Proceedings of 2019 4th International Conference on Information Technology (InCIT) Encompassing Intelligent Technology and Innovation Towards the New Era of Human Life 24 – 25 October, 2019 Thai-Nichi Institute of Technology Bangkok, THAILAND
2019 4th International Conference on Information Technology (InCIT) Copyright © 2019 by the Institute of Electrical and Electronic Engineers, Inc. All rights reserved. Copyright and Reprint Permissions Abstracting is permitted with credit to the source. Libraries are permitted to photocopy beyond the limit of U.S. copyright law for private use of patrons those articles in this volume that carry a code at the bottom of the first page, provided the per-copy fee indicated in the code is paid through Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923. For reprint or republication permission, email to IEEE Copyrights Manager at pubs-permissions@ieee.org. IEEE Catalog Number : CFP19CAH-USB ISBN : 978-1-7281-1018-9 Additional copies of this publication are available from: Curran Associates, Inc. 57 Morehouse Lane Red Hook, NY 12571 USA +1 845 758 0400 +1 845 758 2633 (FAX) Email: curran@proceedings.com
2019 4th International Conference on Information Technology (InCIT2019) Table of Contents Page List of Committees i Conference Chair’s Message v Keynote Speakers vi Conference Program x Oral Presentation Schedule xii List of Full Papers xviii Full Papers 1 List of Reviewers 311
2019 4th International Conference on Information Technology (InCIT2019) List of Committees Honorary Chair Bandhit Rojarayanont President of Thai-Nichi Institute of Technology (TNI), Thailand. General Chair Ruttikorn Varakulsiripunth, TNI, Thailand. Chairman of Council of IT Deans of Thailand (CITT). General Co-chair Burapha University (BUU), Thailand. Krisana Chinnasarn, Phayung Meesad, King Mongkut’s University of Technology North Bangkok (KMUTNB), Thailand. Advisory Board (CITT: alphabet order) Anong Rungsuk, Nakhon Phanom University (NPU), Thailand. Chetneti Srisaan, Rangsit University (RSU), Thailand. Dechanuchit Katanyutaveetip, Siam University (SiamU), Thailand. Kriengkrai Porkaew, King Mongkut’s University of Technology Thonburi (KMUTT), Thailand. Narongdech Keeratipranon, Dhurakij Pundit University (DPU), Thailand. Nopporn Chotikakumtorn, King Mongkut’s Institute of Technology Ladkrabang (KMITL), Thailand. Pattanasak Mongkolwat, Mahidol University (MU), Thailand. Pisit Charnkeitkong, Panyapiwat Institute of Management (PIM), Thailand. Poonpong Boonbrahm, Walailak University (WU), Thailand. Pornthep Rojanavasu, University of Phayao (UP), Thailand. Sasitorn Keawmun, Mahasarakham University (MSU), Thailand. Sinchai Kamolphiwong, Prince of Songkla University (PSU), Thailand. Somsak Chartnamphet, Silpakorn University (SU), Thailand. Teeravisit Laohapensaeng, Mae Fah Luang University (MFU), Thailand. Thana Sukvaree, Sripatum University (SPU), Thailand. Thirapon Wongsaardsakul, Bangkok University (BU), Thailand. Woraphon Lilakiataskun, Mahanakorn University of Technology (MUT), Thailand. i
2019 4th International Conference on Information Technology (InCIT2019) Oversea Advisory Board Chuo University (CHUO-U) / Tohoku University (TOHOKU), Japan. Norio Shiratori, University of Electro-Communications (UEC), Japan. Kenzo Takahashi, Kyoto University (KYOTO-U), Japan. Susumu Yoshida, Hokkaido University (HOKUDAI), Japan. Yoshikazu Miyanaga, University of Sydney (SYDNEY), Australia. Branka Vucetic, National Chiao Tung University (NCTU), Taiwan. Ying-Dar Lin, Yasushi Kato, Sendai Foundation for Applied Information Sciences (SFAIS), Japan. Local Technical Program Supporting Committee Kosin Chamnongthai, King Mongkut’s University of Technology Thonburi (KMUTT), Thailand. Waree Kongprawechnon, Sirindhorn International Institute of Technology (SIIT), Thailand. Virach Sornlertlamvanich, Sirindhorn International Institute of Technology (SIIT), Thailand. Kuntpong Woraratpanya, King Mongkut’s Institute of Technology Ladkrabang (KMITL), Thailand. Chotipat Pornavalai, King Mongkut’s Institute of Technology Ladkrabang (KMITL), Thailand. Noppadol Maneerat, King Mongkut’s Institute of Technology Ladkrabang (KMITL), Thailand. Sakchai Tipchaksurat, King Mongkut’s Institute of Technology Ladkrabang (KMITL), Thailand. Saiyan Saiyod, Khon Kaen University (KKU), Thailand. Wimol San-Um, Thai-Nichi Institute of Technology (TNI), Thailand. Putchong Uthayopas, Kasetsart University (KU), Thailand. Chaodit Aswakul, Chulalongkorn University (CU), Thailand. International Technical Program Supporting Committee Tokai University (TOKAI-U), Japan. Kazuhiko Hamamoto, Samuel S. Chua, Lyceum of the Philippines University (LPU), Philippines. I Wayan Mustika, Universitas Gadjah Mada (UGM), Indonesia. Hitoshi Isahara, Toyohashi University of Technology (TUT), Japan. Kaoru Takahashi, National Institute of Technology, Asahikawa College, Japan. Salahuddin Muhammad Salim Zabir, National Institute of Technology, Tsuruoka College, Japan. Takashi Mitsuishi, Tohoku University (TOHOKU), Japan. Hiroshi Tsunoda, Tohoku Institute of Technology (TOHTECH), Japan. ii
