, Farag A, Montevecchi FM, Morbiducci U (2012) Left ventricle segmentation in cardiac MRI images. Am J Biomed Eng 2:131–135 4. Andreopoulos A, Tsotsos JK (2008) Efficient and generalizable statistical models of shape and appearance for analysis of cardiac MRI. Med Image Anal 12:335–357 5. Souto M, Masip LR, Couto M, Suárez-Cuenca JJ, Martínez A et al (2013) Quantification of right and left ventricular function in cardiac MR imaging: comparison of semiautomatic and manual segmentation algorithms. Diagnostics 3:271–282 6. Lu Y, Radau P, Connelly K, Dick A, Wright GA (2009) Segmentation of left ventricle in cardiac cine MRI: an automatic image-driven method. In: Ayache N, Delingette H, Sermesant M (eds) Functional imaging and modeling of the heart. Lecture notes in computer science, vol 5528. Springer, Berlin, Heidelberg 7. Hu H, Liu H, Gao Z, Huang L (2013) Hybrid segmentation of left ventricle in cardiac MRI using gaussian-mixture model and region restricted dynamic programming. Magn Reson Imaging 31:575–584 8. Dakua SP (2015) LV segmentation using stochastic resonance and evolutionary cellular automata. Int J Pattern Recognit Artif Intell 29:1–26 9. Wang L, Pei M, Codella NCF et al (2015) Left ventricle: fully automated segmentation based on spatio-temporal continuity and myocardium information in cine cardiac magnetic resonance imaging (LV-FAST). Biomed Res Int 36758:1–9 10. Sanchez-ortiz GI (1999) Medical image computing and computer-assisted interventionMICCAI’99, vol 1679 11. Suinesiaputra A, Cowan BR, Finn JP et al (2012) Left ventricular segmentation challenge from cardiac MRI: a collation study. In: Camara O et al (eds) Statistical atlases and computational models of the heart. Imaging and modelling challenges. Lecture notes in computer science, vol 7085, pp 88–97 12. Lebenberg J, Lalande A, Clarysse P, Buvat I, Casta C et al (2015) Improved estimation of cardiac function parameters using a combination of independent automated segmentation results in cardiovascular magnetic resonance imaging. PLoS One 10 13. Ngo TA, Carneiro G (2013) Left ventricle segmentation from cardiac MRI combining level set methods with deep belief networks. In: IEEE international conference on image processing, pp 695–699 14. Constantinides C, Roullot E, Lefort M, Frouin F (2012) Fully automated segmentation of the left ventricle applied to cine MR images: description and results on a database of 45 subjects. In: Proceedings of the annual international conference of the IEEE engineering in medicine and biology society, pp 3207–3210 15. Wu Y, Wang Y, Jia Y (2013) Segmentation of the left ventricle in cardiac cine MRI using a shape-constrained snake model. Comput Vis Image Underst 117:990–1003 16. Lee H, Codella NCF, Cham MD, Weinsaft JW, Wang Y (2010) Automatic left ventricle segmentation using iterative thresholding and an active contour model with adaptation on short-axis cardiac MRI. IEEE Trans Biomed Eng 57:905–913 17. Kass M, Witkin A, Terzopoulos D (1988) Snakes: active contour model. Int J Comput Vis 1:321–331 18. Akram F, Garcia MA, Puig D (2017) Active contours driven by local and global fitted image models for image segmentation robust to intensity inhomogeneity. PLoS One 12 19. Tealab A, Hefny H, Badr A (2017) Forecasting of nonlinear time series using ANN. Future Comput Inform J 2:39–47

546

Md. Al Noman et al.

20. Gencay R, Liu T (1997) Nonlinear modelling and prediction with feedforward and recurrent networks. Phys D 108:119–134 21. Rahman MA, Ahmad M (2016) Movement related events classification from functional near infrared spectroscopic signal. In: International conference on computer and information technology, pp 1–6 22. Kumar R (2010) Snakes: active contour models. MATLAB central file exchange. https://www. mathworks.com/matlabcentral/fileexchange/28109-snakes-active-contour-models?focused= 5156463&tab=function

Chapter 46

Bangla Handwritten Digit Recognition and Generation Md. Fahim Sikder

1 Introduction Recognizing handwritten numerals is one of the emerging problems in the sector of computer vision. Automation of the banking system, postal services, form processing are the practical example of handwritten character recognition [5, 10, 21, 24, 25, 29, 33]. A lot of work already has been done with great accuracy in the recognition of English handwritten digits [3, 19]. Researchers used support vector machine, histogram of gradient oriented, neural network, etc., algorithm to solve these problems. Recently, a lot of focus has been drawn to the neural network architecture due to the wide availability of high-performance computing systems [1]. ANNs are computing system which is influenced by the organic neural network. Convolutional Neural Network is one of the architectures of neural network which makes it easy to recognize image with great accuracy. Besides English, a lot of work also done in Arabic, Chinese, Japanese, and Roman scripts [7–9, 12, 14, 15, 20, 30, 31]. But in the case of Bangla, not many works have been done and there is a chance for improvement. On the other hand, generating images is another outstanding image processing field recently caught the attention of researchers. Image generation can be used in art creation, fraudulent detection also can be applied in law enforcement. Generative Adversarial Network or GAN, another architecture of the neural network is been used to generate the image. Researchers also applied GAN to generate MNIST dataset but not much work has been done in other datasets. To mend this research gap on Bangla, we have implemented an architecture which recognizes Bangla handwritten digits at 99.44% accuracy using BHAND dataset which contains 70000 images of Bangla handwritten digits which are collected from 1750 persons. At the same time, we have implemented a semi-supervised generative adversarial network or SGAN to Md. Fahim Sikder (B) Jahangirnagar University, Savar, Bangladesh e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_46

547

548

Md. Fahim Sikder

generate Bangla digits. The paper is arranged as follows: Sect. 2 reviews the relevant works, Sect. 3 describes the proposed solution, Sect. 4 describes the result and lastly, Sect. 5 concludes the paper.

2 Related Works A lot of research works have been done on Bangla handwritten digit recognition using SVM [6], HOG [4], etc. Recently, loads of attention is being given on deep learning because of easy access to GPU (graphics processing unit). Using multilayer convolutional layer, pooling layer increases the performance of accuracy. Some of the legendary deep learning based architecture such as Alexnet [16], LeNet [18], and Inception V3 [32] took the accuracy of image recognition to the next level. MNIST recognition [17], CIFAR-10 database recognition [16] are some examples of that architecture. For Bangla handwritten recognition, numerous work has been done. But initially, it was troublesome for the researcher because of the limitation of a dataset [2]. But now, some great datasets are available for Bangla digit recognition. A deep belief network is being introduced where the author first used unsupervised feature learning, then it is followed by a supervised fine-tuning [27]. In [11], the author removed the overfitting problem and has an error rate of 1.22%. Besides digit recognition, few works have been done on digit generation. Researchers used different kinds of Generative Adversarial Networks (GAN) to generate digits or characters. Auxiliary Classifier GAN [23], Bidirectional GAN [13], Deep Convolutional GAN [26], and Semi-Supervised GAN [22] were used on MNIST dataset to generate digits.

3 Proposed Work In this work, we have proposed a architecture for digit recognition, which outperforms Alexnet [16] and Inception V3 [32] model at validation accuracy and error on the BHAND [11] dataset. Also, we have implemented Semi-Supervised Generative Adversarial Network (SGAN) for digit generation for the same dataset.

3.1 Dataset Description For recognition and generation, BHAND dataset has been used which contains 70000 handwritten Bangla digits. This is one of the biggest datasets of handwritten Bangla digits. This dataset is divided into three sets: Training set (50000), Testing set (10000), and Validation set (10000). These 70000 data is collected from 1750 persons. The images are gray scale and the dimension is 32 ∗ 32.

46 Bangla Handwritten Digit Recognition and Generation

549

3.2 Network Architecture For recognizing handwritten digit, we have proposed an architecture which consists of several convolutional layers, pooling layers, normalization layers, and dense or fully connected layers. In the first convolutional layer, we took the 32 ∗ 32 images as input from the dataset. As mentioned earlier, the images are gray scale, so it has 1 channel. In this layer, we have taken 32 filters which has the filter size of 2 ∗ 2. The output of this layer then goes into a second convolutional layer, which also has 32 filters and the size of those filters is 2 ∗ 2. Then, the outcome of the second convolutional layer feed into max pooling layer, which has the filter size of 2 ∗ 2 and the stride size is 2. This outcome then goes into a normalization layer. These convolutional layers, pooling layer and normalization layer, together we named it block. In a single block, the number of these layers could vary. The second block is composed of three convolutional layers, one max pooling layer, and another normalization layer. The amount of filters in the second block’s convolutional layers are 64 and the filter size is 3 ∗ 3. This max pooling layer has also 2 ∗ 2 filter size and stride of 2. Then, the third to sixth block consists of two convolutional layers, one pooling layer, and one normalization layer. Third block’s convolutional layer has 128 filters and the size of the filters is 5 ∗ 5, fourth block’s convolutional layer has 256 filters which has 5 ∗ 5 filter size, fifth block’s convolutional layer has 384 filters, sixth block’s convolutional layer has 512 filters and their filter size is 5 ∗ 5. And, all the blocks have the same pooling layer architecture. It has 2 ∗ 2 filter size and stride size 2. Figure 1 shows the blocks used in this architecture. The outcome of the sixth block is then fed into a fully connected layer, which has 1024 units, then we drop the 50% of the neuron for avoiding overfitting, and then the output is fed on the second fully connected layer which has 5120 units. Here, we also drop the 50% of the neuron. Till now, every layer used r elu activation function. The following equation [28] is how r elu works. R(z) = max(0, z) Now, the output is then fed into the last fully connected layer, which has 10 units because we have 10 class as output and here we have used so f tmax activation function. The following equation [28] is how so f tmax works. ez j s(z) j =  K k=1

ezk

The complete architecture of the recognizing part is shown in Fig. 2. Now for the digit generation part, here Semi-Supervised Generative Adversarial Network (SGAN) [22] is used for this task. Here, we have a generator and discriminator. We took random noise as input, then the noise goes to the generator, at the same time we took a sample from training dataset. The generator attempts to forge the sample from training dataset and both the real and fake data goes to the discriminator then the discriminator attempts to distinguish between the genuine and the

550

Md. Fahim Sikder

Fig. 1 Blocks of the architecture

fabricated one. Usually, in GAN, we train generator and discriminator concurrently and after training, we could discard discriminator because it is only used for training the generator. In SGAN, we alter the discriminator into a classifier and we discard the generator after the training. Here, generator is used to aid the discriminator during training. Figure 3 shows the complete architecture of the SGAN. In the generator, first, we took a random vector as input then we reshape it and then batch normalize it. Then we upsample the output. After that, we took a convolutional layer and pass the output through it. The convolutional layer has 128 filters and the filter size is 3 ∗ 3 also we used the same padding. We again use batch normalize and upsample in it. After that, we use another convolutional layer which has the same filter size and padding but it has only 64 filters and we again batch normalize it. The last two convolutional layers used r elu activation function. Now, the output is

46 Bangla Handwritten Digit Recognition and Generation

551

Fig. 2 Our architecture for digit recognition

Fig. 3 Architecture of SGAN

passed through the last convolutional layer which has one filter and the filter size and padding are like the same as others and it used tanh activation function. Now for the discriminator part, it is a multiclass classifier. We have used four convolutional layers. First convolutional layer takes 32 ∗ 32 images and it has 32 filters which have the size of 3 ∗ 3 also the strides of 2 to reduce the dimension of the feature vectors. Here, we have used Leaky Recti f ied LinearU nit activation functions. Then, we drop 25% of neurons for avoiding overfitting. Then, the output goes to the next convolutional layers which have 64 filters and the size and strides are the same as the last one. Then again, we drop 25% of the neuron and use batch normalization. In the third and fourth convolutional layer, the filter size is the same but has 128 and 256 filters, respectively. Then we flatten the output. In the end, we used two dense or fully connected layers. The last layer takes N + 1 units because discriminator could generate N + 1 outputs because of the fake label. Here, N is the number of total class and we used so f tmax activation function. We used binar ycr oss − entr opy loss function and Adam optimizer.

552

Md. Fahim Sikder

4 Experimental Analysis and Result We have implemented our architecture using BHAND dataset which has 50000 training image, 10000 testing image and 10000 validating images of handwritten Bangla numerals. It has 32 ∗ 32 image dimension and the number of the channel was 1. For recognizing the digit, we have also applied this dataset in popular Alexnet and Inception v3 model. We have run a total of 19550 steps in the training and achieved 99.44% validation accuracy. We have used r mspr op optimizer and categoricalcr oss − entr opy as loss function. The learning rate in our architecture was 0.001. A detailed analysis of our experiments is shown in Table 1. The validation accuracy and the validation error of our model is shown, respectively, in Figs. 4 and 5. For generating the Bangla handwritten image, we also used the same dataset. For generating an image, we have used the Semi-Supervised Generative Adversarial Network (SGAN). Here, we have built our model using generator and discriminator. Generator took a random vector as input. On the other hand from the real train dataset, an image goes to the discriminator. Generator tries to fool the discriminator by mimicking the real image. Then the discriminator discriminates the real and forges image. For our generator, we have used a combination of a fully connected layer, convolutional layer. Also, we need to normalize and upsample our data. For the discriminator, it also has a series of a convolutional layer and fully connected layer. Discriminator took the image as input to the input dimension is 32 ∗ 32. It used two Table 1 Comparison of our model with others for recognizing digit Model name Steps Validation accuracy (%) Validation error Alexnet Inception V3 Our Model

19550 19550 19550

97.74 98.13 99.44

Fig. 4 Validation accuracy of our model for recognizing digit

0.1032 0.07203 0.04524

46 Bangla Handwritten Digit Recognition and Generation

553

Fig. 5 Validation Error of our model for recognizing digit

Fig. 6 Output of our generation model at step 0, 100000, 200000, and 300000

loss functions: binar ycr oss − entr opy and categoricalcr oss − entr opy, whereas generator used binar ycr oss − entr opy. Here, we have used Adam optimizer where the learning rate is 0.002. We have also reshaped our data to −1 to 1 because of the usage of sigmoid and tanh activation function. After 300000 steps of training, we have got 0.368 loss of discriminator and 0.694 generator loss. From Fig. 6, we can

554

Md. Fahim Sikder

Fig. 7 Training loss of our model for digit generation

see the output of our SGAN. The first image (a) is from 0 step, the second image (b) is after 100000 and (c) and (d) image are after respectively 200000 and 300000 steps. The training loss is shown in Fig. 7.

5 Conclusion Loads of work have been done in the area of handwritten numeral recognition but still, there is an opportunity to improve and only a few works have been done in the area of digit generation. From that motivation, in this paper, we have proposed an architecture for recognizing Bangla handwritten digits which outperforms popular Alexnet and Inception v3 architecture using BHAND dataset. By adding a more convolutional layer and hyperparameter tuning could result in a better performance. Also, we have implemented the Semi-Supervised Generative Adversarial Network (SGAN) using the same dataset and successfully generate Bangla digits. In the future, we will try to reduce the discriminator’s training loss on SGAN. Acknowledgements The author is grateful to the anonymous reviewers for their comments that improved the quality of this paper, also thankful to Md. Rokonuzzaman Sir from ISTT and Umme Habiba Islam for their support and help.

References 1. Abir B, Mahal SN, Islam MS, Chakrabarty A (2019) Bangla handwritten character recognition with multilayer convolutional neural network. In: Advances in data and information sciences. Springer, pp 155–165

46 Bangla Handwritten Digit Recognition and Generation

555

2. Akhand M, Rahman MM, Shill P, Islam S, Rahman MH (2015) Bangla handwritten numeral recognition using convolutional neural network. In: 2015 International conference on electrical engineering and information communication technology (ICEEICT). IEEE, pp 1–5 3. Bengio Y, Lamblin P, Popovici D, Larochelle H (2007) Greedy layer-wise training of deep networks. In: Advances in neural information processing systems, pp 153–160 4. Bhattacharya U, Chaudhuri BB (2009) Handwritten numeral databases of indian scripts and multistage recognition of mixed numerals. IEEE Trans Pattern Anal Mach Intell 31(3):444–457 5. Bhowmik S, Malakar S, Sarkar R, Basu S, Kundu M, Nasipuri M (2018) Off-line Bangla handwritten word recognition: a holistic approach. Neural Comput Appl 1–16 6. Bhowmik TK, Ghanty P, Roy A, Parui SK (2009) Svm-based hierarchical architectures for handwritten Bangla character recognition. Int J Doc Anal Recognit (IJDAR) 12(2):97–108 7. Bozinovic RM, Srihari SN (1989) Off-line cursive script word recognition. IEEE Trans Pattern Anal Mach Intell 11(1):68–83 8. Broumandnia A, Shanbehzadeh J, Varnoosfaderani MR (2008) Persian/arabic handwritten word recognition using M-band packet wavelet transform. Image Vis Comput 26(6):829–842 9. Bunke H (2003) Recognition of cursive roman handwriting: past, present and future. In: Seventh international conference on document analysis and recognition, 2003. Proceedings. IEEE, pp 448–459 10. Bunke H, Bengio S, Vinciarelli A (2004) Offline recognition of unconstrained handwritten texts using hmms and statistical language models. IEEE Trans Pattern Anal Mach Intell 26(6):709– 720 11. Chowdhury AMS, Rahman MS (December 2016) Towards optimal convolutional neural network parameters for Bengali handwritten numerals recognition. In: Proceedings of 19th international conference on computer and information technology (ICCIT). Dhaka, pp 431–436 12. Dehghan M, Faez K, Ahmadi M, Shridhar M (2001) Handwritten farsi (Arabic) word recognition: a holistic approach using discrete HMM. Pattern Recognit 34(5):1057–1065 13. Donahue J, Krähenbühl P, Darrell T (2016) Adversarial feature learning. arXiv:160509782 14. El Qacimy B, Kerroum MA, Hammouch A (2015) Word-based Arabic handwritten recognition using SVM classifier with a reject option. In: 2015 15th International conference on intelligent systems design and applications (ISDA). IEEE, pp 64–68 15. Koerich AL, Sabourin R, Suen CY (2005) Recognition and verification of unconstrained handwritten words. IEEE Trans Pattern Anal Mach Intell 27(10):1509–1522 16. Krizhevsky A, Sutskever I, Hinton GE (2012) Imagenet classification with deep convolutional neural networks. In: Advances in neural information processing systems, pp 1097–1105 17. LeCun Y, Boser B, Denker JS, Henderson D, Howard RE, Hubbard W, Jackel LD (1989) Backpropagation applied to handwritten zip code recognition. Neural Comput 1(4):541–551 18. LeCun Y, Boser BE, Denker JS, Henderson D, Howard RE, Hubbard WE, Jackel LD (1990) Handwritten digit recognition with a back-propagation network. In: Advances in neural information processing systems, pp 396–404 19. LeCun Y, Jackel L, Bottou L, Brunot A, Cortes C, Denker J, Drucker H, Guyon I, Muller U, Sackinger E et al (1995) Comparison of learning algorithms for handwritten digit recognition. In: International conference on artificial neural networks, vol 60. Perth, Australia, pp 53–60 20. Liu CL, Koga M, Fujisawa H (2002) Lexicon-driven segmentation and recognition of handwritten character strings for Japanese address reading. IEEE Trans Pattern Anal Mach Intell 11:1425–1437 21. Madhvanath S, Govindaraju V, Ramanaprasad V, Lee DS, Srihari SN (1995) Reading handwritten us census forms. In: Proceedings of the third international conference on document analysis and recognition, 1995, vol 1. IEEE, pp 82–85 22. Odena A (2016) Semi-supervised learning with generative adversarial networks. arXiv:160601583 23. Odena A, Olah C, Shlens J (2016) Conditional image synthesis with auxiliary classifier gans. arXiv:161009585 24. Pal U, Roy K, Kimura F (2009) A lexicon-driven handwritten city-name recognition scheme for indian postal automation. IEICE Trans Inf Syst 92(5):1146–1158

556

Md. Fahim Sikder

25. Pal U, Roy RK, Kimura F (2012) Multi-lingual city name recognition for indian postal automation. In: 2012 International conference on frontiers in handwriting recognition (ICFHR 2012). IEEE, pp 169–173 26. Radford A, Metz L, Chintala S (2015) Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv:151106434 27. Sazal MMR, Biswas SK, Amin MF, Murase K (2014) Bangla handwritten character recognition using deep belief network. In: 2013 International conference on electrical information and communication technology (EICT). IEEE, pp 1–5 28. Sharma S (2018) Activation functions: neural networks. https://towardsdatascience.com/ activation-functions-neural-networks-1cbd9f8d91d6 29. Srihari SN, Shin YC, Ramanaprasad V, Lee DS (1995) Name and address block reader system for tax form processing. In: Proceedings of the third international conference on document analysis and recognition, vol 1. IEEE, pp 5–10 30. Srihari SN, Yang X, Ball GR (2007) Offline Chinese handwriting recognition: an assessment of current technology. Front Comput Sci China 1(2):137–155 31. Su T (2013) Chinese handwriting recognition: an algorithmic perspective. Springer Science & Business Media 32. Szegedy C, Liu W, Jia Y, Sermanet P, Reed S, Anguelov D, Erhan D, Vanhoucke V, Rabinovich A (2015) Going deeper with convolutions. In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp 1–9 33. Yacoubi AE (2001) Handwritten month word recognition on brazilian bank checks. In: Proceedings of the sixth international conference on document analysis and recognition. IEEE Computer Society, p 972