2019 4th International Conference on Information Technology (InCIT2019) Organizing Committee Thai-Nichi Institute of Technology (TNI), Thailand. Annop Monsakul, Ekarat Rattagan, Mahanakorn University of Technology (MUT), Thailand. Datchakorn Tancharoen, Panyapiwat Institute of Management (PIM), Thailand. Duangjai Jitkongchuen, Dhurakij Pundit University (DPU), Thailand. Kanyarat Sriwisathiyakun, Sripatum University (SPU), Thailand. Lapas Pradittasnee, King Mongkut’s Institute of Technology Ladkrabang (KMITL), Thailand. Nittaya Kerdyam, Siam University (SiamU), Thailand. Nathaporn Karnjanapoomi, Silpakorn University (SU), Thailand. Nutchanat Buasri, Mahasarakham University (MSU), Thailand. Pakapan Limtrairut, Bangkok University (BU), Thailand. Parkpoom Chaisiriprasert, Rangsit University (RSU), Thailand. Pawitra Chiravirakul, Mahidol University (MU), Thailand. Prajaks Jitngernmadan, Burapha University (BUU), Thailand. Pruegsa Duangphasuk, Mahanakorn University of Technology (MUT), Thailand. Suwitchaya Rattarom, University of Phayao (UP), Thailand. Rattana Wetprasit, Prince of Songkla University (PSU), Thailand. Rawiworn Hongma, Nakhon Phanom University (NPU), Thailand. Suppat Rungraungsilp, Walailak University (WU), Thailand. Suree Funilkul, King Mongkut’s University of Technology Thonburi (KMUTT), Thailand. Watchareewan Jitsakul, King Mongkut’s University of Technology North Bangkok (KMUTNB), Thailand. Worasak Rueangsirarak, Mae Fah Luang University (MFU), Thailand. Local Arrangement Committee Thai-Nichi Institute of Technology (TNI), Thailand. Adisak Suasaming, Thai-Nichi Institute of Technology (TNI), Thailand. Saromporn Charoenpit, Thai-Nichi Institute of Technology (TNI), Thailand. Apichaya Nimkoompai, Thai-Nichi Institute of Technology (TNI), Thailand. Kanda Tiwatthanont, Thai-Nichi Institute of Technology (TNI), Thailand. Patsama Charoenpong, Thai-Nichi Institute of Technology (TNI), Thailand. Amonpan Chomklin, Thai-Nichi Institute of Technology (TNI), Thailand. Nitirat Tanthavech, Thai-Nichi Institute of Technology (TNI), Thailand. Chan Jaruwongrungsee, iii
2019 4th International Conference on Information Technology (InCIT2019) Prajak Chertchom, Thai-Nichi Institute of Technology (TNI), Thailand. Paskorn Apirukvorapinit, Thai-Nichi Institute of Technology (TNI), Thailand. Pramuk Boonsieng, Thai-Nichi Institute of Technology (TNI), Thailand. Sared Wansopa, Thai-Nichi Institute of Technology (TNI), Thailand. Lalita Na Nongkhai, Thai-Nichi Institute of Technology (TNI), Thailand. Sarayut Nonsiri, Thai-Nichi Institute of Technology (TNI), Thailand. Ferdin Joe John Joseph, Thai-Nichi Institute of Technology (TNI), Thailand. Kasem Thiptarajan, Thai-Nichi Institute of Technology (TNI), Thailand. Kanakarn Ruxpaitoon, Thai-Nichi Institute of Technology (TNI), Thailand. Narungsun Wilaisakoolyoug, Thai-Nichi Institute of Technology (TNI), Thailand. Nattagit Jiteurtragool, Thai-Nichi Institute of Technology (TNI), Thailand. Saprangsit Mruetusatorn, Thai-Nichi Institute of Technology (TNI), Thailand. Pranisa Isarasena, Thai-Nichi Institute of Technology (TNI), Thailand. Triratana Metkarunchit, Thai-Nichi Institute of Technology (TNI), Thailand. Thitiporn Lertrusdachakul, Thai-Nichi Institute of Technology (TNI), Thailand. Salinla Chevakidagarn, Thai-Nichi Institute of Technology (TNI), Thailand. Nuchanart Pongpanich, Thai-Nichi Institute of Technology (TNI), Thailand. Tanyaporn Kanignant, Thai-Nichi Institute of Technology (TNI), Thailand. Budsaraphorn Luangmalawat, Thai-Nichi Institute of Technology (TNI), Thailand. Cholrit Luangjinda, Thai-Nichi Institute of Technology (TNI), Thailand. Puwadol Sirikongtham, Thai-Nichi Institute of Technology (TNI), Thailand. Oran Roenshoen, Thai-Nichi Institute of Technology (TNI), Thailand. Nichakan Chaiyajak, Thai-Nichi Institute of Technology (TNI), Thailand. Pakachart Puttipakorn, Thai-Nichi Institute of Technology (TNI), Thailand. Rungphop Preechawit, Thai-Nichi Institute of Technology (TNI), Thailand. Athitaya Thaiyong, Thai-Nichi Institute of Technology (TNI), Thailand. Patchon Sangarun, Thai-Nichi Institute of Technology (TNI), Thailand. iv
2019 4th International Conference on Information Technology (InCIT2019) CONFERENCE CHAIR’S MESSAGE Dr.Ruttikorn Varakulsiripunth Dean, Faculty of Information Technology Thai-Nichi Institute of Technology (TNI) & Chairman of Council of IT Deans of Thailand (CITT) Nowadays, various technologies are researched and developed in order to fulfill the needs of human society. Faced with these new major challenges, it is necessary to rethink and encompassing intelligent technology and innovation towards the New Era of Human Life. There are several important functions and issues to be considered and discussed among experts and researchers. Therefore, 2019 the 4th International Conference on Information Technology (InCIT2019) provides an opportunity for faculty members, researchers and students who have involved with above-mentioned areas to get together for sharing their achievement. The participants will discuss and exchange opinions to realize smart technologies for next generation of innovations regarding to informatics, and related areas. Moreover, InCIT2019 is also continuously approved by IEEE Thailand as well as IEEE head organization for Technical Support as co-sponsor and the conference papers will be included into IEEE Xplore Digital Library. As same as former InCIT2017 and InCIT2018 that be indexed by SCOPUS, InCIT2019 will be indexed by SCOPUS too. It is pleased to have 89 research papers submitted to InCIT2019 from researchers of many countries. All papers were reviewed and evaluated by three experts/paper in related area, and finally 69 papers were accepted for presentation. However, due to some problems, there are 10 accepted papers were withdrawn by authors, finally there are 59 papers will be presented in InCIT2019. It would be my great pleasure that all the participants could have fruitful discussion and get insights for further research. Finally, I would like to thank all authors for your academic contribution to InCIT2019. I also would like to express my sincere gratitude to all local and oversea committee for supporting, and also to all faculty members who worked hard to the success of InCIT2019. v
2019 4th International Conference on Information Technology (InCIT2019) Keynote Speakers 5G Mobile Edge Computing: Research Roadmap of the H2020 5G-Coral Project Prof.Dr.Ying-Dar Lin, IEEE Fellow, IEEE Distinguished Lecturer Editor-in-Chief, IEEE Communications Surveys and Tutorials Distinguished Professor, National Chiao Tung University, Hsinchu, TAIWAN Web page: www.cs.nctu.edu.tw/~ydlin Abstract: 5G promises to deliver enhanced mobile broadband (eMBB), massive machine type communication (mMTC), and ultra reliable low latency communication (URLLC). To support mMTC and URLLC, 5G needs to carry out computations closer to subscribers at the “edge” instead of the cloud, which turns 5G into an infrastructure for both communication and computing. Just like cloud computing, edge computing shall also be virtualized. On the other hand, communication is also being virtualized with software defined networking (SDN) and network function virtualization (NFV) which virtualize control plane and data plane, respectively. When applied to 5G, together they virtualize functions in access and core networks, and release them to run on any virtualized computing platform. Combining virtualization needs in edge computing and communication, 5G mobile edge computing (MEC) is virtualizing eNB (evolved node B), EPC (evolved packet core), and CO (central office) into VeNB, vEPC, and CORD (central office re-architected as a datacenter). They are not just communication devices anymore, but also serve as computing datacenters with many open source resources like OpenDaylight and OpenStack. After streamline the above evolution path, we then introduce 5G-Coral, an H2020 EU-TW project with Taiwanese and European partners, including NCTU, ITRI, ADLink, UC3M, Ericsson, InterDigital, Telecom Italia, SICS, Telcaria, and Azcom. We then give an overview of our research roadmap on 5 key components, including service chain routing, multi-RAT offloading, multi-tenant slicing, horizontal and vertical federation, and capacity optimization. Selected results are then presented. Key findings include (1) the 3-tier architecture with edge computing saves about 20.7% capacity cost over the traditional 2-tier architecture, with 70% of capacity allocated to the edges; (2) multi-RAT offloading reduces about 40% capacity cost with a large number of UEs; (3) some use cases in 5G would capture 1.5 to 2.3 times more resource than required if without slicing; (4) the low-latency authentication with MEC reduces over 90% overhead if done with the cloud. vi