Chapter 47

Portable Mini-Weather Station for Agricultural Sector of Rural Area in Bangladesh Nazib Ahmad, Thajid Ibna Rouf Uday, Md. Toriqul Islam, Rayhan Patoary, Md. Mostasim Billah, Nuhash Ahmed and Farhana Sharmin Tithi

1 Introduction In our day-to-day life, climate and weather play a very important role. Weather forecasting and monitoring system give us the opportunity to know different climates behavior such as temperature, humidity, atmospheric, air pressure, light intensity, rainfall, wind speed, wind direction, UV, amplitude, etc. A recent survey found that everyday 41% check weather forecast one time a day and 24% check multiple times. Technology plays a vital role in weather analysis. People check weather forecast by technology: 66% in online, 48% on TV, and 37% by using a mobile device. And just 7% read from the newspaper or other print media. The weather station is used as a N. Ahmad (B) · T. I. R. Uday · Md. Toriqul Islam Ai Robotics Asia Ltd, Dhaka, Bangladesh e-mail: [email protected] T. I. R. Uday e-mail: [email protected] Md. Toriqul Islam e-mail: [email protected] R. Patoary · Md. Mostasim Billah · N. Ahmed Department of SWE and CSE, Daffodil International University, Dhaka, Bangladesh e-mail: [email protected] Md. Mostasim Billah e-mail: [email protected] N. Ahmed e-mail: [email protected] F. S. Tithi Department of CSE, Daffodil International University and Global Emerging Technology Network, Dhaka, Bangladesh e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_47

557

558

N. Ahmad et al.

climate and weather data monitoring system which gives exact and details information in agriculture sector. Climate is one of the most imperative field parameters that ensures plant growth and its output to fulfill high farming productivity, and every farmer needs to know weather forecasting and correctly check the weather parameter [1]. Each plant is susceptible to certain growing conditions such as air temperature, relative humidity, soil temperature, wind, light, etc. [2]. Therefore, farmers have to understand various climatic conditions of their farms. By weather stations and its various sensors, a farmer can collect accurate environment measurement of weather parameter [3]. Also, circumstances of weather predict disease attacks of crops and help the farmer to increase crop production. The research comes with a portable user-friendly mini-weather station that we design and implement for Bangladeshi firmer. The process of design this prototype is focused on most effective sensor data algorithms with real-time response. The weather elements are continuously sensing using sensors. This research has the process of air temperature and air humidity-B and UV index monitoring, air speed, air pressure, and altitude soil moisture, soil temp, and CO2 detecting. This mini-weather station is prototyped and developed in Ai Robotics Asia Ltd with Global Emerging Technology Network with various functionality weather sensors which focus on four main agricultural-based weather parameters—soil status for crops, air quality status, humidity, weather prediction with a single portable mini compact, cost-effective, user-friendly for farmer solutions, and helpful to increase production Argo product. This mini-weather station is very easy to set up and it will not need any technical knowledge to set up, integrate, and keep up except smart phone usage knowledge. The device will be greatly reduce the hassle of farmers and increase the production along with the developing countries as well as Bangladesh. In this paper, the authors developed a cost-effective weather station using Arduino microcontroller with various sensors and Bluetooth module with visualization using mobile application. The rest of this workshop organized as follows: Sect. 3 gives the materials and explains the methods employed; Sect. 4 presents the system implementation and testing procedures.

2 Related Works Numerous economic portable mini-weather stations are designed and used in many countries [4]. Jain, Abhinav, and K. Selvakumar described irrigation or climate control systems at a landscape site. They also proposed a wireless system that provided observation of environmental, soil, or weather statistics [5]. Kirankumar G. Sutar showed the use of many sensors that are able to continuously read some factors that indicate the weather conditions such as temperature, humidity, and light intensity in an industrial environment [6]. Karimi has proposed Web-based monitoring system using wireless sensor networks [7]. A system design on the reduction of agricultural output helps in reducing the effects of weather changes on agricultural output for the benefits of farmers. Considering aspects like low cost, portability, and durability,

47 Portable Mini-Weather Station for Agricultural …

559

Gaurav et al. [8] has designed a GSM-based weather monitoring system for solar and wind energy generation. Android-Based mini-weather station was discussed briefly by Budianto, Anwar, Yoiko Rashaki, and Ginaldi Ari Nugroho [9] and smart irrigation system was discussed by Ashwini, B., and V. Udayarani [10]. IOT-based field monitoring system was implemented by Gautami, U., and K. Phani Babu [11] and various sensors for weather prediction were included by Basnet, Barun, and Junho Bang [12].

3 Materials and Methods To realize the overall functionality of mini-weather station, a system block diagram is given in Fig. 2 and the ATmega328P microchip is performed as processor with Arduino Uno R3 which acts as the motherboard. All the sensors are linked with ADC to collect data from various sensors and sent to the processor. Processor takes the result from sensors data and mobile applications server retrieve data by HC05 which is a Bluetooth Module. The architecture of mini-weather station (Fig. 1) is prototyped with focusing on below power utilization and energy saving technique. All data of sensors are visualized with icon by graphical user interface (GUI) as farmer [13] or the user can easily understand the process [14] and can use this mini-weather station to increase their cultivation productivity. The developed device that gives us valuable information about air temperature and air humidity using sensor DHT11 sensor, UV index, and UV-A lamp monitoring using UV-B sensor, between 30,000 and 110,000 Pa air pressure and with I2C interface using Barometric Pressure BMP085, combustible gas sensing with high sensitivity LPG detecting threatening gas using MQ2 Gas Sensor, measuring soil moisture using moisture sensor, operating −55 to +125 °C with accuracy of ±0.5 °C of soil using DS1820 sensor, using HC-05 Bluetooth Module to transfer data with Serial Port Protocol with 2.4 GHz radio transceiver with modulation of 3Mbps.

Fig. 1 Architectural diagram of a mini-weather station

560

N. Ahmad et al.

Fig. 2 System block diagram of mini-weather station

3.1 Arduino Uno R3 The Arduino UNO R3 (Fig. 3) is a microcontroller board based on dual-inline package (DIP) ATmega328 AVR microcontroller and is developed by Arduino.cc. Arduino Uno R3 microcontroller is used to control the system and connect with various sensors to read analog data from various weather variables. In this prototype, this Arduino board fulfills the purpose of the research project to become it is more cost-effective. With 14 Digital I/O pins with PWM pints, it is simple and easy to connect with pc and updating the framework is also a great advantage of this board for this project.

3.2 HC-05 Bluetooth Module HC-05 as shown in Fig. 4 is a Bluetooth module. To transfer data of sensors from processor to mobile application, we used HC-05 module which has serial port protocol and designed with wireless serial connection setup. The mobile application is connected with the processor using this Bluetooth and it also operates on serial port protocol (SSP). The module has 38,400 default baud rate, 8-bit data, stop bit by 1, and no parity options. As we are making sure that the baud rate of the Bluetooth module is synchronized with that of the Arduino so that there is no loss of data and proper communication is achieved.

47 Portable Mini-Weather Station for Agricultural …

Fig. 3 Arduino Uno R3

Fig. 4 HC-05 bluetooth module

561

562

3.3 Sensors Used (a) Digital Humidity Temperature Sensor (DHT11) (Fig. 5). (b) UV Sensor (Fig. 6). (c) Barometric Pressure BMP085 Barometric Pressure BMP085 (Fig. 7) (d) Soil Moisture Sensor The Soil Moisture Sensor (Fig. 8) (e) Soil Temp Sensor DS1820 Soil Temp Sensor DS1820 (Fig. 9) (f) MQ2 Gas Sensor Sensitive material of MQ-2 gas sensor (Fig. 10).

Fig. 5 Digital humidity temperature sensor (DHT11)

Fig. 6 UV sensor

N. Ahmad et al.

47 Portable Mini-Weather Station for Agricultural …

563

Fig. 7 Barometric pressure BMP08

Fig. 8 Soil moisture sensor

Fig. 9 Soil temp sensor DS1820

3.4 Android Interface In this project, graphical user interface of android-based mobile application is performed to showcase the information obtained in the form of temperature (air and soil) and humidity, altitude, moisture, gas, UV index, pressure which is shown in Fig. 11. It enables Android devices with any Bluetooth device associated with serial port profile. It acts as VT-100 terminal emulator for Bluetooth connections. It is user-friendly and can measure nonstop weather parameter. We set up 10 s of pause settings to read data correctly.

564

Fig. 10 MQ2 gas sensor

Fig. 11 Mobile application user interface

N. Ahmad et al.

47 Portable Mini-Weather Station for Agricultural …

565

4 System Implementation and Testing 4.1 Circuit Diagram and Simulation The circuit diagram (Fig. 12) is showing all the necessary components and useful sensors are connected with Arduino Uno R3. Moisture, UV, MQ2, Barometric Pressure BMP085 sensors are connected with analog pins as those are giving analog data and DHT11 and DS1820 are digital sensors connected with digital pin 1 and 2. Bluetooth Module HC05 is connected with TX and RX which is required to transfer data. All ground is linked with a common ground. All channel and sensors are connected with +5 V channel.

4.2 Operation Flow Chart When a farmer or user is using mobile application, then they receive data from miniweather station. The mobile application connects with Arduino Microcontroller via Bluetooth to read appropriate data. Microcontroller requests sensors to collect analog data. In condition, Data are process and send to mobile application using Bluetooth module. After the entire process is completed, then this entire process will start again

Fig. 12 Circuit diagram and simulation

566

N. Ahmad et al.

Fig. 13 Operation flow chart of mini-weather station

carry out the operation for data collecting via mini-weather station. In Fig. 13, an operational flow chart is described.

4.3 Core Circuit Implementation and Final Model Implementation See Fig. 14.

Fig. 14 Core circuit implementation and mini-weather station final model implementation

47 Portable Mini-Weather Station for Agricultural …

567

4.4 Program Sample

Start program Add header file, Define Pin number, Define variable, Void Loop { Get temperature; and air Get pressure; Get atm Calculate atm from pressure; Get altitude; Get uv; Get moisture; Get Humadity; Get gas; }

// Get the temperature, bmp085ReadUT, from soil // Get the pressure from air //Uncompensated calculation - in Meters //connect UV sensors to Analog // measuring moisture from soil & air // measuring Humadity from soil & air

End program

4.5 Testing and Results During testing, we use serial monitor to see the sensor’s data. Sensor’s data is tested under various weather circumstances. We have also tested with different soil conditions and different weather conditions like windy, rainy, and normal. During debugging, all data are shown below at serial monitor. This data is then sent via Bluetooth model to mobile Android application which we design for user or farmers so that they can know and analyze the recorded data anytime and anyplace. They also can take a decision based on weather forecasting data and variable and make the right decision for their farms. In Fig. 15 shows the testing and debugging Result.

Fig. 15 Testing and debugging result

568

N. Ahmad et al.

5 Conclusion This mini-weather station combines the features of multiparameter weather sensors which are mainly related to soil and air temperature, humidity, atmospheric pressure, and soil moisture and enables a mobile application for the user or farmer. Mobile application monitoring system is very useful which can easily read the weather forecasting data and take a proper action for their crops. The device design focuses on low cost and high accuracy results to solve the current situation of farmer is purchasing and using the ability. This will be helpful to produce more agricultural product and using this device requires no training, maintenance engineer, technical knowledge, and mostly no need for Internet connection. The device will be greatly reduce the hassle of farmers and prolong the fertility of farming along with the development of the countries as well as Bangladesh. Future Work We have a plan to extend our project mini-weather station with GSM and GPRS system [15] which will be compatible for wireless internet connection. The device will perform well on both offline and online platform. We also wanted to add more parameters like energy saving system by using renewable energy solar panel [16, 17]. We will introduce our dashboard [18] in large farming [18] and also, we will more focus on the stability, portability, and accuracy of the product. We will include IOT-based prototyped to faster with cloud computing [19]. In our current product, we are making efforts for only generalized agriculture solution which might not be possible for each crop observation as every crop has the unique specialization to take care of. We will focus on making some AI-based software [20–22] and also Bangla Agri-bot to know all weather condition and crop condition. Create support for taking special care of every crop with based on their characteristics. So, the farmers can rely on this product and would never have to search for any services in the market. It will bring all solutions in one package with low-cost consideration and will also save lots of money by precure farming. It will also be used for correct observation for each crop by measuring the right value of weather forecasting and ensure smart farming. Acknowledgements The authors would like to express profound gratitude to AI Robotics Asia Ltd and Global Emerging Technology Network for their endorsement as well as technical supervision. This mini-weather station is prototyped and developed in AI Robotics Asia Ltd with the partnership of Global Emerging Technology Network.

References 1. Suškeviˇcs M et al (2018) Learning for social-ecological change: a qualitative review of outcomes across empirical literature in natural resource management. J Environ Plan Manag 61.7:1085–1112 2. Rasiah R et al (2018) Climate change and sustainable development issues: arguments and policy initiatives. J Asia Pac Econ 23.2:187–194

47 Portable Mini-Weather Station for Agricultural …

569

3. Li H et al (2018) Web-based irrigation decision support system with limited inputs for farmers. Agric Water Manag 210:279–285 4. Jain A, Selvakumar K (2018) Intelligent automated IoT based irrigation system. Int J Adv Res Comput Sci 9.2:512 5. Sutar KG (2015) Low cost wireless weather monitoring system. Int J Eng Technol Manag Res 1(1):48–52 6. Karimi, Navab, et al (2018) Web-Based monitoring system using Wireless Sensor Networks for traditional vineyards and grape drying buildings. Comput Electron Agric 144:269–283 7. Gahlot N et al (2015) Zigbee based weather monitoring system. Int J Eng Sci IJES4: 61–66 8. Gaurav D et al (2014) A GSM based low cost weather monitoring system for solar and wind energy generation. In: The fifth international conference on the applications of digital information and web technologies (ICADIWT 2014). IEEE 9. Budianto A, Rashaki Y, Nugroho GA (2018) Mini automatic weather station development for android based weather parameters monitoring. Adv Sci Lett 24.12:9509–9515 10. Ashwini B, Udayarani V. (2018) Surveillance of crop-field with smart irrigation system. Int J Adv Res Comput Sci 9.1 11. Gautami U, Babu KP (2018) Field monitoring system using IoT 12. Basnet B, Junho B (2018) Sensors and data analytics in agriculture: a review 13. Sarkar S (2018) Automatic irrigation system. Dissertation, University of Technology 14. Huynh L (2018) Environmental monitoring system 15. Mahmood SN (2018) GSM interaction based real time climate change monitoring technique. Kirkuk Univ J Sci Stud 13(2):1–17 16. Sun Y (2018) Deep neural network regression and sobol sensitivity analysis for daily solar energy prediction given weather data. Purdue University, Dissertation 17. Li L et al (2018) Sustainable energy management of solar greenhouses using open weather data on MACQU platform. Int J Agric Biol Eng 11(1):74–82 18. XIN L (2018) Statistical modelling and analysis for regional climate change. Dissertation 19. Akhtar S Integrated IoT (Internet of Things) system solution for smart agriculture management. U.S. Patent Application No. 15/451,420 20. Kaneriya S et al (2018) A range-based approach for long-term forecast of weather using probabilistic Markov model. IEEE international conference on communications workshops (ICC Workshops). IEEE, 2018 21. Oliveira I et al (2018) A scalable machine learning system for pre-season agriculture yield forecast. arXiv preprint arXiv:1806.09244 22. Rahul DS et al (2018) IoT based solar powered Agribot for irrigation and farm monitoring: Agribot for irrigation and farm monitoring. 2nd international conference on inventive systems and control (ICISC). IEEE, 2018

Chapter 48

Appliance of Agile Methodology at Software Industry in Developing Countries: Perspective in Bangladesh Abdus Sattar, Arif Mahmud and Sheak Rashed Haider Noori

1 Introduction Improvement in the software industry in Bangladesh has become a critical factor as it makes the business’ advertising and marketing obligations simpler to fulfill. In different phrases, software development is used and accomplished so that it will offer growth and value to the requirement of the consumer. The challenges for software companies are to choose the accurate methods for development and to accomplish the company’s improvement goals. The achievement pace of development works can be improved through the consumption of a methodology which is acceptable for the precise features of these software projects. For the last few years, a huge variety of software design model has been elaborated, consequently, and agile method can be considered an efficient one [1, 2]. Agile in essence is an iterative, lightweight, software design, and development methodology. Agile methodology splits a total project into tiny size iterations which consumes the similar interval length. Unlike other methods, the agile method is supported by consuming feedback that guarantees the better users requirement fulfilment [3]. However, this method has some limitations too such as quality measurement, user feedback, user adaptability, etc. ICTization is an important management level framework [4] and merging these two methodologies can lessen these shortcomings

A. Sattar · A. Mahmud (B) · S. R. H. Noori Daffodil International University, 102, Shukrabad, Mirpur Road Dhanmondi, Dhaka 1207, Bangladesh e-mail: [email protected] A. Sattar e-mail: [email protected] S. R. H. Noori e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_48

571

572

A. Sattar et al.

to some great extent and can be very effective form the developing countries like Bangladesh viewpoint. The primary goal of this study is to discover and explore the major issues which manipulate the decision to choose the most widespread software development model for a specific project. The four research objectives are: 1. Adopting the agile method as one of the services in Bangladesh and receive maximum benefits. 2. To find out the adoption rate of agile methods in Bangladesh and compare with other methods. 3. To find out the main benefits derived from agile methods in Bangladesh to increase popularity. 4. To address the challenges in adopting the agile method and procedure to lessen them. To choose an appropriate methodology is one of the primary concerns for those countries where the ICT industry will play a leading role in the near future. Besides, the challenges need to be addressed and solved. Therefore, this paper is systematized in the following approach: Sect. 2 depicts the cases studied in developing nations; Sect. 3 describes the hypothesis along with the motivations behind; Sect. 4 defines the data collection methodology; Sect. 5 presents the achieved result and conclusions are drawn in Sect. 6 with some ideas for future research.

2 Agile Method in Developing Countries Agile methodology has already received enough interest in developed nations. However, deficiencies of experimental studies have been observed in developing countries. These issues are required to be addressed properly since developing nations are currently playing the lead role in software outsourcing. In developing countries, the business enhancement provides an opportunity to make competitive in global market and dramatic improvements in economic growth. Since agile methods help to increase customer satisfaction though the requirements are continuously updated [2]. In addition, it provides an effective management, quality assurance, and productivity and maintenance practices to develop quality products. Consumption and contribution of agile methodology in four different developing countries were studied to ensure its prospect in Bangladesh.

2.1 Case Study 1 Based on the study [5], it can be found that the agile design model is efficient in Philippine to solve the customers’ necessities, develop knowledge sharing, and build better teamwork. It is also observed that the software developers have generously

48 Appliance of Agile Methodology at Software Industry …

573

adopted the agile design model to solve their assigned project-oriented tasks. Some key issues can be noted: • Agile methodology can play an important role in outsourcing overcoming the cultural divergence between both clients and developers. • The enhanced alliance which is developed through agile design can become more imperative in comparison to our expectations especially in societies where communism is followed.

2.2 Case Study 2 It can be found that [6] the software organizations provided remarkable elasticity in the GDP of Croatia on the same era while the reduction in GDP was observed. The paper has also observed that most of the software firms, large to small, have previously implemented agile design methodology. Main findings mentioned from this research are given below: • 33% of revenues are received from the development firms and these organizations offer software outsourcing services in North America and Europe. • Mostly small to medium size development firms are found to utilize agile development model nationwide. • Primarily medium to small size development firms are found to practice agile design model both in locally and internationally. • The local policies, official provision are presently assisting them to operate internationally and helping to achieve more profits in Croatia for longer period.

2.3 Case Study 3 A quantitative study was used in this paper [7] to analyze the accomplishment of information management task in their companies. Their paper has claimed to use the valid sample and survey constraints. It is observed that 53% of the software organizations are presently operating agile model in Jamaica. Several benefits were also found from the survey like user fulfilment, enhanced team efficiency, and speedy growth rate. The standard approval rate of agile model active in Jamaican software organizations would aid the opportunity for a long-term scrutiny to track the recognition inclination.

574

A. Sattar et al.

2.4 Case Study 4 Several issues related to achievement and failure possibilities of agile methods in Zambia are studied. Several challenges are addressed and solutions are also proposed [8]: • Project management can be considered as one of the vital tasks that are not exceptional to agile model in software companies. • Dearth of resources can be considered important challenges. Therefore it is suggested that open source and cloud platform can be used to curtail the budget of execution. • An obligatory financial allocation is not appropriate for instigating the agile supported projects. But it can be solved using strong interpersonal skill among the parties. • Cordial working relationship can be considered as one of the most challenging issues for agile methods. Frustration might develop among developers when they are not satisfied and it is necessary to develop a good relation with developing team. In addition, this method is currently contributing to software industries of India, Pakistan, Malaysia, Thailand, etc., to some great extent.

3 Motivation and Hypothesis Agile method can improve the rate of success and solve problems of heavyweight methods. On the contrary, problem might arise for longer period project works which can affect the decision development procedure. So, some other factors are observed to coexist with agile methodology and plan-driven development in many organizations [9]. However, in all cases, the primary challenge is the collaboration among various approaches that might affect the larger companies [10]. Therefore, the project team needs to apply the most suitable agile method for the development but the present model cannot always fulfill the customer requirements and fail to address the challenges. We have proposed a managerial framework, ‘ICTization’ which aims to develop strategy and assess the success possibility, harmonization and regulate the managerial accomplishments, dissemination, and allocation of services among the eligible users. Moreover, it is also responsible for taking decisions and examines the success probabilities of the allocated assignments along with possibilities of mistakes, merges the vital services with viable explanations and service consumption becomes easier. This framework is a managerial tool which is used to synchronize the several e-services. The user adaptability and efficiency of any e-services can be measured through user feedback, KPI index, etc. Our idea is to consider this agile methodology as one of the services and receive maximum benefits out of it.

48 Appliance of Agile Methodology at Software Industry …

575

• The framework will help to find out the breakdown factors and challenges faced by the developers in Bangladesh. • This framework can help to distribute and share the Knowledge of Agile method’s tools and technique in Bangladesh. • This framework can promote and increase the success probability of agile implementation in Bangladesh. • This framework will help to evaluate the technical and non-technical factors which manipulate the success and failure of agile implementation in Bangladesh. • This framework will provide better service in terms of sending, receiving and synchronizing information to the deserving and will reduce the gap between the sectors like human resources and information system. • This framework will be user-friendly, less complicated, and will allow knowledge sharing through socialization. • This system will measure the service feasibility and capability along with the popularity and users’ adaptability before launching the services. This study is based on several hypotheses and can be outlined as follows: • Cultural infrastructures in organizations have positive impact that facilitates a learning environment. • Facilitator and coach approach adopted by the IT project managers have positive impact on software development projects. • Room for self-organization within development teams has positive impact on software development projects. • Ability to customize the working methodology has positive impact on to balance between customer requirements and the development team’s needs as the product evolves. • Extreme chaos (i.e., undisciplined agility) has negative impact on extreme predictive planning. • The values of the project as set by the customer at every stage have positive impact on project’s development. • flexible enough to deliver necessary documentation when asked to do so by the customer without drawing heavily on the project’s budgetary constraints have positive impact on project’s development. • Courage to tackle the next project probably using a similar methodology has positive impact on project’s development.

4 Survey and Data Collection Procedure Structured groups of research goals are the most vital parts for the survey as it can simplify many decisions comprised in survey procedure. Every part on the questionnaire was included Research Questions and Survey Questions. All questions are not needed for the purposes of the research. But, to ensure selected questions in the questionnaire for the survey can meet the research goals properly (Table 1).

576

A. Sattar et al.

Table 1 Correspondence between research questions and survey questions Part of questionnaire

1st

2nd

3rd

4th

Survey question

7

4

6

3

5th 3

Research question

5

3

6

2

3

Table 2 “Profile” of the software industry in Bangladesh SL. no.

Company name

Address Established year Before 2000

BASIS certified

2000–2017

CMMI level

✓

Level-1

✓

Level-2

✓

1.

ERAIT Ltd.

Dhaka

2.

IPCP Services

Dhaka

3.

MKB technologies

Rangpur

✓

Level-3

✓

4.

ZSI Bd.

Rangpur

✓

Level-4

✓

5.

PCN (Pvt) Ltd.

Dhaka

Level-5

✓

✓

✓

4.1 The Target Population In Bangladesh, maximum IT companies are established after 2000–2017. Besides, a lower amount IT companies are established before 2000. The main industrial section is IT Consultancy, Financial sector, Game, e-commerce, and others. A tiny amount of IT companies are not certified from BASIS, remaining all companies are certified. Not only certified BASIS but also they follow CMMI maturity level (Table 2).

4.2 The Mode of Administration To develop the survey, two types of interviews were conducted. • Face-to-face Interview Survey: This technique implies the paper questionnaire and the presence of the interviewer. Totally, 25 software companies were initially selected for the survey where 16 companies responded. • Online Survey (Mail & Telephone): surveys paper questionnaires were sent to the respondents by mail, totally 10 software companies were initially selected for the survey where 4 companies responded.

48 Appliance of Agile Methodology at Software Industry …

577

4.3 Developing the Questionnaire A questionnaire was designed with 25 questions on Agile Methods (Scrum, XP, and DSDM) and 9 questions on Company Demography. There are two blocks of total of 34 questions. • Block 1. On the Agile Methods: It includes 24 multiple-choice questions and 1 open question. The title of the questionnaire was 1. Requirements Analysis 2. Planning/Design 3. Development/Coding 4. Testing 5. Miscellaneous. All questions were based on Roles and Responsibilities, Artifacts of Agile Method specially Scrum, XP and DSDM. • Block 2. On the Company: The questionnaire contained 9 questions. It includes 7 multiple-choice questions and 2 open questions. All questions were based on: 1. Information of the company/organization 2. City where the company is located 3. Some information (Membership of BASIS, CMMI level).

4.4 Developing Data Collection and Data Processing Plans When the preliminary design strategy gets completed, collection and processing of data need to be initiated. Our sample consisted of 3 large firms with over 50 employees, 8 medium firms with over 21–50 employees, and 9 small firms with below 20 employees. All the firms had business focus on IT and IT consulting. Later, face-to-face interview and Online Survey (Mail & Telephone) data were combined.

5 Results and Analysis 5.1 Adoption of SDM The respondents were asked whether or not they were following a Software Development Method. As displayed in Fig. 1. 65% of the respondents are currently using and involved with agile methodologies in their developing process. These shows that almost 3/5 of the companies are adopting agile methodologies as a working practice. However, 10% has responded that agile methodologies are not being used in their development process. And 25% companies are used their own techniques and tools.

578 Fig. 1 Percentage of SDM (UTAT = use techniques and tools)

Fig. 2 Percentage of SDM with ICTization framework

5.2 Adoption of SDM and ICTization Framework See Fig. 2.

5.3 Adoption of Scrum See Table 3.

A. Sattar et al.

48 Appliance of Agile Methodology at Software Industry … Table 3 Adoption of Scrum

Scrum activities

No. of responses

Role of product owner

15

23

Customer involvement

12

19

Role of design team

Table 4 Adoption of XP

Table 5 Adoption of DSDM

579

Percent (%)

8

13

Team size of SW developers

13

20

Significance of framework

16

25

Total

64

100

XP activities

No. of responses

Percent

Role of tracker

1

4

Customer involvement

3

11

Role of programmer

7

26

Team size of SW developers

3

11

Significance of method

13

48

Total

27

100

DSDM activities

No. of responses

Percent

Delivery Time

Fixed

3

15

Resource

Fixed

3

15

Functionality

Changeable

Total

14

70

20

100

5.4 Adoption of XP See Tables 4 and 5.

5.5 Adoption of DSDM The survey has come up with a decision that different types of agile methods are used in Bangladeshi IT Companies. It was found that “Scrum” was leading the way at 58%, followed by “XP” at 24% and “DSDM” at 18% as shown in Fig. 3.

580

A. Sattar et al.

Fig. 3 The percentages of different agile methods used by the respondents

6 Discussion and Future Work A method choice for development is most important because implementing an incorrect methodology could result in slippages, lack of communication and administrative overheads and quality of a develop software, leading to customer dissatisfaction. Within the beyond decades, fast-paced evolution of software development methodologies improvement methodologies has affected substantial upgrades in software quality. On the other hand, agile methodologies became general to deal with a few shortcomings of conventional methodologies like heavy documentation, loss of productivity, reliability, and ease. Nowadays developers have been developing different methods which simplify the way of software development and some of the methods can be combined. Agile contains different approaches, however, Scrum (58%), XP (24%), and DSDM (18%) are found to be the most popular software design and development methods in Bangladesh. The software industry in Bangladesh is methodological and technological and our study adds evidence in support of it. Results were achieved in different categories like the percentage of respondents who follow software development method, adoption of Scrum, adoption of XP, adoption of DSDM, and the percentages of different agile methods used by the respondents. Based on the result, it can be stated that the agile method is the most popular software design and development methodologies and its popularity can be enhanced by 81% through merging with ICTization framework. In short, this paper has worked with three issues that meet our research objectives. First, case studies were used to find out the contribution of agile method in software firms in different developing countries. Second, the survey was done to find out the present utilization of this method in software industry in Bangladesh. Lastly, it was tried to increase the popularity of this method in Bangladesh though lessening the challenges and merging with other frameworks. The results achieved from this paper

48 Appliance of Agile Methodology at Software Industry …

581

are anticipated to notify the software developers on the utilization and practice of software development model from the context of Bangladesh. Future research will focus on the framework refinement by way of relating it with some other agile methods if possible supported by pragmatic feedback gained from real project observations, and quantitative data. This work can get enhanced in order to achieve methodological engineering especially through combining CEFAM with situational method engineering (SME) procedure. Acknowledgements We are grateful to students in the Department of Computer Science and Engineering of Daffodil International University for their participation to collect data from different software organizations. We are also thankful to the HR department of ERAIT Ltd, IPCP Services, MKB Technologies, ZSI BD, and PCN (Pvt.) Ltd. We pay our sincere gratitude to the participants who voluntary responded to the emails and attended face-to-face interviews.

References 1. Ramya Krishna TS, PhaniKanth CH, Phani Krishna CHV, VamsiKrishna TV (2011) Survey on extreme programming in software engineering. IJCTT 2 2. Dave R (1997) Patterns of success in the indian software industry 3. Boehm B (2007) A survey of agile development methodologies, 1–19 4. Mahmud A, Sattar A (2013) ICTization framework: a conceptual development model through ICT modernization in Bangladesh. Department of Computer Science and Engineering, Britannia University. In: 2nd International conference on advanced computer science applications and technologies—IEEE 2013. Sarawak, Malaysia (in press) 5. Sison R, Yang T (2007) Use of agile methods and practices in the Philippines. In: 2007 IEEE, college of computer studies 14th Asia-Pacific software engineering conference, De La Salle University, Manila, Philippines 6. Vlahovic N et al (2016) Study of software development in developing countries: a case of croatia. Int J Commun 10 1998–4480 7. Chevers DA, Chevers Whyte C (2015) The adoption and benefits of agile software development methods in Jamaica. In: Adoption of agile software development methods in Jamaica, twentyfirst Americas conference on information systems, Puerto Rico, 2015 8. Douglas K et al 2017 (Challenges of agile development and implementation in a developing country: a Zambia case study. J Comput Sci Mulungushi University, School of Science, Engineering and Technology, Kabwe, Zambia 9. Tomanek M, Cermak R, Smutny Z (2014) A conceptual framework for web development projects based on project management and agile development principles. Faculty of Informatics and Statistics, University of Economics, Prague. In: Czech Republic 10th European Conference on Management Leadership and Governance (ECMLG), Zagreb, Republic of Croatia, 13–14 10. Taromirad M, Ramsin R (2008) CEFAM: comprehensive evaluation framework for agile methodologies. In: 32nd annual IEEE software engineering workshop, 978-0-7695-3617-0, 1550–6215. https://doi.org/10.1109/sew.2008.19

Chapter 49

A Novel Approach for Tomato Diseases Classification Based on Deep Convolutional Neural Networks Md. Ferdouse Ahmed Foysal, Mohammad Shakirul Islam, Sheikh Abujar and Syed Akhter Hossain

1 Introduction Tomato (Solanum lycopersicum) growing is highly profitable in Bangladesh. According to officials at the Department of Agriculture Extension, the cultivation has been increased by five to six times during the last 15–20 years. Tomato farming is gaining popularity everywhere in the region as it is now considered as the second cash crop and plays an important role in the economy, financially benefiting at least 0.2 million families there. Diseases are often considered as one of the major limiting factors in the cultivation of tomato. Tomato crops are highly affected by few diseases, which causes dramatic losses in agriculture economy [1]. There are a plethora of diseases which attacks tomato leaves, and a variety of symptoms such as spots or smudge arise on the tomato leaves. There are few kinds of tomato disease arising in tomato plants, and those diseases are known as early and late blight, septoria spot in leaf, leaf mold, bacterial spot, pith necrosis, buckeye rot, anthracnose, and southern blight. In this paper basically, we work with leaf dataset. One of the serious diseases of potato and tomato is blight, and there are two kinds, namely blight late and early blight which is caused by one kind of fungus. In wear weather, late blight basically attack the plants. All the parts of the plant can Md. Ferdouse Ahmed Foysal (B) · M. Shakirul Islam · S. Abujar · S. Akhter Hossain Department of Computer Science and Engineering, Daffodil International University, Dhaka, Bangladesh e-mail: [email protected] M. Shakirul Islam e-mail: [email protected] S. Abujar e-mail: [email protected] S. Akhter Hossain e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_49

583

584

Md. Ferdouse Ahmed Foysal et al.

be affected by the fungus. Lesions are so small which comes in darkness in young leaf, and water-soaked the spots. Generated leaf spots can enlarge so fast and a white mold came at the end point of the affected portion on the lower parts of folia [2]. Bacterial spot is caused by a bacterium named Xanthomonas vesicatoria, and green tomatoes are attacked by these bacteria not red. This disease is also seen in wet seasons [3]. There is another disease called Septoria Leaf Spot, which is caused by another fungus named Septoria lycopersici. The lower leaves near the land is usually affected after the plants begin to grow as young fruits [3]. Spider mites are tiny, but can be seen with a hand lens or good eyes. They never have wings. They usually attack the undersides of leaves. Tomato leaf curl virus is transmitted by whiteflies. Syndromes in plants are the high-placed curling of leaves, small leaves, yellow edge, flower drop, and stunting plants. There will be no fruit formed, if the tomato plants are infected early [3]. The main objective of our approach is to classify the disease. Our dataset has six classes, first five contains different diseases and the last one is healthy leaf images.

2 Literature Review In this paper, our approach is tomato diseases classification using deep convolutional neural network. Our suggested CNN [4] model can provide a better result. Based on the exploration of formation and parameters in neural network, various algorithms can be applied to train a CNN, and gradient-descent is one of them. A total of 3000 images of the tomato leaves are processed first, and after the initial process, the processed data is sent to the final center of training the CNN [4]. In 2012, Inge M. Hanssen and Moshe Lapidot published their work paper on major tomato viruses in the Mediterranean basin. They described about various causes and effects of tomato diseases [5]. In 2013, a research group from Pakistan proposed an Automated Plant Disease Analysis (APDA). They compared the performance of several ML techniques to identify plant disease swatch from the images [6]. The previous detection of diseases in tomato crops using the approach of intelligent systems and electronic nose was proposed back in 2010 [7]. They used statistical and intelligent systems techniques, which were employed to process the data. For visualizing clusters within the dataset methods of k-means clustering, PCA and Fuzzy CM clustering were applied. ANNs were used to classify and hence categorizes the datasets. J. Amara et al. proposed a method of renowned CNN-based study for banana disease recognition. They use LeNet architecture as a CNN to classify images [8]. In [9], Y. Lu and many more raised unification of diseases in rice using neural nets. Their neural nets were learned to unify 10 most important rice afflictions. Mohammad Brahimi and few others formed another deep learning-based method for tomato diseases classification and symptoms visualization [10]. Their dataset was pretty good, and they also used visualization methods to understand symptoms. Many tomato diseases are described in detail with syndromes in an article [3]. In [11–16], different approaches of different leaf-based plant diseases classification and recognition works has been shown.

49 A Novel Approach for Tomato Diseases Classification …

585

3 Proposed Methodology The process of tomato diseases classification is discussed in this section. In our method, we formed a deep Convolutional Neural Networks. The construction has few sequences such as data collection, data preparation, proposed model description, and training procedure of the model. The last section describes the performance and result of our work.

3.1 Convolutional Neural Networks Convolutional Neural Network (CNN) is one kind of deep artificial neural network. The well-known computer scientist Yann LeCun made convolutional neural network inspired by biological processes. There are several types of layers in a Convnet. Types of layers: Let us take an image which dimension is 100 × 100 × 3 pass through a covnet. Input Layer: It contains raw input of image, which dimension is 100 × 100 × 3. Convolution Layer: In convolution layer, it calculates the dot product between all filters and image patch. If the input image dimension is 100 × 100 × 3 and if filter number that is used is 12, then the output dimension is 100 × 100 × 12. Activation Function Layer: It is also called Transfer Function Layer. There are two types of activation function. One is Linear Activation Function and Nonlinear Activation Function. The output dimension of this layer is 100 × 100 × 12. Pool Layer: Pool layer generally sits after activation function layer. There are several types of pool layer. Max pooling is one of the most used pool layers. If the max pooling layer is used with pool size 2 × 2 and stride 2, then the output dimension is 50 × 50 × 12 (Fig. 1). Fully Connected Layer: Fully connected layer takes input from the immediate previous layer. It converts the array into 1D array that is equal to the size of number of classes.

Fig. 1 Max pool layer

586

Md. Ferdouse Ahmed Foysal et al.

3.2 Dataset Collection We have collected image of tomato leaves that are infected from five different diseases and also collected the image of healthy tomato leaves. The following list is the classes of our dataset: a. b. c. d. e.

Bacterial spot. Late bright. Septoria leaf spot. Spider mites. Tomato leaf curl virus.

We made a dataset from the collected images, which contains 3000 images. There are 6 classes. We have used 2400 images for training and 600 images for testing (Fig. 2).

3.3 Dataset Preperation Machine learning is all about to train the model based on current data to predict the future values. So, we need the proper amounts to train our model. So in real life, we do not always have the correct data to work with. If the data is not processed correctly, then we need to prepare it, and then start training our model. For our proposed model, we resized our images into 100 × 100 pixels and converted them to gray scale from RGB.

Fig. 2 Example of dataset

49 A Novel Approach for Tomato Diseases Classification …

587

Fig. 3 Architecture of our model

3.4 Proposed Model We proposed a convolutional neural network, which is developed by us was used to classify the diseases of tomato. It is a 15-layer network. The layers are given as below (Fig. 3): Convolution layer: In our model, there are five convolution layers, which is described below. a. b. c. d. e.

Convolution 32 – 3 × 3 filter. Convolution 64 – 3 × 3 filter. Convolution 128 – 3 × 3 filter. Convolution 256 – 3 × 3 filter. Convolution 512 – 3 × 3 filter.

Pooling Layer: Most commonly pooling layer sits next to convolution layer. There are five max pooling layers. Each layer’s pool size is 2 × 2. Dropout Layer: Generally, “dropout” means to dropping out units in a neural network. Dropout layer is mainly used to avoid overfit on neural networks. Our model has two dropout layers. One layer rate is 0.4 and the other is 0.5. Flatten Layer: Flattening is the process of converting all the resultant twodimensional arrays into a single long continuous linear vector. The model has a flatten layer. Dense Layer: The last stage of a Convolutional Neural Network (CNN) is a classifier. It is called a dense layer, which is just an Artificial Neural Network (ANN) classifier. Our model has two dense layer. The first dense layer has 512 units and the last dense layer has 6 units, which is also called output layer.

3.5 Training the Model Our model was compiled by one of the most renowned algorithms called Adam optimizer. Adam optimizer is one of the most used optimizers. It is faster and more

588

Md. Ferdouse Ahmed Foysal et al.

reliable and for that reason it is used mostly. The loss type we have used is known as “Categorical cross entropy”. It is mainly needed for multi-class classification. Keras fit( ) function was used to train our model. Our model was trained for 40 epochs and batch size was 64. 80% (2400 images) of the dataset was used for training. From 2400 images, 80% was used for training and the remaining 20% was used for validation.

4 Performance Evaluation After training and validation, we obtain our results. Training accuracy is usually the accuracy when the model is applied on the training data. When the model is applied on test data from different classes, it is known as validation accuracy. Figure 4 shows a graph which contains training and validation accuracy of our model. Figure 5 shows a graph which contains training and validation loss of our model.

Fig. 4 Training and validation accuracy

Fig. 5 Training and validation loss

49 A Novel Approach for Tomato Diseases Classification …

589

5 Result Discussion After all the training processes finally our model is tested by 600 data, the confusion matrix was generated. In the field of machine learning and specifically the classification problem, confusion matrix used to visualize the performance of an algorithm. In Fig. 6 graph shows the normalized confusion matrix. We also calculated Precision, Recall, and F1- score from testing the result. Average Precision, Recall, and F1-score is 0.76. And our accuracy is 76% which is a decent accuracy (Table 1). Next figure shows the Confusion matrix without normalization (Fig. 7).

Fig. 6 Normalized confusion matrix

Table 1 Classification report Class Precision Bacterial spot Late bright Septoria leaf spot Spider mites Tomato leaf curl virus Healthy leaf Avg/total

0.78 0.67 0.71 0.80 0.74 0.86 0.76

Recall

F1-score

0.87 0.73 0.59 0.82 0.75 0.80 0.76

0.82 0.70 0.64 0.81 0.75 0.83 0.76

590

Md. Ferdouse Ahmed Foysal et al.

Fig. 7 Confusion matrix without normalization

6 Future Work In our proposed method, we can classify five kinds of tomato disease, and we have used convolution neural networks algorithm to build a model for our tomato data. Our future goal is to optimize accuracy, collect huge amount of data to make our own dataset on various plant diseases. Our country is an agricultural land, and here, more than half of the total population depends on farming. Using new technologies, we want to make a better solution for our future world, for betterment. Hope one day our country, our agriculture industry will be more beneficial for us. We will make automated plant virus defender and disease detector to take prior steps.