2019 4th International Conference on Information Technology (InCIT2019) Autobiography: YING-DAR LIN is a Distinguished Professor of computer science at National Chiao Tung University (NCTU), Taiwan. He received his Ph.D. in computer science from the University of California at Los Angeles (UCLA) in 1993. He was a visiting scholar at Cisco Systems, San Jose, during 2007–2008, CEO at Telecom Technology Center, Taiwan, during 2010-2011, and Vice President of National Applied Research Labs (NARLabs), Taiwan, during 2017-2018. Since 2002, he has been the founder and director of Network Benchmarking Lab (NBL, www.nbl.org.tw), which reviews network products with real traffic and has been an approved test lab of the Open Networking Foundation (ONF) since July 2014. He also cofounded L7 Networks Inc. in 2002, later acquired by D-Link Corp, and O’Prueba Inc. in 2018. His research interests include network security, wireless communications, and network softwarization. His work on multi-hop cellular was the first along this line, and has been cited over 850 times and standardized into IEEE 802.11s, IEEE 802.15.5, IEEE 802.16j, and 3GPP LTE-Advanced. He is an IEEE Fellow (class of 2013), IEEE Distinguished Lecturer (2014–2017), ONF Research Associate, and received in 2017 Research Excellence Award and K. T. Li Breakthrough Award. He has served or is serving on the editorial boards of several IEEE journals and magazines, and is the Editor-in-Chief of IEEE Communications Surveys and Tutorials (COMST). He published a textbook, Computer Networks: An Open Source Approach (www.mhhe.com/lin), with Ren-Hung Hwang and Fred Baker (McGraw-Hill, 2011). vii
2019 4th International Conference on Information Technology (InCIT2019) Natural Language Processing Research in Thai Context — A 29-Year Journey of Thai NLP — Prof.Dr. Virach Sornlertlamvanich School of Information, Computer and Communication Technology Sirindhorn International Institute of Technology (SIIT), Thammasat University, Thailand and Professor of Musashino University, Japan Chair of Digital Cluster, Research University Network (RUN) virach@siit.tu.ac.th, virach@gmail.com Abstract: The talk introduces a brief history of Thai NLP:- about how it began, developed and evolved in the research community. Some of the Thai language particular issues will be discussed together with some approaches which have been introduced in the past with good success and failure. Beginning with the rule-based approaches, many kinds of dictionaries and grammar rules have been designed and proposed with their coverage to fulfill the Thai language depending characteristics such as no explicit word/sentence boundary, no inflection, no grammatical markers, etc. Thousands of CFG rules have been created but still allow exceptional cases which are left unexplainable. In the earliest time, we realized that a large enough corpus is needed to capture such flexible language phenomena. We still do not know how large the corpus we need, but we do need. Based on the long time research experience of the community, the language resources such as computational dictionaries, ontology, POS tagged corpora, treebank, named entity corpora, speech corpora for TTS and ASR research, have been accumulated. Statistical and probabilistic language models have been studied since then. Hidden Markov model and Bayesian networks are widely used in word/sentence segmentation tasks. Probabilistic GLR parser is also reported to outperform the results from PCFG, where both can capture the sequential (word order) and structural (syntactic) contexts. Thanks to the tremendous efforts in designing and collecting the annotated corpora, the approaches in the scheme of machine learning, the larger the corpus is the better the model can predict. Today, with the main stream of deep learning techniques, many NLP problems are being re-challenged and the results are significantly improved. Some NLP applications will also be re-introduced to understand its potential in applying to many other research fields. viii
2019 4th International Conference on Information Technology (InCIT2019) Autobiography: In 2003, he achieved the “National Distinguished Researcher Award” in Information Technology and Communication from the National Research Council of Thailand, following by the “ASEAN Outstanding Engineering Achievement Award” from ASEAN Federation of Engineering Organizations (AFEO) in 2011. He was also esteemed “The Researcher of the Year 2001” by the Nation Newspaper in 2001. He started his research career in the field of Knowledge Engineering and Artificial Intelligence during his study in Kyoto University in 1980-1986. He started his research in Natural Language Processing by participating in the Multi-lingual Machine Translation project during 1988-1995, and received his Ph.D. from Tokyo Institute of Technology in 1998. Some of his long-running research contributions can be seen in the initiative in the development of Thai POS tagged corpus (ORCHID, 1997), the first corpus based Thai-English dictionary (LEXiTRON, 1997), and the first English-Thai online machine translation web service (ParSit, 2000) based on the Inter-lingual approach. His recent efforts are on the research and development of the technologies for digital content creation and understanding. He proposed the Digitized Thailand project in 2009 to establish an intelligent service platform for being a fundamental framework for digital content sharing and application mashup. Some of the achievements have already been publicized in culture and local wisdom digitization and the applications on the digital content services for tourism, product design and education. His research interest includes Natural Language Processing, Human Language Technology, Information Retrieval, Data Mining, Artificial Intelligence, Machine Learning, Deep Learning, Social Media Analytics and the related fields. ix
2019 4th International Conference on Information Technology (InCIT2019) Conference (InCIT2019) Program 24th October 2019 (Day 1) Thai 08.00-08.45 Registration at Convention Hall 6th floor in building E 08.45-09.15 Opening Ceremony in Convention Hall -A 09.15-10.15 Keynote Speech Prof. Dr. Ying-Dar Lin “5G Mobile Edge Computing: Research Roadmap of the H2020 5G-Coral Project” 10.15-10.30 coffee break at Convention Hall 6th floor in building E 10.30-11.30 Keynote Speech Prof. Dr. Virach Sornlertlamvanich “Natural Language Processing Research in Thai Context — A 29-Year Journey of NLP —” 11.30-12.40 Lunch Break at Convention Hall-B 6th floor in building E 12.40-14.50 Oral Presentation (at 4th floor in building E) 14.50-15.20 coffee break at 4th floor in building E 15.20-17.30 Oral Presentation (at 4th floor in building E) 18.00–21.00 Banquet (dinner and shows) at Convention Hall-A, B 6th floor in building E x
2019 4th International Conference on Information Technology (InCIT2019) 25th October 2019 (Day 2) 08.00-08.30 Registration at Convention Hall 6th floor in building E 08.30-10.30 Oral Presentation (at 4th floor in building E) 10.30-10.50 coffee break at 4th floor in building E 10.50-12.30 Oral Presentation (at 4th floor in building E) 12.30-14.30 Closing Ceremony and Lunch (best paper awarding) at Convention Hall-A, B 6th floor in building E xi
2019 4th International Conference on Information Technology (InCIT2019) ORAL PRESENTATION SCHEDULE for InCIT2019 4th Floor, Building E 24th October 2019 (12.40 – 14.50) Room E404: Image Processing-1 Session Chair: Asst. Prof. Dr. Thitiporn Lertrusdachakul ID Title Presenter 1570560058 Segmentation of Shinbone Interosseous Space Using GVF Techniques Siwakorn Artraksa 1570560059 Information of Sulci Vector for Classifying Hydrocephalus and Cerebral Atrophy Onsiri Singkorn Symptom 1570565263 MicroRNA-Gene Signatures Prediction for Cancers with Drug Discovery Prathan Phumphuang 1570565658 RGB-D Depth Inpainting with Color Guide Inverse Distance Weight Yossawee Kaeomanee 1570559445 Comparison of 3D Point Cloud Processing and CNN Prediction Based on RGBD Ananya Kuasakunrungroj Images for Bionic-eye's Navigation Room E405: Machine Learning Session Chair: Dr. Paskorn Apirukvorapinit ID Title Presenter 1570561042 Chaiyapat Sirisin 1570563544 i-Sleep: Intelligent Sleep Detection System for Analyzing Sleep Behavior Karn Yongsiriwit 1570564993 Surveillance System for Abnormal Driving Behavior Detection Worawit Saetan Power Allocation for Sum Rate Maximization in 5G NOMA System with 1570565698 Imperfect SIC: A Deep Learning Approach Supawadee Srikamdee Forecasting Daily Air Quality in Northern Thailand Using Machine Learning 1570565750 Techniques Katika Kongsil Physical Activity Recognition Using Streaming Data from Wrist-worn Sensors xii