7 Conclusion A human being cannot live without food, and for better production of healthy food, we can apply new technologies like detection and prevention. Nowadays, we can achieve good accuracy using CNN models, and there are so many challenges to solve a real-life problem like this tomato disease. There are many pattern-recognition methods in the problem-solving world, but CNN can be applied both in theory and application. An innovative method is applied in this classification or recognition process to enlarge the ability of deep learning. Detecting the five most common tomato diseases from leaves was our main goal. It was challenging and difficult to work with images of different resolution, size, pose, and orientation. Our system was able to find good results after finalizing the raw data, completing the training, and

49 A Novel Approach for Tomato Diseases Classification …

591

validation problem. We can say from above all experiments and experiences that our applied method is able to detect affected leaves accurately. We can easily understand that a computational effort can give us a better solution.

References 1. Hanssen IM, Lapidot M (2012) Major tomato viruses in the Mediterranean basin. In: Loebenstein G, Lecoq H (eds) Advances in virus research, vol 84. Academic Press, San Diego, pp 31–66 2. Blancard D (2012) Tomato diseases. Academic Press, The Netherlands; Breitenreiter A, Poppinga H, Berlin TU, Technik FN (2015) Deep learning. Nature 521:2015 3. Blake JH, Keinath AP, Kluepfel M, Williamson J (2018) Tomato diseases and disorders by Clemson University Home and Garden Information Center. https://hgic.clemson.edu/factsheet/ tomato-diseases-disorders 4. Alex K, Sutskever I, Hinton GE (2012) Imagenet classification with deep convolutional neural networks. In: Pereira F, Burges CJC, Bottou L, Weinberger KQ (eds) Neural information processing systems (NIPS). Curran Associates Inc., Lake Tahoe, pp 1097–1105 5. Hanssen IM, Lapidot M (2012) Major tomato viruses in the Mediterranean basin. Adv Virus Res 84:31–66 6. Akhtar A, Khanum A, Khan SA, Shaukat A (2013) Automated plant disease analysis (APDA): performance comparison of machine learning techniques. In: Proceedings of the 11th international conference on Frontiers of information technology, pp 60–65 7. Ghaffari R, Zhang F, Iliescu D, Hines EL, Leeson MS, Napier R, Clarkson PJ (2010) Early detection of diseases in tomato crops: an electronic nose and intelligent systems approach. In: International Joint Conference on Neural Networks (IJCNN) 8. Amara J, Bouaziz B, Algergawy A (2017) A deep learning-based approach for banana leaf diseases classification. In: Mitschang B et al (eds) BTW 2017-Workshopband. Lecture notes in informatics (LNI), Gesellschaft für Informatik, Bonn 9. Lu Y, Yi S, Zeng N, Liu Y, Zhang Y (2017) Identification of rice diseases using deep convolutional neural networks. Neurocomputing 267:378–384 10. Brahimi M, Boukhalfa K, Moussaoui A (2017) Deep learning for tomato diseases: classification and symptoms visualization. Appl Artif Intell 31(4):299–315. https://doi.org/10.1080/ 08839514.2017.1315516 11. Fuentes A, Yoon S, Kim SC, Park DS (2017) A robust deep-learning-based detector for realtime tomato plant diseases and pest recognition. Sensors 17:2022. https://doi.org/10.3390/ s17092022 12. Al-Hiary H, Bani-Ahmad S, Reyalat M, Braik M, ALRahamneh Z (2011) Article: fast and accurate detection and classification of plant diseases. Int J Comput Appl 17(1):31–38 13. Rumpf T, Mahlein AK, Steiner U, Oerke EC, Dehne HW, Plümer L (2010) Early detection and classification of plant diseases with support vector machines based on hyper spectral reluctance. Comput Electron Agric 74(1):91–99 14. Dandawate Y, Kokare R (2015) An automated approach for classification of plant diseases towards development of futuristic decision support system in Indian perspective. In: Proceedings of the international conference on advances in computing, communications and informatics (ICACCI), Kochi, India. IEEE, pp 794–799 15. Koike ST, Gladders P, Paulus AO (2007) Vegetable diseases: a color handbook. Academic Press, San Diego 16. Le T-L, Duong N-D, Nguyen VT, Vu H (2015) Complex background leaf-based plant identification method based on interactive segmentation and kernel descriptor. In: Proceedings of the 2nd international workshop on environmental multimedia. Conjunction with ACM conference on multimedia retrieval (ICMR), Shanghai, China. ACM, pp 3–8

Chapter 50

Classification by Clustering (CbC): An Approach of Classifying Big Data Based on Similarities Sakib Shahriar Khan, Shakim Ahamed, Miftahul Jannat, Swakkhar Shatabda and Dewan Md. Farid

1 Introduction In the current digital era, due to the rise of digital technology, the volume of digital data is rapidly increasing day by day. It has opened a new research area in the field of data science and attracts computational intelligence researchers to work on big data [6]. Big data generally refer to 3Vs: volume, variety, and velocity [14]. Volume is the quantity of the data, variety is the different types of attributes in the data, and velocity is the continuous incoming data in and out. There are few other Vs that can be added for big data analytics such as vision, verification, and validation [5]. Vision is the planning/purpose of the data that would be achieved. The verification ensures that the data conforms to a set of specifications and validation checks whether the big data generation is fulfilled or not. In case of big data, there could be several different sources (e.g., customer analysis, medical research, government research, security and law enforcement, and financial trading, etc.) [15]. Mining big data employing data mining algorithms is a challenging and difficult task. Knowledge extraction from big data is the process of discovering patterns or unknown structures in big data S. S. Khan · S. Ahamed · M. Jannat · S. Shatabda · D. Md. Farid (B) Department of Computer Science & Engineering, United International University, United City, Madani Avenue, Badda, Dhaka 1212, Bangladesh e-mail: [email protected] URL: http://cse.uiu.ac.bd S. S. Khan e-mail: [email protected] S. Ahamed e-mail: [email protected] M. Jannat e-mail: [email protected] S. Shatabda e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_50

593

594

S. S. Khan et al.

[20]. Machine learning for mining big data can be bifurcated into descriptive and predictive tasks [9, 10]. Descriptive mining is the task of analysing, summarising, and finding historical patterns in big data. On the contrary, predictive mining is the process of decision-making or prediction of unseen data by building a learning model from the big data [12]. In general, predictive mining is categorised into supervised and unsupervised learning. Supervised learning focuses on classification and regression techniques. Whereas, unsupervised learning looks for structures in the data. In classification/regression, a predictive model is created from the labelled training data having the output variables or target predictions [8, 19]. Labelling training data is a costly and time-consuming process. In many realworld applications, human expert manually label the training instances like Bioinformatics, genomic data processing, intrusion classification, etc. [6]. Big data contain thousands/millions of the instances in the train set data and labelling these large numbers of training instances is really very difficult task. In this paper, we have proposed an approach of classifying unlabelled big data employing cluster analysis. The main objective of this study is the classification of big data by clustering (CbC). The idea is to build a classifier based on the similarities of the instances instead of class labels of the instances for classifying big data. The proposed method clusters the big data into several groups/clusters and builds a classification model based on the clusters. Clustering is a process of grouping a set of instances into a set of subclasses that are called as clusters, and it is also known as learning by observation, automatic classification, data segmentation, unsupervised learning, etc. Similarities and dissimilarities of instances can be determined by the feature values. Clustering groups the big data into several clusters so that instances in the same cluster have high similarity to each other and instances in different clusters are high dissimilar [7]. To design and build classifiers, we have used several decision tree learning algorithms: ID3, C4.5, and CART, as decision tree induction is one of the most popular machine learning algorithms for data mining task [11]. We have collected 10 big datasets from the UC Irvine machine learning repository for experimental analysis and validated the proposed method. The experimental results show that the proposed method scale up the classification accuracy for most of the datasets. The rest of the paper is arranged as follows. Section 2 presents the decision tree induction algorithms. Section 3 presents the clustering methods. Section 4 introduces the proposed big data classification method. Section 5 provides the experimental results and datasets details that used in experiments. Conclusion with future work is presented in Sect. 6.

2 Decision Tree Induction Decision tree (DT) is one of the most popular machine learning algorithms for data mining task [11]. DT can be used for solving many real-life classification and regression problems in supervised learning. It presents a set of rules; each path from root to leaf node represents a classification rule. DT is basically a top-down recursive divide and conquer approach. DT starts with a root node chosen by the highest entropy of

50 Classification by Clustering (CbC): An Approach of Classifying Big Data …

595

feature and follows the tree down the branches using gain, until a leaf node represents a single class. Information gain, which is also known as mutual information that is a ratio of information gain to reduce the bias towards multi-valued features by taking the number and size of branches into account when choosing a node of the tree [18]. The following Eqs. 1–3 exhibit the gain calculation: Entr opy(X ) = −

n 

pi log2 ( pi )

(1)

i=1 n  |X j |

× Entr opy(X j )

(2)

Gain(A) = Entr opy(X ) − Entr opy A (X )

(3)

Entr opy A (X ) =

j=1

|X |

The Gain Ratio is applied in C4.5 algorithm [17], which is an extension of the entropy-based method that is the successor of ID3 (Iterative Dichotomiser 3) algorithm [18]. Gain Ratio uses “split information” a kind of normalisation of entropy estimation. The feature with maximum Gain Ratio is selected for further splitting in the tree. Equations 4 and 5 present the Gain Ratio calculation. Split I n f o A (X ) = −

n  |X j | j=1

Gain Ratio(A) =

|X |

 × log2

|X j | |X |



Gain(A) Split I n f o(A)

(4)

(5)

Another DT induction algorithm is called Classification and Regression Tree (CART) [2], which uses Gini Index in Eqs. 6 and 7. The attribute, A j that maximises the reduction in impurity is selected as the splitting feature. Gini(X ) = 1 −

n 

p 2j

(6)

j=1

The splitting of X into subsets X 1 and X 2 is dependent on the following equation: Gini A (X ) = 1 −

| X 1 | ×Gini(X 1 ) | X 2 | ×Gini(X 2 ) + |X| |X|

(7)

3 Clustering Analysis Clustering is the process of grouping or partitioning unlabelled data/objects [1]. Unknown hidden patterns/structure can be found in the data by applying clustering analysis. Clustering algorithms try to group the data into many clusters so that

596

S. S. Khan et al.

instances/data points in the same cluster have very high similarity to each other and instances in different clusters are highly dissimilar. Another principle of cluster analysis is that every cluster should have minimum one instance and one instance should belong to only one cluster [3]. Let X be the set of unlabelled instances. X = {x1 , x2 , x3 , . . . , xn }

(8)

After partitioning X into k clusters: C p = ∅, p = 1, . . . , k

(9)

X = C1 ∪ C2 ∪ C3 ∪ · · · ∪ Ck

(10)

C p ∩ Cq = ∅, p = q, p, q = 1, . . . , k

(11)

Similarities and dissimilarities of instances can be determined by the feature values in the dataset. Clustering refers to the automatic classification, which is also known as data segmentation, unsupervised learning, learning by observation, etc. Clustering methods are divided into four categories: (1) partitioning method, (2) hierarchical method, (3) density-based method, and (4) grid-based method [7, 12]. Partitioningbased clustering method clusters N instances into K number of clusters of convex shape (N ≥ K ). It uses traditional distance measures (e.g., Euclidean distances, Manhattan distance) to cluster the data points. It uses mean or medoid to find cluster center and an iterative relocating approach to improve the cluster validation (e.g., k-means clustering) [13].

4 Method for Classifying Big Data Classifying big data is a challenging and difficult task, as there are thousands/millions of instances in the data and we cannot analysis and store the full data at a time. The most common problem to deal with big data is to store the total data into the computer memory as big data are so big in volume. Therefore, parallel and distributed computing becomes very useful for handling big data. In general, big data is divided into several small subsets that we can deal with and each of which fits into the computer memory. Most real-world datasets are unlabelled and labelling data is very costly and time consuming. Most of the time human experts label the unlabelled the data instances. In this paper, we have proposed an approach for classifying big data employing Classification by Clustering (CbC) method that does not consider the class labels of data instances for building classifiers. Initially, the big data is divided into several small sub-datasets. Then, clustering technique is applied to each sub-dataset to cluster the data. The proposed method used similarity-based clustering algorithm, which is a robust method for clustering instances based on the similarities. The similarity-based clustering method automatically generates the cluster numbers and form different volumes of clusters. After that, several decision trees are built

50 Classification by Clustering (CbC): An Approach of Classifying Big Data …

597

from sub-datasets. Finally, all the decision trees are examined and analysed to form a single decision tree that turns out to be very close to the tree that would have been generated if the big data had fit in memory. To select an informative attribute/feature for constructing a decision tree, we can use any attribute selection techniques, e.g., ID3, C4.5, and CART. Algorithm 1, outlines the proposed algorithm for classifying unlabelled big data using similarity-based clustering with decision tree. Algorithm 1 Classifying Big Data by CbC Input: Big Data, D Big ; Output: A classification model, M; Method: 1: divide D Big into several sub-datasets {D1 , D2 , . . . , Dn }; 2: for each Di = {x1 , x2 , . . . , x N } do 3: // find clusters, C = {C1 , C2 , . . . , Ck }; 4: C = ∅, k = 1, & Ck = {x1 }; 5: C = C ∪ Ck ; 6: for i = 2 to N do 7: for l = 1 to k do 8: find maximum similarity, sim(xi , xl ) with lth cluster center xl ∈ Cl ; 9: end for 10: if sim(xi , xl ) ≥ threshold then 11: Cl = Cl ∪ x i ; 12: else 13: k = k + 1, & Ck = {xi }; 14: C = C ∪ Ck ; 15: end if 16: end for 17: // create a decision tree: DTBuild(Di ); 18: DTi = find the root node with best splitting attribute, A j ∈ Di , & add arc to root node for each split predicate and label with cluster, Ck ; 19: for each arc do 20: create D j by applying splitting predicate of Di ; 21: if stopping point reached for this path then 22: DT  = create leaf node by label Ck ; 23: else 24: DT  = DTBuild(D j ); 25: end if 26: DTi = add DT  to arc; 27: end for 28: end for 29: construct a new tree, DTN ew using decision trees, DTs = {DT1 , DT2 , . . . , DTn }; 30: M = DTN ew ;

The proposed algorithm is an adaptive and scalable approach. It can take new instances and update the decision tree to reflect the changes in data without constructing a new decision tree from the scratch. The main idea of this paper is to build a scalable and adaptive classification model for mining big data based on the similarities of the instances instead of class labels of instances. We have divided the big data, D Big into several small sub-datasets, {D1 , D2 , . . . , Dn } and applied clus-

598

S. S. Khan et al.

tering technique to cluster each of sub-data, Di based on similarity of instances. The similarities are determined based on the features of the data, so that similar instances are grouped together. Then, we have built a decision tree from each sub-dataset using cluster analysis without considering class labels of the data. The DTBuild() is a recursive function, which builds the decision tree from a given training data. It can be used, information gain, gain ratio, or gini index for selecting the best/root attribute/feature. Resulting we have n number of decision trees, DTs . To form a single classification model we merge all the decision trees and build a new decision tree, DTN ew . In this study, we have considered decision tree learning algorithms for classifier construction, because decision tree algorithms are very well known and commonly using many real-world data mining applications. It is easy to build and easy to understand.

5 Experiment Analysis This section presents the datasets, experimental setup, and results.

5.1 Dataset Descriptions We have selected the following 10 real benchmark datasets from the UCI machine learning repository [4]. Table 1 presents the details of the datasets. 1. 2. 3. 4. 5. 6. 7. 8. 9. 10.

Covertype Data Set (Covertype). APS Failure at Scania Trucks Data Set (APS). Gas Sensor Array Drift Dataset Data Set (Gas Sensor). Character Font Images Data Set (Character Font). Semeion Handwritten Digit Data Set (Semeion Digit). Adult Data Set (Adult). KDD Cup 1999 Data Set (KDD Cup). ISOLET Data Set (ISOLET). UJIIndoorLoc Data Set (UJIIndoorLoc). Dota2 Games Results Data Set (Dota2 Games).

5.2 Experimental Setup We have used scikit-learn machine learning library in Python for experimental setup and design [16]. We have used Spyder 3.2.6 for Python coding. Spyder is a part of Anaconda distribution, which is cross-platform and includes most of the Python

50 Classification by Clustering (CbC): An Approach of Classifying Big Data … Table 1 Datasets description No. Datasets No. of features Types of features 1 2 3 4 5 6 7 8 9 10

Covertype APS Failure Gas Sensor Character Font Semeion Digit Adult KDD Cup ISOLET UJIIndoorLoc Dota2 Games

54 171 128 411 256 41 42 617 529 117

Categorical, Integer Integer, Real Real Integer, Real Integer Categorical, Integer Categorical, Integer Real Integer, Real Integer, Real

599

Instances

Classes

581012 60000 13910 745000 1593 48842 4000000 7737 21048 102944

7 2 6 153 10 2 23 26 5 2

packages. Scikit-learn library includes a range of machine learning algorithms that contains preprocessing, feature selection, dimensionality reduction, and model selection tools. In our experiments, ID3 and CART algorithms are directly used from scikit-learn library. We have implemented C4.5 in python. We have used the classification accuracy, precision, recall, F-score, and tenfold CV (cross-validation) to evaluate our proposed method that are shown in Eqs. 12–15 where TP, TN, FP, and FN are true positive, true negative, false positive, and false negative, respectively. |X | accuracy =

i=1

assess(xi ) , xi ∈ X |X |

pr ecision = r ecall = F − scor e =

TP T P + FP

TP T P + FN

2 × pr ecision × r ecall pr ecision + r ecall

(12)

(13)

(14)

(15)

5.3 Result At first, we have evaluated the performance of the existing decision tree algorithms such as ID3, C4.5, and CART on benchmark datasets from Table 1. We have measured the accuracy, precision, recall, and F-score with tenfold cross-validation to test the performance of these classifiers. We have considered a weighted average of precision,

600

S. S. Khan et al.

Table 2 Classification results of ID3 algorithm using tenfold CV on benchmark datasets No.

Datasets

Accuracy (%)

Precision (weighted average)

Recall (weighted average)

F-score (weighted average)

1

Covertype

93.13

0.927

0.927

0.928

2

APS Failure

98.66

0.986

0.986

0.986

3

Gas Sensor

98.58

0.984

0.984

0.984

4

Character Font

99.99

0.999

0.999

0.999

5

Semeion Digit

74.81

0.765

0.748

0.748

6

Adult

81.40

0.815

0.814

0.814

7

KDD Cup

99.90

0.998

0.998

0.998

8

ISOLET

83.35

0.839

0.834

0.834

9

UJIIndoorLoc

98.18

0.982

0.982

0.982

10

Dota2 Games

52.40

0.524

0.524

0.524

Table 3 Classification results of C4.5 algorithm using tenfold CV on benchmark datasets No.

Datasets

Accuracy (%)

Precision (weighted average)

Recall (weighted average)

F-score (weighted average)

1

Covertype

91.70

0.915

0.915

0.915

2

APS Failure

98.62

0.986

0.985

0.986

3

Gas Sensor

98.63

0.986

0.986

0.986

4

Character Font

99.70

0.997

0.997

0.997

5

Semeion Digit

74.25

0.753

0.747

0.748

6

Adult

81.53

0.816

0.812

0.813

7

KDD Cup

99.95

0.999

0.999

0.999

8

ISOLET

83.13

0.838

0.831

0.832

9

UJIIndoorLoc

98.33

0.983

0.983

0.983

10

Dota2 Games

52.34

0.523

0.523

0.523

recall, and F-score. The results are summarised in Tables 2, 3, and 4 for ID3, C4.5, and CART respectively. The ID3 and C4.5 algorithms work well with more than 98% accuracy on APS Failure, Gas Sensor, Character Font, KDD Cup, and UJIIndoorLoc datasets. Contrarily, CART performs better on APS Failure, Gas Sensor, Character Font, and KDD Cup datasets. As the decision tree is a strong classifier, the overall performance of ID3, C4.5, and CART algorithms are good on benchmark datasets except for the Dota2 Games dataset. Then, we have tested the performance of the proposed Classification by Clustering (CbC) approach with ID3, C4.5, and CART on benchmark datasets from the Table 1. The results of CbC are tabulated in Tables 5, 6, and 7, respectively. The proposed approach improves the classification results

50 Classification by Clustering (CbC): An Approach of Classifying Big Data …

601

Table 4 Classification results of CART algorithm using tenfold CV on benchmark datasets No.

Datasets

Accuracy (%)

Precision (weighted average)

Recall (weighted average)

F-score (weighted average)

1 2

Covertype

92.30

APS Failure

98.41

0.925

0.922

0.925

0.983

0.984

3

Gas Sensor

0.983

98.54

0.985

0.985

4

0.925

Character Font

99.90

0.998

0.998

0.998

5

Semeion Digit

74.82

0.759

0.748

0.747

6

Adult

74.82

0.812

0.811

0.812

7

KDD Cup

99.93

0.999

0.999

0.999

8

ISOLET

82.19

0.828

0.822

0.822

9

UJIIndoorLoc

97.84

0.979

0.978

0.978

10

Dota2 Games

52.10

0.521

0.521

0.521

Table 5 Classification results of CbC with ID3 algorithm using tenfold CV on benchmark datasets No.