2019 4th International Conference on Information Technology (InCIT2019) Room E406: Speech Recognition Session Chair: Asst.Prof.Dr. Kuntpong Woraratpanya ID Title Presenter 1570558883 Yeh Huann Goh 1570562643 Audio-Visual Speech Recognition System Using Recurrent Neural Network Suphanut Thattinaphanich Thai Named Entity Recognition Using Bi-LSTM-CRF with Word and Character 1570565488 Representation Chayanin Tongphasook 1570565595 Two Recognition Models for Thai Dancing Data Set Nattaporn Triemvitaya Sound Tooth: Mobile Oral Health Exam Recording Using Individual Voice 1570566730 Recognition Nitis Monburinon A Novel Hierarchical Edge Computing Solution Based on Deep Learning for Distributed Image Recognition in IoT System 24th October 2019 (15.20 – 17.30) Room E404: Neural Network Session Chair: Dr.Ferdin Joe John Joseph ID Title Presenter 1570563547 Kietikul Jearanaitanakij Predicting Short Trend of Stocks by Using Convolutional Neural Network and 1570556151 Candlestick Patterns Sally Goldin 1570560301 Sugar Cane Grading from Photos Using Convolutional Neural Networks John C. Valdoria iDahon: An Android Based Terrestrial Plant Disease Detection Mobile 1570564814 Application Through Digital Image Processing Using Deep Learning Neural Sukrit Jaidee Network Algorithm 1570558894 Very Short-Term Solar Power Forecasting Using Genetic-Algorithm-Based Deep Sukrit Jaidee Neural Network Very-Short Term Solar Power Forecast Using Data from NWP Model xiii
2019 4th International Conference on Information Technology (InCIT2019) Room E405: Image Processing-2 Session Chair: Dr. Saprangsit Mruetusatorn ID Title Presenter 1570565132 Vinh Truong Hoang Dimensionality Reduction Based on Feature Selection for Rice Varieties 1570565246 Recognition Vinh Truong Hoang Data Augmentation Based on Color Features for Limited Training Texture 1570560060 Classification Saowalak Thamnawat Region of Interest Identification on Low-Resolution Lateral Spine Radiography 1570560062 Image Using Density-based and Ellipse-like Method Jiraporn Wongwarn Tunica Media Localization in IVUS Image with Shadow Artifact Constraints Using Circular-like Estimating Techniques Room E406: E-education-1 Session Chair: Dr. Sarayut Nonsiri ID Title Presenter 1570560732 Yuichi Ohkawa Development and Evaluation of Smartphone Learning Material for Blended 1570560917 Language Learning Samuel S. Chua Online Examination System with Cheating Prevention Using Question Bank 1570560923 Randomization and Tab Locking Chanakarn Phandan Design and Evaluation of Interactive Learning Story and User Interface 1570561017 Prototyping for Mobile Responsive Learning Application Konomu Dobashi A Heat Map Generation to Visualize Engagement in Classes Using Moodle 1570565709 Learning Logs Kuniaki Yajima Detection of Concentration State Using Biosignals xiv
2019 4th International Conference on Information Technology (InCIT2019) 25th October 2019 (8.30 – 10.30) Room E404: AI Session Chair: Dr.Suppakarn Chansareewittaya ID Title Presenter 1570559414 Ferdin Joe John Joseph Twitter Based Outcome Predictions of 2019 Indian General Elections Using 1570539912 Decision Tree Yonten Jamtsho Bhutanese License Plate Recognition Using Hu's Moments and Centroid 1570565294 Difference Pongsak Thuankhonrak 1570565431 Machine Trading by Time Series Models and Portfolio Optimization Pannawat Thanapirompokin Applied Genetic Algorithm per Environment Zone to Solve Problem of Crops 1570566739 Selection for Intercropping by Modified Parameter of Fitness Function Krisana Kotprom Extracting Components from Thai-Official Documents Using Image Processing and Machine Learning Techniques Room E405: IT Medical System Session Chair: Acting Sub-Lt. Dr.Charoenchai Wongwatkit ID Title Presenter 1570556446 Sakada Sao 1570560264 Long Short-Term Memory for Bed Position Classification YB Dwi Setianto Medical Device Authentication and Authorization Protocol in Indonesian 1570560573 Telemedicine Systems Toshihiro Kita 1570560878 Implementation of Voice User Interfaces to Enhance Users' Activities on Moodle Thitinan Tantidham 1570566377 Rehabilitation Exercise Prescription on Android System Automatic Celebrity Weight Estimation Jian Qu xv
2019 4th International Conference on Information Technology (InCIT2019) Room E406: Software Application-1 Session Chair: Asst.Prof.Dr. Lapas Pradittasnee ID Title Presenter 1570552780 Application of Large Neighborhood Search and Differential Evolution for Nuttachat Wisittipanit Solving Vehicle Routing Problem in Post Office Delivery for the Post Office 1570558127 Chiang Rai Branch Rinaldi Munir 1570560566 A Secure Fragile Video Watermarking Algorithm for Content Authentication Pongpat Rakdej 1570565591 Based on Arnold Cat Map Kunihiko Sakurai Coin Recovery from Inaccessible Cryptocurrency Wallet Using Unspent Transaction Output An Evaluation of Virtual Machine Placement by Using Live Migration 25th October 2019 (10.50 – 12.30) Room E403: LAN Session Chair: Dr. Pranisa Israsena ID Title Presenter 1570564589 A Novel Combination Method of Beamforming and OSTBC for MIMO Paeiz Azmi Broadcast System 1570564727 Zhiqing Zhang 1570565229 Unveiling Malicious Activities in LAN with Honeypot Yuwei Sun 1570565676 Detection and Classification of Network Events in LAN Using CNN Mio Kobayashi Development of Water Temperature Measuring Application Based on LoRa/LoRWAN Room E404: E-education-2 Session Chair: Dr. Prajaks Jitngernmadan ID Title Presenter 1570564704 Class Attendance Recording via QR Code Amonpan Chomklin 1570564946 A Proposal of a Students' Voluntary Growth System of Generic Skills by Kuniaki Yajima 1570565197 Objective Evaluation Methods Shin-nosuke Suzuki 1570565747 Examination of A-txt System Independent from OSs After Developed the iOS and Android Version Yunarso Anang Implementation of Computer-Based Test in a Countrywide New Student Recruitment Process xvi
2019 4th International Conference on Information Technology (InCIT2019) Room E405: E-commerce Session Chair: Asst.Prof.Dr. Datchakorn Tancharoen ID Title Presenter 1570554607 Anonymity Supporting Tool for Community of Inquiry-Based Platform in Erni Juraida Encouraging People to Share Knowledge 1570556288 Recommender System Based on User Evaluations and Cosmetic Ingredients Yoko Nakajima 1570560735 E-Commerce for the Preservation of Traditional Thai Craftsmanship Prisana Mutchima 1570566521 Key Factors of Usability of Science and Technology Faculties' Website: Prawit Boonmee Marketing Purpose Room E406: Software Application-2 Session Chair: Dr.Pakapan Limtrairut ID Title Presenter 1570560002 Disaster Risk Management Training Simulation for People with Hearing Arlene R. Caballero Impairment A Design and Implementation of ASL Assisted Model Using Virtual 1570563746 Reality Suppakarn Chansareewittaya Improved TTC per Fuel Cost with DGs by Using Evolutionary Programming Wantana Sisomboon Apisit Saengsai 1570565300 Engage and Motivate Developers for Scrum Adopting with Gamification 1570565765 AppDOSI: An Application for Analyzing and Monitoring the Personal Software Process xvii
2019 4th International Conference on Information Technology (InCIT2019) List of Full Papers) ID Manuscript Page 1570539912 Bhutanese License Plate Recognition Using Hu’s Moments and Centroid Difference 1 1570552780 7 Application of Large Neighborhood Search and Differential Evolution for Solving Vehicle 1570554607 Routing Problem in Post Office Delivery for the post office, Chiang Rai branch 12 1570556151 Anonymity Supporting Tool for Community of Inquiry-Based Platform in Encouraging 18 1570556288 People to Share Knowledge: A Case Study of Pusilkom Universitas Indonesia 22 1570556446 28 1570558127 Sugar Cane Grading from Photos Using Convolutional Neural Networks 32 1570558883 Recommender System Based on User Evaluations and Cosmetic Ingredients 38 1570558894 44 1570559414 Long Short-Term Memory for Bed Position Classification 50 1570559445 54 A Secure Fragile Video Watermarking Algorithm for Content Authentication Based on 1570560002 Arnold Cat Map 60 1570560058 Audio-Visual Speech Recognition System Using Recurrent Neural Network 65 1570560059 71 1570560060 Very Short-Term Solar Power Forecast using Data from NWP Model 77 1570560062 Twitter Based Outcome Predictions of 2019 Indian General Elections Using Decision Tree 83 1570560264 Comparison of 3D Point Cloud Processing and CNN Prediction Based on RGBD Images 89 for Bionic-eye’s Navigation 1570560301 94 Disaster Risk Management Training Simulation for People with