Datasets

Accuracy (%)

Precision (weighted average)

Recall (weighted average)

F-score (weighted average)

1 2

Covertype

99.53

APS Failure

98.65

0.994

0.993

0.994

0.986

0.986

3

Gas Sensor

0.986

97.90

0.978

0.978

4

0.978

Character Font

81.79

0.818

0.887

0.887

5

Semeion Digit

77.58

0.785

0.775

0.775

6

Adult

99.99

0.999

0.998

0.998

7

KDD Cup

99.92

0.999

0.999

0.998

8

ISOLET

84.65

0.852

0.847

0.846

9

UJIIndoorLoc

99.99

0.999

0.999

0.999

10

Dota2 Games

98.55

0.984

0.985

0.985

Table 6 Classification results of CbC with C4.5 algorithm using tenfold CV on benchmark datasets No.

Datasets

Accuracy (%)

Precision (weighted average)

Recall (weighted average)

F-score (weighted average)

1 2

Covertype

99.48

APS Failure

98.65

0.994

0.994

0.994

0.986

0.986

3

Gas Sensor

0.986

97.70

0.976

0.976

4

0.976

Character Font

80.75

0.807

0.807

0.807

5

Semeion Digit

78.13

0.799

0.788

0.789

6

Adult

99.40

0.994

0.998

0.998

7

KDD Cup

99.99

0.999

0.999

0.999

8

ISOLET

84.75

0.853

0.857

0.857

9

UJIIndoorLoc

99.99

0.999

0.999

0.999

10

Dota2 Games

98.86

0.988

0.988

0.988

of Dota2 Games dataset from 52% to greater than 98%, Adult dataset from 81 to 99% and Semeion Digit dataset from 74 to 78%. It classifies the Covertype, KDD Cup, and UJIIndoorLoc datasets with 99% accuracy and APS Failure dataset with

602

S. S. Khan et al.

Table 7 Classification results of CbC with CART algorithm using tenfold CV on benchmark datasets No.

Datasets

Accuracy (%)

Precision (weighted average)

Recall (weighted average)

F-score (weighted average)

1

Covertype

99.50

0.995

0.994

0.994

2

APS Failure

99.90

0.999

0.999

0.999

3

Gas Sensor

98.62

0.986

0.986

0.986

4

Character Font

80.25

0.803

0.802

0.802

5

Semeion Digit

78.65

0.798

0.786

0.786

6

Adult

99.99

0.999

0.999

0.999

7

KDD Cup

99.99

0.999

0.999

0.999

8

ISOLET

83.04

0.835

0.830

0.832

9

UJIIndoorLoc

99.99

0.999

0.999

0.999

10

Dota2 Games

99.52

0.996

0.995

0.995

1 0.95

ClassificaƟon Accuracy

0.9 0.85 0.8 0.75 0.7 0.65 0.6 0.55 0.5

ID3

CbC + ID3

Fig. 1 Accuracy of ID3 and CbC with ID3 algorithms on benchmark datasets

98% accuracy. Figures 1, 2 and 3 show the comparison of classification accuracies of decision tree algorithms (ID3, C4.5, and CART) with proposed CbC approach. Figure 4 shows the comparison of accuracies of classifiers.

50 Classification by Clustering (CbC): An Approach of Classifying Big Data … 1 0.95

ClassificaƟon Accuracy

0.9 0.85 0.8 0.75 0.7 0.65 0.6 0.55 0.5

C4.5

CbC + C4.5

Fig. 2 Accuracy of C4.5 and CbC with C4.5 algorithms on benchmark datasets 1 0.95

ClassificaƟon Accuracy

0.9 0.85 0.8 0.75 0.7 0.65 0.6 0.55 0.5

CART

CbC + CART

Fig. 3 Accuracy of CART and CbC with CART algorithms on benchmark datasets

603

604

S. S. Khan et al. 1 0.95

ClassificaƟon Accuracy

0.9 0.85 0.8 0.75 0.7 0.65 0.6 0.55 0.5

ID3

C4.5

CART

CbC + ID3

CbC + C4.5

CbC + CART

Fig. 4 Classification accuracy of the different classifiers

6 Conclusion In this paper, we have introduced a method for classifying unlabelled big data based on classification by cluster (CbC) technique. We have used similarity-based clustering and decision tree algorithms in the proposed method. Both decision tree and similarity-based clustering methods are very popular data mining techniques that are used in many real-world data mining applications. The objective of this paper is to build a classification model based on the similarity of instances instead of the class labels of the instances. The idea is to avoid contradictory and redundant class labels in the training data while building a classifier. As noisy class labels are very common in real-life datasets. The proposed method builds a classifier based on the data, not on the class labels. The experimental results proved that the proposed method of similarities of instances for classification improves the classification results. We have taken 10 benchmark datasets from the UCI machine learning repository with large number of features and instances for experimental analysis that we can consider the datasets as big data. We have used scikit-learn machine learning library in Python for experimental setup, which is a simple and efficient tool for data mining and data analysis. The proposed algorithm overall improves the classification results in comparing with traditional decision tree algorithms (ID3, C4.5, and CART) for all datasets except Character Font dataset. Specially, it improves the classification accuracy of the Dota2 Games dataset from 52 to 98% and Adult dataset from 81 to

50 Classification by Clustering (CbC): An Approach of Classifying Big Data …

605

99%. In the future, we will apply adaptive ensemble classifiers with the proposed method for classifying big data under dynamic feature set and also for novel class classification in concept drifting.

References 1. Aggarwal CC, Reddy CK (eds) (2013) Data clustering: algorithms and applications. Chapman and Hall/CRC data mining and knowledge discovery series. Chapman and Hall/CRC, Boca Raton 2. Breiman L, Friedman J, Stone CJ, Olshen RA (1984) Classification and regression trees. Chapman and Hall/CRC, Boca Raton 3. Chen X, Ye Y, Xu X, Huang JZ (2012) A feature group weighting method for subspace clustering of high-dimensional data. Pattern Recognit 45(1):434–446 4. Dheeru D, Taniskidou EK (2017) UCI machine learning repository. http://archive.ics.uci.edu/ ml 5. Fan W, Bifet A (2013) Mining big data: current status, and forecast to the future. ACM SIGKDD Explor Newsl 14(2):1–5 6. Farid DM, Al-Mamun MA, Manderick B, Nowe A (2016) An adaptive rule-based classifier for mining big biological data. Exp Syst Appl 64:305–316 7. Farid DM, Nowé A, Manderick B (2016) A feature grouping method for ensemble clustering of high-dimensional genomic big data. In: Future technologies conference, San Francisco, United States, pp 260–268 8. Farid DM, Rahman CM (2013) Assigning weights to training instances increases classification accuracy. Int J Data Min Knowl Manag Process 3(1):129–135 9. Farid DM, Rahman CM (2013) Mining complex data streams: discretization, attribute selection and classification. J Adv Inf Technol 4(3):129–135 10. Farid DM, Zhang L, Hossain A, Rahman CM, Strachan R, Sexton G, Dahal K (2013) An adaptive ensemble classifier for mining concept drifting data streams. Exp Syst Appl 40(15):5895– 5906 11. Farid DM, Zhang L, Rahman CM, Hossain M, Strachan R (2014) Hybrid decision tree and naïve bayes classifiers for multi-class classification tasks. Exp Syst Appl 41(4):1937–1946 12. Han J, Kamber M, Pei J (2011) Data mining concepts and techniques, 3rd edn. Morgan Kaufmann, Waltham 13. Jain AK (2010) Data clustering: 50 years beyond k-means. Pattern Recognit Lett 31(8):651–666 14. L’heureux A, Grolinger K, Elyamany HF, Capretz MAM (2017) Machine learning with big data: challenges and approaches. IEEE Access 5:7776–7797 15. Özköse H, Arı ES, Gencer C (2015) Yesterday, today and tomorrow of big data. Procedia-Soc Behav Sci 195:1042–1050 16. Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, Blondel M, Prettenhofer P, Weiss R, Dubourg V, Vanderplas J, Passos A, Cournapeau D, Brucher M, Perrot M (2011) Édouard Duchesnay: Scikit-learn: Machine learning in python. J Mach Learn Res 12:2825– 2830. http://dl.acm.org/citation.cfm?id=1953048.2078195 17. Quinlan J (1993) C4.5: programs for machine learning. Morgan Kaufmann, San Mateo 18. Quinlan JR (1986) Induction of decision tree. Mach Learn 1(1):81–106 19. Witten IH, Frank E, Hall MA (2011) Data mining: practical machine learning tools and techniques, 3rd edn. Morgan Kaufmann, Amsterdam 20. Wu X, Zhu X, Wu GQ, Ding W (2014) Data mining with big data. IEEE Trans Knowl Data Eng 26(1):97–107

Chapter 51

Brain–Machine Interface for Developing Virtual-Ball Movement Controlling Game Md. Ochiuddin Miah, Al Maruf Hassan, Khondaker Abdullah Al Mamun and Dewan Md. Farid

1 Introduction Brain–Machine Interface (BMI) is a direct communication pathway between wired brain and machines. BMI system takes complex neurophysiological activity as input signal then it processes this signal to perform actions by an external device [7, 14]. It revolutionized the health, rehabilitation industry and the medical treatment of disabled people [2, 12]. There are three types of BMI technologies such as noninvasive, partially invasive, and invasive BMI [8]. Noninvasive BMI places a set of electrodes on the surface of the scalp and measured activities from a huge group of electrodes. Partially invasive BMI wires are placed inside the brain and above the gray matter of it. In invasive BMI, single or a few neurons are placed inside the gray matter of the brain for generating better input signal. Invasive BMI is sophisticated and sensitive. Noninvasive BMI cannot generate better-input signal as invasive BMI, but it is less sophisticated and easy to manipulate with [8, 10]. BMI applications are developed widely in the medical, education, self-regulation, gaming, and entertainment fields. A robotic agent like thought-controlled robot and a robotic hand can be controlled and commended by the thought of a human. A thought-controlled wheelchair can be Md. Ochiuddin Miah · A. M. Hassan · K. A. A. Mamun · D. Md. Farid (B) Department of Computer Science and Engineering, United International University, United City, Madani Avenue, Badda, Dhaka 1212, Bangladesh e-mail: [email protected] Md. Ochiuddin Miah e-mail: [email protected] URL: http://cse.uiu.ac.bd A. M. Hassan e-mail: [email protected] K. A. A. Mamun e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_51

607

608

Md. Ochiuddin Miah et al.

moved by the thought of a user. BMI technologies are also revolutionized the gaming industry and make gaming more fun and lively [11, 13]. In this paper, we have explored the BMI technologies and developed a virtual-ball movement controlling game to control the movements of a ball using brain signal of voluntary movements. Initially, we have recorded the brain signals and extracted features from these signals to engender the training and testing data for binaryclass and three-class classifications. Then, informative patterns are extracted from the training data. We have applied the following machine learning for data mining algorithms: OneR, naïve Bayesian (NB) classifier [6], decision tree (DT) classifier [4], Random Forest [5], and Bagging on the training data and tested the performances of the above classifiers using testing data. Random Forest and DT (C4.5) classifier achieved better accuracy for binary-class and three-class classifications, respectively. Random Forest achieved 93.16 and 62.84% accuracy for binary-class and threeclass classification. On the contrary, DT (C4.5) classifier achieved 90.89 and 65.66% accuracy for binary-class and three-class classification. Then, we have applied C4.5 decision tree induction for developing a virtual-ball movement controlling game to control the direction of movement of ball using brain signals of voluntary movements. The rest of the paper is organized as follows. Section 2 presents the related works. Section 3 presents the signal acquisitions technique and datasets information. Section 4 provides experimental results for testing performance of different classifiers. Section 5 introduces the game that is controlled by brain signals of voluntary movements. Finally, Sect. 6 concludes the findings and proposed directions for future work.

2 Related Work Pattnaik and Jay [10] were extracted features from the raw EEG signals and transform it into the informative data for left- and right-hand movements. The hand movement data were recorded at 500 Hz from 21 years young person. Then, they applied lowpass filtering technique to remove artifacts and compared with the original datasets that gave satisfactory results. The frequency ranges theta (0–8 Hz), alpha (8–16 Hz), beta (16–32 Hz) and delta were extracted from the EEG signal using Discrete Wavelet Transform. Shanechi [12] presented the decoding algorithms for BMI with particular focus on closed-loop control ideas. This paper focuses on the Bayesian filtering and closed-loop adaptation techniques and concluded that an invasive BMI system is noisy, stochastic, and often consist of spiking activity. Yang et. al. [15] proposed a module called multichannel spike sorting with Neural Signal Processor (NSP) that deals with feature extractor, spike detector, and spike classifier blocks. In this paper, the authors identified that a training procedure is required to acquire spike templates (for template matching) or classification model coefficients through a spike clustering step. They minimized the memory usage for high-channel-count applications by adopting the feature-space based spike sorting. They chose non-linear energy operator (NEO) algorithm for their spike detector

51 Brain–Machine Interface for Developing Virtual-Ball Movement Controlling Game

609

design. They compared the features of each spike to the mean feature of each cluster for spike classification by using a distance-based metric function. This experimental analysis showed that the Decision Tree (DT) based method can provide 50% memory reduction compared to the distance-based classification models. An automatic clustering employing k-means clustering with the gap statistic technique was used to complete the spike sorting system. In this paper, the hardware-efficient with featurespace based multichannel spike sorting with NSP module are targeted for the future high-channel-count of BMI applications. Davis et. al. [3] have applied the signal processing techniques to analyze brain dynamics for deeper understanding of Neural Correlates of Consciousness (NCC) in different cognitive states like Mental Arithmetic Exercises, Music- Induced Meditation, and Meditation. This research can measure multichannel EEG brain dynamics using MINDO48 equipment associated with three experimental modalities measured both in the laboratory and the natural environment. The spatiotemporal dynamics and transitions from and into different frequency bands are measured on 19 different areas of the scalp to find the differences in modalities that are smaller in the occipital and parietal regions of the scalp. Andreu-Perez et. al. [1] introduced a system to control the left and right side movements of the robot using EEG brain signals. Noninvasive BMI was used that electrodes were placed surface of the scalp using six electrodes for recording EEG motor imagery signals and two electrodes for ground and reference electrodes. This system used self-adaptive GT2 fuzzy classification algorithm for classifying EEG brain signals with 85.78% accuracy when compared to other 13 methods such as: oLDA, AdIM, McIT2FIS, SA-GT2FGG-NS, SA-GT2FGG-T1FS, T1FSIMclass, DBN, SCGNF, ARTMAP, T1FGG, RF, RBFSVM, LinSVM. Mamun et. al. [9] decoded human voluntary movements from subthalamic local field potentials (LFPs) using optimized signal processing and classification methods. Three different spectral measures named fast Fourier transform, continuous wavelet transform, and the statistical properties of wavelet spectra were used to evaluate the important features of the STN LFPs. Left or right movements were sequentially classified using a support vector machine (SVM) with optimized parameters and average correct classification of movement reached 91.5 ± 2.3%.

3 Experimental Paradigm and Signal Acquisitions 3.1 EEG Emotiv EPOC+ Neuroheadset The EEG neuroheadset Emotiv EPOC+ generates quantifiable electric potentials that can measure brain activities. It is equipped with a total of 14 electrodes which are metallic with a plastic base and placed on the scalp of the brain to measure brain activities. The electrodes are distributed according to the 10–20 international system as shown in Fig. 1. In 10–20 systems, the actual distance between two adjacent electrodes be can either 10 or 20% of the total left-right or front-back distance of the

610

Md. Ochiuddin Miah et al.

Fig. 1 Emotiv EPOC+ neuroheadset electrodes distribution according to the 10–20 international system. Source: emotiv systems

scalp [10]. The electrodes are located in AF3, F7, F3, FC5, T7, P7, O1, O2, P8, T8, FC6, F4, F8, and AF4 location and with two CMS (Common Mode Sense) and DRL (Driven Right Lag) reference electrodes are located in P3 and P4 locations [8]. The Motor Cortex is the primary contributor for generating neural impulses which are passed down via spinal cord and control the execution of movements [9]. We have recorded brain signal of voluntary movements so we take F3, FC5, FC6, and F4 locations which are located in motor cortex area.

3.2 Participants Six left- and right-handed healthy subjects (four males and two females) participated in this research. All participants had stated that they have no neurological disorders or movement-related illness. All participants gave informed consents, informed about the procedure of the experiment and trained before the experiment.

3.3 Data Recording We have developed a Java program using Emotiv community SDK to take brain signals from Emotiv Neuroheadset. The program can acquire the average band power

51 Brain–Machine Interface for Developing Virtual-Ball Movement Controlling Game

611

of theta, alpha, low beta, high beta and gamma frequency bands. The average band power is acquired for F3, FC5, FC6, and F4 location from the latest epoch with 0.5 s step size and 2 s window size. We have recorded data for binary-class and three-class classifications. Binary-classes are right hand and steady and three-classes are left hand, right hand, and steady. During the recording, subjects wore Emotiv Epoc+ EEG neuroheadset that was connected via wireless connection to the USB dongle on the program computer. From every subject, we have taken two trials for recording training data and one trial for testing data. Every trial for training data, we have recorded data for binary-class and three-class classifications and the duration was 30 s. We also take a trial for recording testing data for binary-class and three-class classifications and the duration was 15 s.

3.4 Data Description The frequency of different activities inside the brain provide informations about those activities. We can see significant activities on many frequency bands and those are mentioned below: Theta: Its band value is 4–8 Hz and activities are seen in arousal, drowsiness, and often during meditation. This frequency range is involved in sleeping, daydreaming, feeling deep, and raw emotions [12]. Alpha: Its band value is 8–12 Hz. It is the frequency range between beta and theta and it bridges the gap between our conscious thinking and subconscious mind. It involves the feelings of deep relaxation, associated with inhibition control and closing the eyes [3]. Beta: Its band value is 12–25 Hz and activities are associated with busy or anxious thinking, task-oriented, active concentration, conscious thought, and logical thinking. It is observed while we are awake and known as high-frequency low- amplitude brain waves. The Emotiv SDK provides access of low beta (12–18 Hz) and high beta (18– 25 Hz) subbands in beta zone for allowing the intensity and the type of processing to be better understood [9]. Gamma: Its band value is 25–45 Hz. It is the fastest brain waves and pass information rapidly and quietly. It is involved in high processing tasks and cognitive functioning. It is important for learning and information processing [11]. We have distributed dataset for binary-class and three-class from the acquired dataset. Binary-class and three-class datasets informations are mentioned in Table 1. We have visualized data and find attributes statistics using R programming mentioned in Table 2.

612

Md. Ochiuddin Miah et al.

Table 1 Hand movement datasets description No. Datasets Name of Types of features features 1.

2.

Three-class hand movement data Binary-class hand movement data

Theta, Real Alpha, Low beta, High beta, Gamma Theta, Real Alpha, Low beta, High beta, Gamma

Training data Testing data points points

Classes

473676

124492

Lefthand, Righthand, Steady

472060

52304

Righthand, Steady

Table 2 Hand movement dataset attributes statistics No. Attribute name Min value Max value 1. 2. 3. 4. 5.

Theta Alpha Low beta High beta Gamma

0.009 0.006 0.006 0.01 0.01

112.32 92.89 64.35 67.4 37.2

Mean value

STD value

6.83 4.99 5.52 3.31 1.63

13.14 10.21 11.34 8.31 3.83

4 Experiment and Result We have applied classification methods like: OneR, naïve Bayesian (NB) classifier, decision tree (DT) induction, Random Forest, and Bagging. We have trained the models supplying training data and test the models supplying testing data using R programming. We have used the classification accuracy, precision, recall, F-score to test the performance of the classifiers that are shown in Eqs. 1–4 where TP, TN, FP, and FN are true positive, true negative, false positive, and false negative, respectively. We have considered the weighted average for precision, recall, and F-score analysis. The performance of classifiers for binary-class and three-class are tabulated in Tables 3 and 4, respectively. |X | assess(xi ) , xi ∈ X (1) accuracy = i=1 |X | precision =

recall =

TP TP + FP

TP TP + FN

(2)

(3)

51 Brain–Machine Interface for Developing Virtual-Ball Movement Controlling Game

613

Table 3 The accuracy, precision, recall, and F-score of classifiers with test data on binary-class hand movement dataset No.

Name of classifier

Accuracy (%)

Precision (weighted average)

Recall (weighted average)

F-score (weighted average)

1.

OneR

73.69

0.76

0.73

0.73

2.

NB classifier

73.31

0.83

0.73

0.71

3.

C4.5

90.89

0.91

0.9

0.9

4.

CART

91.33

0.91

0.91

0.91

5.

Random forest

93.16

0.93

0.93

0.93

6.

Bagging

90.30

0.91

0.9

0.9

Table 4 The accuracy, precision, recall, and F-score of classifiers with test data on three-class hand movement dataset No.

Name of classifier

Accuracy (%)

Precision (weighted average)

Recall (weighted average)

F-score (weighted average)

1.

OneR

51.29

0.52

0.51

0.51

2.

NB classifier

59.94

0.73

0.59

0.58

3.

C4.5

65.66

0.66

0.65

0.66

4.

CART

52.56

0.53

0.52

0.53

5.

Random forest

62.84

0.63

0.62

0.62

6.

Bagging

60.59

0.61

0.60

0.60

F − score =

2 × precision × recall precision + recall

(4)

Accuracy of movement classification varied between participants. We have tested classifier models supplying different participant brain signals. It was really difficult to classify left- and right-hand movements, because they both are generated from motor cortex area of brain and samples are related. Random Forest performs better with 93.16% accuracy and decision tree (C4.5) and CART achieved 90.89 and 91.33% accuracy for binary-class hand movement data. Decision tree (C4.5) also achieved 65.66% accuracy for three-class hand movement data. We have used decision tree (C4.5) classifier model to develop the targeted application system that can control the direction of movements of a ball using brain signal of voluntary movements.

5 Developing Game Using Emotiv Epoc+ We build models by supplying training data using R programming, then these models were used to predict the possible classes. We have developed a virtual-ball movement controlling game using JavaFX and C4.5 decision tree model. This game can control the direction of movements of a ball using brain signal of voluntary movements. We have developed two versions of this game. One is for binary-class classifications and

614

Md. Ochiuddin Miah et al.

Fig. 2 Graphical User Interface (GUI) of developed game, showing movements and steady respectively

actions are steady and right-hand movements. Another one is for three-class classifications and actions are steady, left-hand and right-hand movements. The animated balls shown in Fig. 2 represents the movements and steady respectively.

6 Conclusions In this paper, we have analyzed BMI devices such as Emotiv Epoc+ 14 channel mobile EEG, Emotiv Insight 5 Channel Mobile EEG, Aurora Dream Headband, MindWave, and Muse Headband. We have used Emotiv Epoc+ EEG neuroheadset in this research. We have applied Emotiv SDK and Java technology to make a program in order to acquire the brain signals. We have used R programming for data visualization and model construction. We have applied following classifiers: OneR, naïve Bayesian classifier, decision tree (DT) classifier, Random Forest, and Bagging on both binaryclass and three-class classification problems. DT classifier (C4.5) achieved 90.89 and 65.66% for binary-class and three-class classifications, respectively. After evaluating overall results, we have considered C4.5 decision tree model to build the targeted application system. We have developed a virtual-ball movement controlling game that can control the movements of a ball using brain signal of voluntary movements without any need of conventional input devices like mouse, keyboard, joystick, etc. By playing this game user can exercise and enhance attention of his/ her mind. It helps the user to increase the attention, improves the quality of his/ her works and makes his/ her life more productive. Moreover, playing this game also provides with amusement to the user. We have developed a single-player game. In future, it can be upgraded into a multi-player interactive game in order to make it more challenging and entertaining to play. The finding of our research can be enhanced

51 Brain–Machine Interface for Developing Virtual-Ball Movement Controlling Game

615

and manipulate the movements of robotic agents and open up new opportunities in the health, rehabilitation and well being of our society. It will also have the opportunities to create humanoid robot from the understanding of brain process and engineering. Conflict of Interest The authors declare that they have no conflict of interest. Ethical Approval All applicable international, national, and/or institutional guidelines for the care and use of animals were followed. Informed Consent Informed consent was obtained from all individual participants included in the study.

References 1. Andreu-Perez J, Cao F, Hagras H, Yang GZ (2016) A self-adaptive online brain machine interface of a humanoid robot through a general type-2 fuzzy inference system. IEEE Trans Fuzzy Syst 26(1):101–116 2. Chatterjee R, Bandyopadhyay T, Sanyal DK, Guha D (2018) Comparative analysis of feature extraction techniques in motor imagery EEG signal classification. In: First international conference on smart system, innovations and computing. Springer, pp 73–83 3. Davis JJJ, Lin CT, Gillett G, Kozma R (2017) An integrative approach to analyze EEG signals and human brain dynamics in different cognitive states. J Artif Intell Soft Comput Res 7(4):287– 299 4. Farid DM, Al-Mamun MA, Manderick B, Nowe A (2016) An adaptive rule-based classifier for mining big biological data. Expert Syst Appl 64:305–316 5. Farid DM, Zhang L, Hossain A, Rahman CM, Strachan R, Sexton G, Dahal K (2013) An adaptive ensemble classifier for mining concept drifting data streams. Expert Syst Appl 40(15):5895–5906 6. Farid DM, Zhang L, Rahman CM, Hossain M, Strachan R (2014) Hybrid decision tree and naïve bayes classifiers for multi-class classification tasks. Expert Syst Appl 41(4):1937–1946 7. Guler I, Ubeyli ED (2007) Multiclass support vector machines for EEG-signals classification. IEEE Trans Inf Technol Biomed 11(2):117–126 ˇ 8. Madoš B, Ádám N, Hurtuk J, Copjak M (2016) Brain-computer interface and arduino microcontroller family software interconnection solution. In: IEEE 14th international symposium on applied machine intelligence and informatics. IEEE, pp 217–221 9. Mamun KA , Vaidyanathan R, Lutman ME, Stein J, Liu X, Aziz T, Wang S (2011) Decoding movement and laterality from local field potentials in the subthalamic nucleus. In: 5th International IEEE/EMBS conference on neural engineering. IEEE, pp 128–131 10. Pattnaik PK, Sarraf J (2018) Brain computer interface issues on hand movement. J King Saud Univ-Comput Inf Sci 30(1):18–24 11. Rahman NA, Mustafa M, Samad R, Abdullah NRH, Sulaiman N, Pebrianti D (2018) Classification of EEG signal for body earthing application. J Telecommun, Electron Comput Eng (JTEC) 10(1–2):81–85 12. Shanechi MM (2017) Brain-machine interface control algorithms. IEEE Trans Neural Syst Rehabil Eng 25(10):1725–1734 13. Sotnikov P, Finagin K, Vidunova S (2017) Selection of optimal frequency bands of the electroencephalogram signal in eye-brain-computer interface. Procedia Comput Sci 103:168–175

616

Md. Ochiuddin Miah et al.

14. Trofimov AG, Shishkin SL, Kozyrskiy BL, Velichkovsky BM (2018) A greedy feature selection algorithm for brain-computer interface classification committees. Procedia Comput Sci 123:488–493 15. Yang Y, Boling S, Mason AJ (2017) A hardware-efficient scalable spike sorting neural signal processor module for implantable high-channel-count brain machine interfaces. IEEE Trans Biomed Circuits Syst 11(4):743–754

Chapter 52

Vehicle Tracking and Monitoring System for Security Purpose Based on Thermoelectric Generator (TEG) Md. Fahim Newaz, Abu Tayab Noman, Humayun Rashid, Nawsher Ahmed, Mohammad Emdadul Islam and S. M. Taslim Reza

1 Introduction Energy harvesting or Ambient Power refers to the procedure by which a small amount of energy is generated from outer sources [1, 2]. To fulfill the continuously increasing power demands of the several power-based devices, several energy harvesting methods have emerged nowadays by light, sound, vibration, air flow, friction, fire, scavenging the mechanical, waste heat, biological etc. The produced small amount of ambient power is then supplied to an energy management system which is usually microelectronic system to convert the energy to a suitable form for supplying power to several electronic loads. Technology is probably a solution to reduce costs and prevent loss of resources. If this energy harvesting system can be implemented, it can bring a revolutionary change in automobile field. In this competitive world, it is important to preserve Md. Fahim Newaz · A. T. Noman (B) · H. Rashid · N. Ahmed · S. M. Taslim Reza Department of Electrical and Electronic Engineering, International Islamic University Chittagong, Kumira, Sitakunda, Chittagong, Bangladesh e-mail: [email protected] Md. Fahim Newaz e-mail: [email protected] H. Rashid e-mail: [email protected] N. Ahmed e-mail: [email protected] S. M. Taslim Reza e-mail: [email protected] M. Emdadul Islam Department of Electrical and Electronic Engineering, University of Science and Technology Chittagong, Khulshi, Chittagong, Bangladesh e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_52

617

618

Md. Fahim Newaz et al.

energy for the future generations. So, researchers are trying to find new paths that can fulfill the energy requirements and non-polluting too. TEG is one of them. Thermoelectric Generator (TEG) or Seebeck Effect Generator, one of the smartest ways of harvesting energy, is a gadget that change over temperature contrasts into usable power [3, 4]. For the energy consumption of the automobile two-thirds of fuel is dissipating to the surrounding as waste heat. The TEG can be used to convert this wasted heat energy to electricity to improve the total efficiency. Using that energy, a 12 V battery can be charged and different features (GPS-GSM, Music system, etc.) of an automobile and it can be easily operated by this energy. Baskar et al. [5] did an examination to consider and investigate the plausibility of retrofitting the waste warmth recuperation framework to a two-stroke petroleum motor. The test execution testing has demonstrated that the general effectiveness of two-stroke petroleum motor introduced with and without the waste warmth recuperation framework is 29.67 and 29.2% individually when the control extraction was 90%. W. Jadhao et al. [6] in his paper had broken down the likelihood of warmth recuperation from an IC motor also, the different approaches to accomplish it. He has looked at the different conceivable outcomes like the Thermoelectric age, Thermo Ionic Generation, and Piezoelectric Generation. He discovered that in Thermoelectric age, ideal temperature contrast is adequate to create the required control. Birkholz et al. [7] have actualized a similar TEG guideline to recuperate the warmth and to create control; however, the material they utilized was FeSi2. Xiaodong Zhang et al. proposed a system [8] Thermoelectric Photovoltaic Hybrid Energy Source for Hybrid Electric Vehicles, a brilliant way to recover waste heat energy. Yu and Chau [11] has proposed and executed a car thermoelectric squander warm recuperation framework by receiving a Cuk converter and a greatest power point tracker (MPPT) controller into its proposed framework as instruments for control molding and exchange. They detailed that the power change is recorded from 7.5 to 9.4% when the hot-side temperature of the TEG is warmed from 1000 to 2500 C. Likewise, when the hot-side temperature of the TEG is settled at 2500 C, the power change as much as 4.8–17.9% can be accomplished. For operating a GPS and GSM tracking system, automobile company provides power by a 24 V battery with vehicle which usually recharged by the fuel. This fuel also runs the whole system of vehicle and generate enough amount of wasted heat. After development of TEG, energy can be harvested from the wasted heat by TEG. It can be reduced the amount of waste and dimension of greenhouse effect. For the completion of this, a tracking system is designed which is powered by wasted heat. A circuit is also designed to charge the battery from TEG and a heat chamber. Haiso et al. [9] hypothesized that the efficacy of fuel usage will be more, if thermoelectric generators are employed to convert wasted heat energy into electricity. A thermoelectric module, compounding both thermoelectric generator and cooler, has been prepared to enhance efficiency of internal combustion engine. Results inferred that 51.13 mW/cm2 is engendered from the module at 290LC temperature difference and TEG was placed in both radiator and exhaust pipe. However, TEG exhibited better performance on the exhaust pipe.

52 Vehicle Tracking and Monitoring System for Security Purpose …

619

Karri et al. [10] study two distinct commercially available thermoelectric materials, bismuth telluride (Bi2Te3) or quantum well (QW), were used to harvest energy from SUV and CNGs exhaust steam. Outcome of this paper has two cases, first for SUV, using Bi2Te3 has 6.25 times more relative fuel savings than QW generators, second for CNG, both of the generators Bi2Te3 and QW s relative fuel savings below 5%. It is also inferred from the result that due to parasitic losses, SUV generator provides negative fuel, ironically, this phenomenon is absent in CNG generators trite. Yu et al. [11] introduced with TE waste heat energy recovery system. In this system maximum power point tracking technique is attributed for utilizing the battery voltage and current to maximize the output power rather than using direct TEG output voltage and power. The regulation of power flow between TEG and battery pack is maintained by DC-DC C0uk converter. This system may have functioned well under different working condition according to their result. Wang et al. [12] delineated a mathematical model of a TEG where exhaust gas of vehicle considered as heat source. It can be inferred from the result, in contrast with low-temperature side and high-temperature side, the output power and efficiency of high temperature side can be augmented by changing the convection heat transfer coefficient of that side. The result also reveals that electrical output at most 1.9 kW, plant efficiency obtained 8.9% and a peak value at 1023 K of PN temperature while the figure of merit of TE material is unity. This model, based on Fouriers law and Seebeck effect, has theoretically disseminated that maximum output power and efficacy of TEG can be obtained if the external resistance is greater than the internal resistance. Dr. Faruk Yildiz et al. [13] intended to find the most temperature differences occurring place during the operation of condenser. To conduct the measurement: a number of thermometers were kept inside the condenser unit with small finite spacing among meters. Data has been recorded during amid day and evening time, while the trite of outside open air temperature considered as temperature variation. Enumerate number of sub-studies affiliated with the sustainable power source ventures foundations has introduced a thermoelectric generator comprised with an aerating and cooling condenser unit. Different aspects of empirical study like efficiency, power generation prospect, potential consumer applications, expense, and system installation complexity were implied by this paper. To simulate a balance system several components: aluminum, copper, brass, and bronze composed in the system even though the copper augmented the expense. Finally, authors suspected more 20 W can be harvested by TEG.

2 Thermoelectric Generator (TEG) Thermoelectric Generator converts wasted heat energy into electricity without any need of moving power-generating turbines. The block diagram of the TEG is shown in Fig. 1. The main working principle of the TEG is the temperature difference on the both ends of the module.

620

Md. Fahim Newaz et al.

Fig. 1 The block diagram of TEG

Fig. 2 The internal construction of TEG [14]

Figure 2 shows the internal structure of the TEG. This module consists of a heat collecting side, a heat releasing side, p-type and n-type pallets sandwiched vertically in between both sides. The combined set of this is known as thermoelectric modules which are strong state coordinated systems that utilize three set up thermoelectric impacts known as the Peltier, Seebeck, and Thomson impacts. TEGs work by exploiting a temperature gradient between the two sides of the generator. If heat is applied on an end of a metal and simultaneously cool the other

52 Vehicle Tracking and Monitoring System for Security Purpose …

621

end, the electrons surrounding the metal of the hot end will have more energy than the equivalent electrons of cold end. So, the electrons of the hot end jiggling around faster than the cold end electrons. As a result, the hot end becomes more positively charged and the cold end becomes more negatively charged. Nowadays, ambient power is getting popularity for its harmless nature. It reduces the requirement of energy storage devices or enhances the time between energy storage device replacements. Temperature change effect (thermoelectric-Seebeck), one of the ambient power generating method, is found in buildings, machines, bridges, staircases, indoor, and outdoor temperature differences, and the human body.

3 Methodology The main purpose of this project is development of a tracking system powered by TEG. A 12 V 60 W TEC1-12706 Thermoelectric Cooler Peltier is used for harvesting energy from the wasted heat. Figure 3 shows the functional block diagram of the proposed system. This block diagram represents the whole working procedure of the TEG powered tracking system. This device will notify about the vehicle position and ambient temperature via SMS. It also shows its position on a display. This is done by the GPS and GSM module. This module is run by a 12 V battery which is charged by TEG via booster circuit. Energy from the wasted heat of automobile is being harvested by the TEG. That harvested energy is boosted and utilized by the booster circuit for recharging the battery. Battery is giving power to the microcontroller, GPS and GSM

Fig. 3 Functional block diagram of the proposed system

622

Md. Fahim Newaz et al.

module. Microcontroller is being notified about the position by the GPS and GSM module. Microcontroller also transfers the notification to the display and emergency contacts. Position of the vehicle is showed by the display.

4 System Development PIC16F876A used as the microcontroller. The circuit also consists of TEG, GPS and GSM module, Display, Heat Sensor, Crystal oscillator, Switching Transistor, and Voltage regulator. Figure 4 demonstrates the circuit diagram of the proposed vehicle Tracking and monitoring device. Pin 1 of the microcontroller or MCLR (Master Clear) is used as reset pin. RA0 and RA1 are connected to heat sensor LM35 and Thermo Electric Generator respectively. RB0 and RB1 are coupled with Rs and Enable pin of a 16 × 2 LCD display. RB4RB7 pin coupled with four data pin of LCD display. RC6 and RC7 are connected to Tx and Rx pin of GPS-GSM module, respectively. Switching Transistor is attached to RC2.

Fig. 4 Circuit diagram of the proposed system

52 Vehicle Tracking and Monitoring System for Security Purpose …

623

Fig. 5 Flow chart of the proposed system

Charging and Discharging curve of the Battery from waste heat is shown in Figs. 12 and 13. When the temperature difference creates between two sides of TEG, current will flow through the circuit as well as a voltage difference will be developed between the sides. Eventually, a battery of 12 V will be charged by this voltage. That is why a boost converter has been used in between TEG and battery. This converter consists of two semiconductor devices (BJT and Diode) and energy storage elements (capacitor and inductor) in two combinations.

624

Md. Fahim Newaz et al.

TEG provides energy to microcontroller and controller gives a PWM signal to turn on the switching transistor. Switching operation of the boost converter is done by this switching transistor. After boosting up the voltage current goes down to mA level TEG starts to recharge the battery. Two voltage regulators are connected with the battery via capacitor. One of them concentrate on the circuit and other one provides sufficient voltage to the GPS and GSM module. TEG is directly connected with the microcontroller via pin 3. Temperature sensor LM35 is connected by pin 2 of port A. It senses the temperature and provides the data to GSM. SIM 808 GPS and GSM module is connected with microcontroller by port C. Display module is connected with port B. While GPS and GSM module search the satellite and BTCL towers, display shows their status. After successful operation result is showed in display and it sends the information via SMS to the emergency contact. A crystal oscillator of 16 MHz is connected with pin 9 and 10 of the microcontroller. When 0 V appears in the microcontroller pin or pin no 1, microcontroller resets the operation. For this purpose, a switch is connected between pin no 1 and ground. This switch resets the system. Figure 5 illustrates the flow chart of the system. After initializing the TEG, if there is any temperature difference, charge the battery. If there is no temperature difference on TEG, activate boost converter. Then sequentially, detect GPS location, Display the coordinates on LCD display and send the coordinates to the user phone at a fixed interval. This process will continue until the system turns off.

5 Result Analysis A. Physical Output. Figure 6 shows the Overview of the proposed system. In the figure, charging of battery is indicated by the LED which is marked by red circle. While temperature difference appears between two sides of TEG, LED starts to blink. Blue mark portion of above picture is showing two pods. If we connect an ammeter in between two pods it will show the current of TEG. Figure 7 shows the position of the vehicle displaying on LCD. Figure 8 shows the message containing GPS location sent through GSM to the Owners phone.

52 Vehicle Tracking and Monitoring System for Security Purpose …

625

Fig. 6 Complete overview of the project Fig. 7 Displaying GPS location of the vehicle on LCD

B. TEG Performance. Some graphical representation of TEG output illustrates in Figs. 9, 10 and 11 containing Temperature versus Power, Temperature versus Voltage and Temperature versus Current curve.

626 Fig. 8 Message sent to the user phone containing the GPS location of the vehicle through GSM

Fig. 9 Temperature versus power curve of the system

Fig. 10 Temperature versus voltage curve of the system

Md. Fahim Newaz et al.

52 Vehicle Tracking and Monitoring System for Security Purpose …

Fig. 11 Temperature versus current curve of the system

Fig. 12 Time versus percentage curve of battery charging from waste heat Fig. 13 Time versus percentage curve of battery discharging or consumption

627

628

Md. Fahim Newaz et al.

6 Comparison with Previous Works Hsiao et al. [9] concentrate on heat recovery from automobile engine. They did not utilize the heat which was recovered from engine. In our system, we designed a tracking system which is powered by the wasted heat of automobile engine. We utilize the power by operating a tracking system. Wang Yuchao et al. [12] harvested energy from the exhaust gas of automobile engine. Exhaust gas does not maintain a constant temperature. This is difficult to get uniform output from TEG without constant temperature difference. In our project, we used only the heat it may be from gas, metal, smoke, etc. It is far better to use heat from any part of automobile rather than gas. Dr. Faruk Yildiz et al. [13] discussed about harvesting low power from air conditioner. This harvested power was very low and unable to recharge a battery. We have charged a battery by the harvested power. We also design a tracking system which operates with this harvested power.

7 Conclusion Energy is the one of the most important resource since the development of a country measures on how much power the country consume. It is important to preserve energy for upcoming generations. But most of the power-generating methods are not eco-friendly. As a result, some change in power generating methods were needed. In this paper, we tried to develop a self-powered, green powered system to mitigate the pollution and pressure on the power generation. The hardest part of the project was harvesting energy and recharging the battery from the harvested power. But with great motivation and effort, finally successful implementation of the project has become possible. In this project, design and development of a tracking system of vehicle powered by using the wasted heat of vehicle harvested through TEG. However, making a tracking system for vehicle using TEG is done but more features are not possible to include due to limitation of cost and availability of equipment. So many more features can be added to the device in future. a. Music system, Air conditioner, display, other indicating meters etc. of a vehicle can be run through this technology. b. Biomedical device can be developed on this energy harvesting method. Animal or Human tracking device can be developed on this energy harvesting method with cost-effective cooler for microprocessor can be designed by TEG.

52 Vehicle Tracking and Monitoring System for Security Purpose …

629

References 1. Kang T, Kim S, Hyoung C, Kang S, Park K (2015) An energy combiner for a multi-input energy harvesting system. IEEE Trans Circuits Syst II: Express Briefs 62(8):911–915 2. Ammar A, Reynolds D (2018) Energy harvesting networks: energy versus data cooperation. IEEE Commun Lett 22(10):2128–2131 3. Chalasani S, Conrad JM, A survey of energy harvesting sources for embedded systems 4. Auckland DW, Shuttleworth R, Luff AC, Axcell BP, Rahman M (1995) Design of a semiconductor themoelectric genertor for remote subsea wellheads. IEE Proc - Electric Power Appl 142(2):65–70 5. Baskar P, Seralathan S, Dipin D, Thangavel S, Norman I, Francis C, Arnold C (2014) Experimental analysis of thermoelectric waste heat recovery system retrofitted to two stroke petrol engine. Int J Adv Mech Eng 4(1):9–14 6. Jadhao J, Thombare D (2013) Review on exhaust gas heat recovery for I.C. engine. Int J Adv Mech Eng 2(12):93–100 7. Birkholz E, Grob U, Voss K (1988) Conversion of waste exhaust heat in automobiles using FeSi2 thermo-elements. In: 7th international conference on thermoelectric energy conversion, pp 124–128 8. Zhang X, Chau K, Chan C (2009) Design and implementation of a thermoelectric photovoltaic hybrid energy source for hybrid electric vehicles. World Electric Veh J 3(2):271–281 9. Hsiao Y, Chang W, Chen S (2010) A mathematic model of thermoelectric module with applications on waste heat recovery from automobile engine. Energy 35(3):1447–1454 10. Karri M, Thacher E, Helenbrook B (2011) Exhaust energy conversion by thermoelectric generator: two case studies. Energy Convers Manag 52(3):1596–1611 11. Yu C, Chau K (2009) Thermoelectric automotive waste heat energy recovery using maximum power point tracking. Energy Convers Manag 50(6):1506–1512 12. Wang Y, Dai C, Wang S (2013) Theoretical analysis of a thermoelectric generator using exhaust gas of vehicles as heat source. Appl Energy 112:1171–1180 13. Yildiz F, Coogler KL (2014) Low power energy harvesting with a thermoelectric generator through an air conditioning condenser. In: ASEE anual conference on exposition 14. Bos JW (2012) Thermoelectric materials: efficiencies found in nanocomposites. Edinburgh (2012)

Chapter 53

Improved Subspace Detection Based on Minimum Noise Fraction and Mutual Information for Hyperspectral Image Classification Md. Rashedul Islam, Md. Ali Hossain and Boshir Ahmed

1 Introduction Hyperspectral datacube contains hundreds of image bands with a fine spectral resolution e.g., 0.01 μm covering visible to near infrared range of the frequency spectrum [1]. Each of these image bands is termed as individual feature as they contain different intensities for each of the ground objects [1]. Therefore, hyperspectral imageries are the rich source of information for the detection of ground objects. However, this high dimensionality presents many difficulties such as the input image bands of hyperspectral datacube are highly correlated. In addition, for a specific application, all the bands are not similarly significant [2]. Since the hyperspectral sensor captures the images on a continuous spectral range, some image bands contain less discriminatory information about the ground objects [3]. Another important challenge is the classification of this high dimensional data cube since enough training samples are not available. For example, if the proportion of the training d and the number of input images becomes very small, the classification accuracy of the test samples starts to decrease gradually and this effect is termed as ‘Hughes phenomena’ or curse of dimensionality [4]. Therefore, it is important to reduce the high dimensional data to relevant subspace for improving the classification accuracy. Thus, an effective technique is a concern in this paper. Different feature reduction approaches can be applied for effective classification of hyperspectral images and it can be achieved by feature extraction and/or feature Md. Rashedul Islam (B) · Md. Ali Hossain · B. Ahmed Computer Science & Engineering, Rajshahi University of Engineering & Technology, Rajshahi, Bangladesh e-mail: [email protected] Md. Ali Hossain e-mail: [email protected] B. Ahmed e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_53

631

632

Md. Rashedul Islam et al.

selection. However, feature extraction maps the input images to a new space K from original space N where K < < N through linear or nonlinear transformation. The most commonly used unsupervised feature extraction approach is Principal Component Analysis (PCA) [5, 6]. Although PCA is based on the fact that adjacent bands are highly correlated and used to transform the original higher dimensional data to lower dimension by eradicating the correlation among the bands using the higher variance, it does not consider the noise factor of an image [7]. The contraposition of hyperspectral images did not consider the actual Signal to Noise Ratio (SNR) [7]. Therefore, small variance does not mean the poor image quality, higher variance may have lower SNR as compared to other bands [8]. While PCA has been effectively used in many remote sensing data, it is not scale invariant, is quite variable with respect to the information content of a particular image, and does not surety good class differentiation in the transformed space [9, 10]. Hence, minimum noise fraction is proposed as the better technique for feature extraction as it can minimize the disadvantages of PCA depending on image quality. In MNF, the components are arranged in terms of SNR, no matter how noise is distributed in spectral bands [7]. Although feature extraction transforms the original large data to a new space with few features but ranking the new features is the major concern shown in some research [1, 2, 11–13]. Since MNF is unsupervised technique which solely considering SNR, some of the class may affect the classification accuracy which is not present in the first few features. Therefore, only MNF is not an effective way for dimensional reduction. So, feature selection is applied for effective feature reduction. Mutual information (MI) is a popular supervised feature selection methods and able to measure in both the linear and nonlinear associations between the images bands and the target classes which makes it suitable for effective subspace detection [2, 14–16]. But it is difficult to rank features based solely on MI between two variables because the MI does not have any range. Therefore, a comparison of two MI values may not be always preferable. Therefore, this paper scales the MI value to a specific range i.e., 0–1 in order to compare two values [17, 18]. Thus, an improved subspace detection technique is proposed using a normalized mutual information (NMI) over the new generated features. This new feature space is able to maximize the relevant features and avoid the redundant of the chosen features.

2 Methodology 2.1 Minimum Noise Fraction (MNF) MNF can estimate the inherent feature dimension of image availability, and its existence is a superposition of two PCA. The MNF is suitable as it selects the signal to noise ratio rather than the global variance in order to measure relevant features [7].

53 Improved Subspace Detection Based on Minimum Noise Fraction …

633

If the input hyperspectral image is denoted as X, where X = [x1 , x2 … xp ] T . Here, p is the total amount of spectral bands. If noise is present in the signal then, X = S + N, where S and N are the signal component and the noise component in the images. The covariance matrix can be expressed as C(X) =



=

 S

+

 N

(1)

where  S and  N are the covariance matrices of the signal and noise, respectively. Using the ratio of noise variance, the linear transformation MNF can be defined as Y = AT X

(2)

where the matrix A is the eigenvector matrix of  −1  N = Λ A. The diagonal matrix Λ is an eigenvalues matrix. The noise ratio of the corresponding component of Λ is  V ar (aiT N) aiT N ai  = V ar (aiT X) aiT X ai

(3)

where Var{} calculates the variance and ai is the ith component of the eigenvector matrix A. The hyperspectral images are arranged in accordance with higher SNR through the corresponding MNF transformation. The first few components carry the less noisy data of the image and the decreasing component contains more noises in information unlike PCA which arrange with global variance.

2.2 Mutual Information Based Feature Selection Criteria The mutual information is used to measure the dependency of two input variables X and C and defined as I(X, C) =

 c∈C

 x∈X

p(x, c) log

p(c, x) p(x) p(c)

(4)

where p(x), p(c) are the marginal probability distributions and p(x, c) is the joint probability distribution of X and C. If X be an input image band and C is the input class label then MI can be expressed in terms of entropy as well. I(X, C) = H(X) + H(C) − H(X, C)

(5)

where H(X) and H(C) are the entropy and H (X, C) is their joint entropy of X and C.

634

Md. Rashedul Islam et al.

The MI value in (4) or (5) is used as the selection criterion for selecting the most effective and informative features. To select the subset of relevant features, MI is measured between the new features generated from MNF Yi and the available training class labels C. Thus, most informative feature is calculated as [19]: Max V =I(Yi , C)

(6)

However, the selected features using (6) may have some redundancy. Therefore, the redundancy is also measured. The objective is to minimize the redundant and maximize the relevant features among the chosen features. Therefore, the selected model of subspace detection can be defined as G(Yi , K ) = Maxi∈X [I (Yi , C) −

1  I (Yi , Y j )] i, j∈S |S|

(7)

However, the MI value G (Yi , k) in the above method is difficult to use directly because it is influenced by entropy of two variables and not constrained to a specific range. To measure how good a MI value is, it is normalized to the range [0, 1] [17, 20]. I (Y, C) Iˆ(Y, C) = √ H (Y )H (C)

(8)

Thus, the proposed subspace detection method has been defined as  ˆ i , k) = Maxi∈X [ Iˆ(Yi , C) − 1 G(Y Iˆ(Yi , Y j )] |S| i, j∈S

(9)

Following is a summarized algorithm for the proposed feature reduction method. Algorithm a. MNF is applied on original datasets Xi to obtain Yi . New feature is selected based on the higher value of MNF components. b. Extract training data from each MNF component Yi . c. Evaluate the max Iˆ(Y, C) as defined in (8). d. Now select a subset of MNF components or output features based on Eq. (9). Output the subset S having the selected features for classification.

53 Improved Subspace Detection Based on Minimum Noise Fraction …

635

3 Experiments 3.1 Experimental Procedure For the experiment presented here, two real hyperspectral image data sets were used. The data set 1 consist of 220 bands is collected by AVIRIS sensor over the Indian Pines test site in USA having 145 × 145 spatial resolution where sixteen classes are defined [20]. For the insufficient training data, “Oats” and “Grass/Pasture mowed” were not used in the experiment of the AVIRIS data. The data set 2 consisted of 191 channels with 1280 × 307 pixels and was collected by the Hyperspectral digital imagery collection experiment (HYDICE) sensor over the Washington DC MALL in 1995 [21]. For data set 2, “paths” were not used for inadequate training samples. Tables 1 and 2 represent the training and testing samples for AVIRIS and HYDICE respectively and were used in the experimental procedure. The training and testing samples are selected for classification using the ground-truth of the original image as shown in Figs. 1 and 2. Here, the kernel support vector machine (SVM) classifier with RBF kernel used for classification accuracy measurement. The classifier has been trained using 10fold cross-validation to select the cost parameter C and kernel width γ [22]. The kernel parameters (C = 10 and γ = 2.9) for AVIRIS data and (C = 10 and γ = 3) for HYDICE were selected for classification. The top ranked 9 features of AVIRS data and 10 features of HYDICE data are used in classification.

Table 1 Training and testing datasets (AVIRIS)

Class

Training data

Testing data

Alfalfa

16

Wheat

43

65

Bldge-grass

22

17

Soybean-min

131

166

Stone-steel

16

26

21

Soybean-notill

108

88

Grass/Pasture

109

73

Corn-notill

47

42

Soybean clean

20

11

Corn-min

65

65

Hay-windrowed

166

137

Woods

279

248

Grass/Trees

96

80

Corn

66

64

636 Table 2 Training and testing datasets (HYDICE)

Md. Rashedul Islam et al.

Class Shadow

Training data

Testing data

20

16

Tree

367

1056

Roof

117

108

Water

425

476

Street

299

369

Grass

361

850

Fig. 1 AVIRIS 92AV3C data with RGB image (left side) and reference dataset (right side)

3.2 Feature Extraction Results In this analysis, new features are generated through MNF, then the feature selection is accomplished on the new features based on the normalized mutual information. We call this approach as MNF-nMI. The proposed approach (MNF-nMI) is compared with two standard methods such as MNF and PCA. For each algorithm, the raking of the selected features is listed in Table 3. It can be noticed that the proposed MNFnMI select the MNF component-2 as the first ranked feature as MNF component-1 is the noisy feature. Figure 3 visually shows that the MNF component-1 is noisy as compared to MNF component-2. Figure 3 visually shows the benefits of applying nMI over traditional MNF image of the MNF component of 1 and MNF component of 2 of AVIRIS data. Figures 4 and 5 show the subspace projection of the AVIRIS data and indicate that the MNF component-2 is more separable than MNF component-1.

3.3 Classification Results The achievement of the proposed method is assessed in terms of classification accuracy. The selected features listed in Table 3 is applied for classification with kernel SVM. The proposed method was compared with popular PCA and MNF approach.

53 Improved Subspace Detection Based on Minimum Noise Fraction …

637

Fig. 2 HYDICE data with RGB image (left side) and reference dataset (right side)

Table 3 Selected features for classification

Data set

Methods

Orders of ranked features

AVIRIS (Indian Pines)

PCA

PC: 1, 2, 3, 4, 5, 6, 7, 8, 9

MNF

MNF-C: 1, 2, 3, 4, 5, 6, 7, 8, 9

MNF-nMI

MNF-C: 2, 4, 3, 6, 5, 7, 8, 9, 10

PCA

PC: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10

MNF

MNF-C: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10

MNF-nMI

MNF-C: 2, 4, 5, 3, 6, 1, 191, 189, 188, 11

HYDICE (Washington DC)

638

Md. Rashedul Islam et al.

Fig. 3 MNF component 1 (left) and MNF component 2 (right)

Fig. 4 Feature space of MNF component-1 and MNF component-2 of AVIRIS data

Figures 6 and 7 represent the overall classification accuracy of AVIRIS and HYDIC datasets for individual approach and dataset with respect to the order of selected features. The classification accuracy without feature reduction for the original AVIRIS data with first 10 features is 64.41% which motivate the feature reduction. The classification accuracy of conventional PCA and MNF with first 9 features is 92.84% and 88.07% respectively. The proposed method shows the classification accuracy of 96.82% for AVIRIS data, which is higher than the other two existing methods. For

53 Improved Subspace Detection Based on Minimum Noise Fraction …

639

ClassificaƟon accuracy (%)

Fig. 5 Feature space of MNF component-2 and MNF component-3 of AVIRIS data 100 90 80 70 60 50 40 30 20 10 0

PCA MNF MNF-nMI 1

2

3

4

5

6

7

8

9

Number of ranked features Fig. 6 Classification accuracy in percentage of AVIRIS data

the HYDICE data, the conventional PCA and MNF with 10 features show the classification accuracy of 93.36% and 96.10% respectively. The proposed method shows the classification accuracy of 99.30% with 10 selected features, which is higher than the other two existing methods. Table 4 listed the classification accuracy of PCA, MNF, and proposed MNF-nMI.

Md. Rashedul Islam et al. ClassificaƟon accuracy (%)

640 100 90 80

PCA MNF MNF-nMI

70 60 50

1

2

3

4

5

6

7

8

9

10

Number of featured ranks

Fig. 7 Classification accuracy in percentage of HYDICE data Table 4 Classification accuracy of PCA, MNF, and MNF-nMI

Data set

PCA (%)

MNF (%)

MNF-nMI (%)

AVIRIS

92.84

88.07

96.82

HYDICE

93.36

96.10

99.30

4 Conclusion Hyperspectral sensor generates a voluminous high dimensional data that needs dimensional reduction procedure. A combination of feature extraction (MNF) and feature selection (MI) is proposed in this paper for the objective of dimensionality reduction in hyperspectral remote sensing images. Feature selection over feature extraction improves the quality of the output features from traditional MNF. This is because MNF-nMI finds the subspace which is less noisy and provide relevant information about the desired ground objects. The improvement in accuracies proves the suitability of the proposed approach. The proposed MNF-nMI is also capable to provide better results when only a few training samples are available.

References 1. Jia X, Kua B, Crawford MM (2013) Feature mining for hyperspectral image classification. Proc IEEE 101(3):676–679 2. Hossain MA, Jia X, Pickering M (2014) Subspace detection using a mutual information measure for hyperspectral image classification. IEEE Geosci Remote Sens Lett 11(2):424–428 3. Guo B, Gunn SR, Damper RI, Nelson JDB (2006) Band selection for hyperspectral image classification using mutual information. IEEE Geosci Remote Sens Lett 3(4):522–526 4. Hughes G (1968) On the mean accuracy of statistical pattern recognizers. IEEE Trans Inf Theory 14(1):55–63 5. Pechenizkiy M, Tsymbal A, Puuronen S (2004) PCA-based feature transformation for classification: issues in medical diagnostics. In: Proceedings of the 17th IEEE symposium on computer-based medical systems, pp 535–540 6. Rodarmel C, Shan J (2002) Principal component analysis for hyper-spectral image classification. ACM Surv Land Inf Sci 62(2):115–122

53 Improved Subspace Detection Based on Minimum Noise Fraction …

641

7. Green AA, Berman M, Switzer P, and Craig MD (1988) A transformation for ordering multispectral data in terms of image quality with implications for noise removal. IEEE Trans Geosci Remote Sens 26(1):65–74 8. Chang C, Du Q (1999) Interference and noise-adjusted principal components analysis. IEEE Trans Geosci Remote Sens 37(5):2387–2396 9. Gao L, Zhao B, Jia X, Liao W, Zhang B (2017) Optimized Kernel minimum noise fraction transformation for hyperspectral image classification. Remote Sens 9(6):548 10. Luo G, Chen G, Tian L, Qin K, Qian S (2016) Minimum noise fraction versus principal component analysis as a preprocessing step for hyperspectral imagery denoising. Can J Remote Sens 42(2):106–116 11. Hossain MA, Pickering M, Jia X (2011) Unsupervised feature extraction based on a mutual information measure for hyperspectral image classification. Proc Int Geosci Remote Sens Symp 1720–1723 12. Hossain MA, Jia X, Pickering M (2012) Improved feature selection based on a mutual information measure for hyperspectral image classification. Proc Int Geosci Remote Sens Symp 3058–3061 13. Hossain MA, Jia X, Pickering M (2013) Subspace detection based on the combination of nonlinear feature extraction and feature selection. In: IEEE workshop on hyperspectral image and signal processing: evolution in remote sensing 14. Richards JA, Jia X (2006) Remote sensing digital image analysis, 4th edn. Springer, Berlin 15. Qi C, Zhou Z, Wang Q, Hu L (2016) Mutual information-based feature selection and ensemble learning for classification. In: 2016 international conference on identification, information and knowledge in the internet of things (IIKI), pp 116–121 16. Fu Y, Jia X, Huang W, Wang J (2014) A comparative analysis of mutual information based feature selection for hyperspectral image classification. In: 2014 IEEE China summit & international conference on signal and information processing (ChinaSIP), pp 148–152 17. Estevez PA, Tesmer M, Perez CA, Zurada JM (2009) Normalized mutual information feature selection. IEEE Trans Neural Netw 20(2):189–201 18. Vinh LT, Lee S, Park YT et al (2012) Appl Intell 37:100. https://doi.org/10.1007/s10489-0110315-y 19. Peng H, Long F, Ding C (2005) Feature selection based on mutual information: criteria of maxdependency, max-relevance, and min-redundancy. IEEE Trans Pattern Recognit Mach Learn 28(8):1226–1238 20. Tarabalka Y, Benediktsson JA, Chanussot J (2009) Spectral-spatial classification of hyperspectral imagery based on partitional clustering techniques. IEEE Trans Geosci Remote Sens 47(8):2973–2987 21. Landgrebe DA. https://engineering.purdue.edu/~biehl/MultiSpec/hyperspectral.html 22. Hsu C, Chang C, Lin C (2003) A practical guide to support vector classification. In: Appendix: Springer-author discount

Chapter 54

A RSA-Based Efficient Dynamic Secure Algorithm for Ensuring Data Security Himadri Shekhar Mondal, Md. Tariq Hasan, Md. Mahbub Hossain, Md. Mashrur Arifin and Rekha Saha

1 Introduction Cyberattack has become more diverse with its increasing significant features. The modern digital transformation is facing major challenges due to cyber threat. As the cyberattack is increasing day by day, digital infrastructure needs more initiatives for facing the increasing malicious cyberattack [11]. Ensuring cyber security has become a major concern, nowadays, as the necessary data related to health, personal, professional as well as financial information are stored in the web for various purposes. So for providing better privacy of data in the cyberspace, proper planning is necessary with updated knowledge and it needs to implemented using developed tools. Many years ago, data networking mainly deals with connectivity issue, not security, but with the passing of time system needs more security by detecting and preventing the threats quickly. Being a moving target, the cyberattacks always try to increase their capability of damaging a system in a big scale. So, the quick response against the ever-increasing threat is necessary to prevent or destroy them. Cybersecurity is needed for better protection of individual or any professional organization. By preventing cyberattacks, it ensures faster growth of an organization as well as adds great value in the digital world. Cryptography is all about encrypting data for ensuring privacy. Cryptography can’t confirm secured communication always, it depends on how one uses it [10]. If anyone hires a security equipment or technique from any untrusted organization and uses it in the security purpose, then the system may be vulnerable as the endpoint can be controlled and used to wreak the network [6].

H. Shekhar Mondal (B) · Md. Tariq Hasan · Md. Mahbub Hossain · Md. Mashrur Arifin · R. Saha Electronics and Communication Engineering Discipline, Khulna University, Khulna, Bangladesh e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4_54

643

644

H. Shekhar Mondal et al.

Fig. 1 A scheme for encryption and decryption

Various cryptographic methods are used for ensuring data security in the networking system. Each method has its own encryption and decryption process which allows only authorized persons reading the encrypted message [15]. Figure 1 represents a generalized scheme of encryption and decryption process. The encryption and decryption are mainly based on different algorithmic processes established by security analyzers. The security system varies from one network to another network for having better security and performance based on the design of the network. In this research, two algorithms are used for increasing data security. One algorithm is RSA algorithm which was invented by Ron Rivest, Adi Shamir, and Len Adleman in 1977 [16]. This has become a popular algorithm because of its encryption process without the need of separate key for exchanging purpose. On the other hand, Diffie–Hellman algorithm [8] is mainly used for identical shared secret of both systems. The interesting fact is that the communications can take place among them who have never communicated with each other. The traffic is encrypted between the two-end devices. This research combines these two algorithms and ensures better security. The outline of this paper is designed as, Sect. 2, related works, in which, cryptography and cyber security related previous researches are discussed. The proposed methodology is discussed in the Sect. 3, where the two algorithms are also discussed. Based on the proposed methodology, results and features of this proposed algorithm are presented in Sect. 4. Finally, the conclusion is presented in Sect. 5.

2 Related Works Ciphering is considered as the secrecy of data, which is used to send secret message using secret code which is unintelligible to all except the receiver. The history says the use of cryptography was popular during the war as the central direction came as cipher for ensuring its security. Also, if we look back in early Egyptian civilization, we can see they used the cryptography with the help of hieroglyphs. To prevent huge number of increasing attacks, various algorithms are developing day by day. The algorithms are mainly divided into two categories, one is symmetric and the other is asymmetric. Symmetric key cryptography is defined as a method in which the same key is shared by both sender and receiver as it is implemented as stream cipher or block cipher [4]. Some of popular symmetric encryption-based algorithms are AES, DES, RC5, RC6, Blowfish, etc. Two common cryptographic technique Data Encryption Standard (DES) and Advance Encryption Standard (AES) use block cipher for ensuring better security. For having numerous success, DES was a popular algorithm and used in a wide range of applications including ATM encryption, increasing email privacy and secure remote access [7]. In the case of asymmetric

54 A RSA-Based Efficient Dynamic Secure Algorithm for Ensuring Data Security

645

key cryptography mainly, there are two keys among them one is public and the other is private or secret key. Although different, the keys are mathematically linked as one key which is used for encryption and the other is used for decryption mechanism [2]. Some asymmetric encryption-based algorithms are RSA, Diffie–Helman, ECC, etc. As the RSA algorithm is considered as asymmetric key cryptography, it has two keys involved, one is public key and the other is private key. Mathematically, it is tough to guess the public key and the complex private key by human but it is comparatively easy for the computer to calculate and generate the keys. The encryption process is done using public key, but the decryption process is done using private keys and to guess the private keys having known the public key is extremely tough [17]. An Enhanced and Secure RSA Key Generation Scheme (ESRKGS) was proposed by Karatas et al. [13]. They used an alternate private key to break a security system. The results presented by ESRKGS was similar to RSA algorithm. Barani et al. [3] developed a mechanism, which was used for transferring data in Vehicular Ad-hoc Networks (VANET). RSA algorithm is used here for security purpose, which ensures the safe transformation of network data. A mechanism for Ambulance Traffic Control System (ATCS), which was helpful for an ambulance to reach the hospital quickly was developed by Leu et al. [12]. In this system, communication among ambulances can also be done and the communication system used RSA algorithm for ensuring better security. A secure authentication system was proposed by Mittal [14], which uses Diffie–Hellman algorithm along with hash function and XOR operations. This system can be useful for preventing various malicious attacks such as man in the middle attack, insider attack, etc., by establishing a secure communication scheme between server and user. Abdalla et al. [1] proposed a security scheme based on Diffie– Hellman algorithm which provides security against ciphertext- based attacks. The proposed mechanism combines hash function along with an authentication code.

3 Methodology Implementing meticulous network security has become demandable in this modern era. As various attacks are rushing in the modern networking system to make it vulnerable, high level of encryption is needed to ensure higher security as well as culminate performance. To prevent the modern attacks, various encryption techniques are developing. This research is mainly featured with RSA algorithm. Two other popular algorithms also studied which are DES and AES. The reason behind choosing RSA algorithm is mainly because of its features based on key size, block size, simulation speed, etc., which are summarized in Table 1.

646

H. Shekhar Mondal et al.

Table 1 Comparison of DES, AES, and RSA algorithm Key parameters AES DES Developed (year) Key size (bits) Block size (bits) key type (Enc. and dec.) Algorithm features Simulation performance Trojan horse (virus)

2000 128, 192, 256 128 Same Symmetric Faster Not proved

1977 56 64 Same Symmetric Faster No

RSA 1978 >1024 Minimum 512 Different Asymmetric Faster No

3.1 The RSA Algorithm RSA algorithm generally uses modular arithmetic process for encryption and decryption. The security criteria of this RSA algorithm is mainly based on factorizing n and obtaining prime numbers, say, p and q. Obtaining n is easy as it can be easily calculated by multiplying p and q, it is very tough to obtain p and q, which is a reverse dynamism of factorization of n, especially in the case of having large values of p and q. Because of having numerous advantages, RSA algorithm is now using in various electronic commerce protocols ensuring better protection by generating long keys and updated utilization of implements [5]. The RSA algorithm basically has two parts, which are, 1. Key generation. 2. Encryption and decryption process. Key Generation 1. Take the two prime numbers p and q indiscriminately considering p! = q. 2. Now calculate modulus n = p*q (Should use prime numbers having large values, for increasing difficulty for factorization). 3. Find  = ( p − 1) (q − 1). 4. Next exponent e should be chosen in a way that 1 Alt4 > Alt1 > Alt2 > Alt5

Alt3

By [11]

Alt3 > Alt1 > Alt4 > Alt2 > Alt5

Alt3

By [6]

Alt3 > Alt1 > Alt2 > Alt4 > Alt5

Alt3

By IFLWA

Alt3 > Alt1 > Alt4 > Alt2 > Alt5

Alt3

Alt3 > Alt1 > Alt4 > Alt2 > Alt5 Hence, alternative Alt3 is the appropriate supplier, while on other hand Alt5 is the worst supplier among the alternatives. From Table 7, we see that when proposed operator IFLWA is compared with other operators mentioned in the literature, it also gives Alt3 as the best alternative.

6 Conclusion In an Atanassov Intuitionistic fuzzy sets, each expert has assigned both membership value and non-membership value with a certain degree of hesitation. The hesitancy in the opinion of the experts appear due to incompleteness in information. Therefore, preciseness in both membership value and non-membership value also becomes highly difficult. MCDM problems have many conflicting criteria on which decision makers knowledge is imprecise and incomplete. Fuzzy set theory is inadequate to deal with all such problems and therefore, AIFS theory is used to overcome this complex situation. Handling of linear operators is simple in comparison to any nonlinear operator. Therefore, with easy implementation of linear operator, the aggregation process also becomes easy. In this paper, SCM is solved while utilizing the proposed IFLWA operator in the TOPSIS method. On comparing our results with the numerical problem stated with other aggregation operators, we find that the similar types of results could be obtained by IFLWA operator.

61 An Approach to Aggregate Intuitionistic Fuzzy Information …

745

References 1. Adak AK et al (2016) TOPSIS in generalized intuitionistic fuzzy environment. Handb Res Mod Optim Algorithms Appl Eng Econ. IGI Global, pp 630–642 2. Albino V, Garavelli AC (1998) A neural network application to subcontractor rating in construction firms. Int J Proj Manag 16.1: 9–14 3. Assari A, Mahesh TM, Assari E (2012) Role of public participation in sustainability of historical city: usage of TOPSIS method. Indian J Sci Technol 5(3):2289–2294 4. Atanassov KT (1986) Intuitionistic fuzzy sets. Fuzzy Sets Syst 20(1):87–96 5. Atanassov KT (1989) More on intuitionistic fuzzy sets. Fuzzy Sets Syst 33(1):37–45 6. Boran FE et al (2009) A multi-criteria intuitionistic fuzzy group decision making for supplier selection with TOPSIS method. Expert Syst Appl 36(8):11363–11368 7. Dickson G (1966) An analysis of vendor selection and the buying process. J Purch 2(1):517 8. Grzegorzewski P, Mrwka E (2005) Some notes on (Atanassov’s) intuitionistic fuzzy sets. Fuzzy Sets Syst 156(3):492–495 9. Harsanyi JC (1955) Cardinal welfare, individualistic ethics, and interpersonal comparisons of utility. J Polit Econ 63(4):309–321 10. Hong DH, Choi C-H (2000) Multicriteria fuzzy decision-making problems based on vague set theory. Fuzzy Sets Syst 114(1):103–113 11. Hung C-C, Chen L-H (2009) A multiple criteria group decision making model with entropy weight in an intuitionistic fuzzy environment. In: Intelligent automation and computer engineering. Springer, Netherlands, pp 17–26 12. Hwang C-L, Yoon K (1981) Methods for multiple attribute decision making. In: Multiple attribute decision making. Springer, Berlin, pp 58–191 13. Joshi D, Kumar S (2014) Intuitionistic fuzzy entropy and distance measure based TOPSIS method for multi-criteria decision making. Egypt Inform J 15(2):97–104 14. Kaushal M, Solanki R, Lohani QD, Muhuri PK (2018) A Novel intuitionistic fuzzy set generator with application to clustering. In: 2018 IEEE international conference on fuzzy systems (FUZZIEEE). IEEE, pp 1–8 15. Khan MS, Lohani QD (2016) A similarity measure for atanassov intuitionistic fuzzy sets and its application to clustering. In: International workshop on computational intelligence (IWCI). IEEE, pp 232–239 16. Khan MS, Lohani QD (2017) A novel sequence space related to L p defined by Orlicz function with application in pattern recognition. J Inequalities Appl 2017(1):300 17. Khan MS, Lohani QD, Mursaleen M (2017) A novel intuitionistic fuzzy similarity measure based on double sequence by using modulus function with application in pattern recognition. Cogent Math 4(1):1385374 18. Khan MS et al (2017) Sequence spaces M(φ) and N (φ) with application in clustering. J Inequalities Appl 2017(1):63 19. Li M, Jin L, Wang J (2014) A new MCDM method combining QFD with TOPSIS for knowledge management system selection from the user’s perspective in intuitionistic fuzzy environment. Appl Soft Comput 21:28–37 20. Lin L, Yuan X-H, Xia Z-Q (2007) Multicriteria fuzzy decision-making methods based on intuitionistic fuzzy sets. J Comput Syst Sci 73(1):84–88 21. Liu H-W, Wang G-J (2007) Multi-criteria decision-making methods based on intuitionistic fuzzy sets. Eur J Oper Res 179(1):220–233 22. Liu J, Ding F-Y, Lall V (2000) Using data envelopment analysis to compare suppliers for supplier selection and performance improvement. Supply Chain Manag: Int J 5(3):143–150 23. Lohani QD, Solanki R, Muhuri PK (2018) A convergence theorem and an experimental study of intuitionistic fuzzy c-mean algorithm over machine learning dataset. Appl Soft Comput 24. Lohani QD, Solanki R, Muhuri PK (2018) Novel adaptive clustering algorithms based on a probabilistic similarity measure over atanassov intuitionistic fuzzy set. IEEE Trans Fuzzy Syst 25. Mursaleen M, Lohani QD, Mohiuddine SA (2009) Intuitionistic fuzzy 2-metric space and its completion. Chaos, Solitons Fractals 42(2):1258–1265

746

M. Kaushal et al.

26. Sirbiladze G, Badagadze O (2017) Intuitionistic fuzzy probabilistic aggregation operators based on the choquet integral: application in multicriteria decision-making. Int J Inf Technol Decis Mak 16(1):245–279 27. Solanki R et al (2016 A correlation based Intuitionistic fuzzy TOPSIS method on supplier selection problem. In: 2016 IEEE international conference on fuzzy systems (FUZZ-IEEE) 28. Solanki R, Lohani QD, Muhuri PK (2015) A novel clustering algorithm based on a new similarity measure over Intuitionistic fuzzy sets. In: 2015 IEEE international conference on fuzzy systems (FUZZ-IEEE). IEEE, pp 1–8 29. Sugeno M (1993) Fuzzy measures and fuzzy integralsa survey. In: Readings in fuzzy sets for intelligent systems. pp 251–257 30. Tan C, Chen X (2010) Intuitionistic fuzzy Choquet integral operator for multi-criteria decision making. Expert Syst Appl 37(1):149–157 31. Weber CA, Current JR, Benton WC (1991) Vendor selection criteria and methods. Eur J Oper Res 50.1:2–18 32. Wei G (2010) Some induced geometric aggregation operators with intuitionistic fuzzy information and their application to group decision making. Appl Soft Comput 10(2):423–431 33. Wu DD (2009) Supplier selection in a fuzzy group setting: a method using grey related analysis and Dempster-Shafer theory. Expert Syst Appl 36(5):8892–8899 34. Xu Z (2007) Intuitionistic fuzzy aggregation operators. IEEE Trans Fuzzy Syst 15(6):1179– 1187 35. Xu Z (2011) Approaches to multiple attribute group decision making based on intuitionistic fuzzy power aggregation operators. Knowl-Based Syst 24(6):749–760 36. Yager RR (1979) On the measure of fuzziness and negation part I: membership in the unit interval. pp 221–229 37. Yager RR (2001) The power average operator. IEEE Trans Syst, Man, Cybern-Part A 31:724– 731 38. Yang Z et al (2017) Developing dynamic intuitionistic normal fuzzy aggregation operators for multi-attribute decision-making with time sequence preference. Expert Syst Appl 82:344–356 39. Ye J (2017) Intuitionistic fuzzy hybrid arithmetic and geometric aggregation operators for the decision-making of mechanical design schemes. Appl Intell 1–9 40. Zeng S, Xiao Y (2016) TOPSIS method for intuitionistic fuzzy multiple-criteria decision making and its application to investment selection. Kybernetes 45(2):282–296 41. Zhang S-F, Liu S-Y (2011) A GRA-based intuitionistic fuzzy multi-criteria group decision making method for personnel selection. Expert Syst Appl 38(9):11401–11405

Author Index

A Abdul Hamid, Md., 395 Abdur Rahman, Mahmud, 1 Abdur Razzaque, Md., 431 Abir, Tasnim Azad, 371 Abujar, Sheikh, 279, 583 Acharjee, Uzzal Kumar, 85 Afrin, Fahmida, 519 Ahamed, Shakim, 593 Ahmad, Mohiuddin, 11, 111, 151 Ahmad, Nazib, 557 Ahmad, Shadman Fahim, 679 Ahmed, Boshir, 631 Ahmed, Bulbul, 25 Ahmed, Eshtiak, 359 Ahmed, Fahmida, 61 Ahmed, Farruk, 417 Ahmed, Kawsar, 207 Ahmed, Mohiuddin, 163 Ahmed, Nawsher, 617 Ahmed, Nuhash, 557 Akbar, Md. Mostofa, 265 Akhter Hossain, Syed, 583 Alam, Firoj, 407 Alam, Md. Mahbub, 443 Ali Akber Dewan, M., 705, 721 Ali Hossain, Md., 235, 631 Al Imran, Abdullah, 455 Al-Mamun Bulbul, Abdullah, 175 Amirul Islam, Md., 85 Anjum, Anika, 11 Al Noman, Md., 533 Aowlad Hossain, A. B. M., 533 Arefin, Mohammad Shamsul, 705

Asadur Rahman, Md., 11, 533 Asaduzzaman, Md., 395 Asa, Tania Akter, 235 Ashfakur Rahman Arju, Md., 289 Azam, Sami, 443 B Bappy, Akash Shingha, 151 Biswas, Trisa, 383 C Chandra Das, Badhan, 495 Choity, Tasnim Ara, 317 Chowdhury, Linkon, 137 Chowdhury, Muhammad Mahir Hasan, 255 Chowdhury, S. M. Mazharul Hoque, 279 Chowdhury, Tausif Uddin Ahmed, 61 D Das, Sree Krishna, 693 Debnath, Amar, 431 Dey, Samrat Kumar, 483 E Emdadul Islam, Mohammad, 617 F Fahim Newaz, Md., 617 Fahim Sikder, Md., 547 Fardoush, Jannatul, 345 Farin, Nusrat Jahan, 679 Farjana, Nishat, 215 Ferdouse Ahmed Foysal, Md., 583 Furhad, Md. Hasan, 61

© Springer Nature Singapore Pte Ltd. 2020 M. S. Uddin and J. C. Bansal (eds.), Proceedings of International Joint Conference on Computational Intelligence, Algorithms for Intelligent Systems, https://doi.org/10.1007/978-981-13-7564-4

747

748

Author Index

G Ghosh, Priyanka, 279 Gilkar, Gulshan Amin, 1 Goswami, Dhiman, 327 Gupta, Kishor Datta, 49

Kowsher, Md., 519 Kuang, Hongbo, 247

H Habib, Md. Tarek, 417 Hakak, Saqib, 1 Hanif Ali, Mohammad, 197 Haque, Chowdhury Azimul, 383 Haque, Sabrina, 665 Hasan Furhad, Md., 163 Hasan, K. M. Zubair, 471 Hasan, Md. Arid, 407 Hasan, Md. Kamrul, 123, 665 Hasan, Md. Zahid, 289 Hasib, Sultan Abdul, 97 Hassan, Al Maruf, 607 Hayder Ali, Md., 197 Hoq, Md. Nazmul, 265 Hoque, Kazi Ekramul, 289 Hossain, Faria, 317 Hossain, Md. Bellal, 175 Hossain, M. Delowar, 345 Hossain, Md. Mahbub, 643 Hossain, Syed Akhter, 279 Huda, Mohammad Nurul, 359, 519 Huq, Fazlul, 235

M Mahbubur Rahman, Md., 483 Mahi, Md. Julkar Nayeen, 215 Mahmud, Arif, 317, 571 Mahmud, Khizir, 305 Mahmudul Haque, Md., 11 Majumder, Anup, 417 Mamun, Khondaker Abdullah Al, 607 Manjurul Ahsan, Md., 49 Mansoor, Nafees, 679 Marium-E-Jannat, 255 Mashrur Arifin, Md., 643 Masud Rana, Md., 207 Md. Farid, Dewan, 593, 607 Md. Mahbub -Or-Rashid, 25 Meem, Nishat Tasnim Ahmed, 255 Mekayel Anik, Mohammad, 721 Miah, Abdullah Arafat, 111 Mia, Shisir, 227 Mim, Khatuna Zannat, 111 Mim, Mahbina Akter, 39 Mofijul Islam, Md., 431 Mohammad, Rafeed, 455 Moharram, Md. Shakil, 507 Moni, Mohammad Ali, 235 Moon, Nazmun Nessa, 443 Morsalin, Sayidul, 305 Mostafiz, Rafid, 183 Mostasim Billah, Md., 557 Mridha, M. F., 395 Mudi, Ratna, 693 Muntasir Shahriar, Md., 705 Musfique Anwar, Md, 495

I Islam, Islam, Islam, Islam, Islam, Islam,

Ashraful, 359 Md. Mezbahul, 183 Md. Milon, 665 Md. Tobibul, 151 Mohammad Mohaiminul, 73, 265 Monira, 383

J Jannat, Miftahul, 593 Jisan, Arif Rizvi, 289 Johora, Fatema T., 25 Juhin, Faija, 317 K Kabir, Enamul, 247 Kabir, Md. Fasihul, 359 Kamal, Abrar Hasin, 679 Karim, Asif, 443 Kaushal, Meenakshi, 735 Khan, Naib Hossain, 289 Khan, Sakib Shahriar, 593

L Lohani, Q. M. Danish, 735

N Nandi, Rabindra Nath, 417 Nasim Akhtar, Md., 655 Nasrin, Samia, 279 Nazi, Zabir Al, 371 Neehal, Nafis, 507 Nigar, Nishargo, 137 Nipa, Nadira Anjum, 265 Noman, Abu Tayab, 617 Noori, Sheak Rashed Haider, 407, 571 Nur, Fernaz Narin, 443 Nurul Islam, Mohammad, 655

Author Index Nurunnabi Mollah, Md., 11 Nyeem, Hussain, 97 O Ochiuddin Miah, Md., 607 Ontika, Nazmun Nisat, 359 P Parvez, H. M. Shahriar, 1 Patoary, Rayhan, 557 Peng, Weilun, 305 Podder, Etu, 175 Q Quinn, Julian M. W., 235 R Rabeya, Tapasy, 519 Rafi, Md. Erfanul Haque, 665 Rahman, Md. Mijanur, 227 Rahman, Md. Motiur, 183 Rahman, Mohammad Motiur, 227 Rahman, Redoan, 431 Raihan Uddin, Md., 483 Rashedul Islam, Md., 631 Rashid, Humayun, 617 Ravishankar, Jayashri, 305 Rifatul Islam Rifat, Md., 455 Riya, Rokeya Islam, 471 Roy, Shanto, 215 Ruheen Bristi, Warda, 327 S Saad, Abdul Munem, 383 Sadi, Muhammad Sheikh, 665 Saha, Proshib, 443 Saha, Rekha, 643 Saha, Swapnil, 443 Saha, Tumpa Rani, 25 Salah Uddin Yusuf, Md., 383 Sarker, Nur Alam, 345 Sarker, Orvila, 123

749 Sattar, Abdus, 571 Sen, Sajib, 49 Shakirul Islam, Mohammad, 583 Shamsul Arefin, Mohammad, 721 Shatabda, Swakkhar, 593 Shawkat Zamil, K. M., 39 Shekhar Mondal, Himadri, 175, 643 Shoaib Ahmed, Md., 495 Shoaib Khan, Mohd, 735 Shultana, Shahana, 507 Siuly, Siuly, 247 Sultana, Naznin, 73 Sultana, Nishat, 327 Sultana, Tangina, 345 T Tanzila Rahman, Kh., 471 Tareque, Saifuddin M., 289 Tariq Hasan, Md., 643 Taslim Reza, S. M., 617 Tithi, Farhana Sharmin, 519, 557 Toriqul Islam, Md., 557 Tusher, Mohammad Obaidullah, 679 U Uday, Thajid Ibna Rouf, 557 Uddin, Mohammad Shorif, 417 Ullah, Abu S. S. M. Barkat, 163 W Wang, Hua, 247 Whaiduzzaman, Md, 215 Whittaker, Frank, 247 Y Yousuf, Mohammad A., 25 Z Zahan, Nusrat, 471 Zahid Hasan, Md., 471 Zead, Niamul Hasan, 345