Hearing Impairment: A 1570560566 Design and Implementation of ASL Assisted Model Using Virtual Reality 99 1570560573 Segmentation of Shinbone Interosseous Space using GVF Techniques 104 Information of Sulci Vector for classifying Hydrocephalus and Cerebral Atrophy Symptom Region of Interest Identification on Low-Resolution Lateral Spine Radiography Image using Density-based and Ellipse-like Method Tunica Media Localization in Intravascular Image with Shadow Artifact Constraint using Circular-like Estimating Techniques Medical Device Authentication and Authorization Protocol in Indonesian Telemedicine Systems iDahon: An Android Based Terrestrial Plant Disease Detection Mobile Application Through Digital Image Processing Using Deep Learning Neural Network Algorithm Coin Recovery from Inaccessible Cryptocurrency Wallet Using Unspent Transaction Output Implementation of Voice User Interfaces to Enhance Users’ Activities on Moodle xviii
2019 4th International Conference on Information Technology (InCIT2019) ID Manuscript Page 1570560732 Development and Evaluation of Smartphone Learning Material for Blended Language 108 Learning 1570560735 114 1570560878 E-Commerce for the Preservation of Traditional Thai Craftsmanship 120 1570560917 126 Rehabilitation Exercise Prescription on Android System 1570560923 132 Online Examination System with Cheating Prevention Using Question Bank 1570561017 Randomization and Tab Locking 138 1570561042 144 1570562643 Design and Evaluation of Interactive Learning Story and User Interface Prototyping for 149 Mobile Responsive Learning Application 1570563544 155 1570563547 A Heat Map Generation to Visualize Engagement in Classes Using Moodle Learning Logs 159 1570563746 i-Sleep: Intelligent Sleep Detection System for Analyzing Sleep Behavior 163 1570564589 168 1570564704 Thai Named Entity Recognition Using Bi-LSTM-CRF with Word and Character 173 1570564727 Representation 179 1570564814 184 Surveillance System for Abnormal Driving Behavior Detection 1570564946 190 Predicting Short Trend of Stocks by Using Convolutional Neural Network and Candlestick 1570564993 Patterns 195 1570565132 Improved TTC per Fuel Cost with DGs by Using Evolutionary Programming 199 1570565197 203 A Novel Combination Method of Beamforming and OSTBC for MIMO Broadcast System 1570565229 208 1570565246 Class Attendance Recording using QR Code via Smartphone 213 1570565263 217 1570565294 Unveiling Malicious Activities in LAN with Honeypot 222 1570565300 228 Very Short-Term Solar Power Forecasting Using Genetic Algorithm Based Deep Neural Network A Proposal of a Students’ Voluntary Growth System of Generic Skills by Objective Evaluation Methods Power Allocation for Sum Rate Maximization in 5G NOMA System with Imperfect SIC : A Deep Learning Approach Dimensionality Reduction Based on Feature Selection for Rice Varieties Recognition Examination of A-txt System Independent from OSs After Developed the iOS and Android Version Detection and Classification of Network Events in LAN Using CNN Data Augmentation Based on Color Features for Limited Training Texture Classification MicroRNA-Gene Signatures Prediction for cancers with Drug Discovery Machine Trading by Time Series Models and Portfolio Optimization Engaging and Motivating Developers by Adopting Scrum Utilizing Gamification xix
2019 4th International Conference on Information Technology (InCIT2019) ID Manuscript Page 1570565431 Applied Genetic Algorithm per environment zone to solve problem of crops selection for 233 intercropping by modified parameter of fitness function 1570565488 238 1570565591 Two Recognition Models for Thai Dancing Data Set 243 1570565595 248 1570565658 An Evaluation of Virtual Machine Placement by Using Live Migration 254 1570565676 259 1570565698 Sound Tooth: Mobile Oral Health Exam Recording Using Individual Voice Recognition 264 1570565709 269 1570565747 RGB-D Depth Inpainting with Color Guide Inverse Distance Weight 273 1570565750 Development of Water Temperature Measuring Application Based on LoRa/LoRWAN 279 1570565765 285 1570566377 Forecasting Daily Air Quality in Northern Thailand Using Machine Learning Techniques 289 1570566521 294 Detection of Concentration State Using Biosignals 1570566730 299 Implementation of Computer-Based Test in a Countrywide New Student Recruitment 1570566739 Process 305 Physical Activity Recognition Using Streaming Data from Wrist-worn Sensors AppDOSI: An Application for Analyzing and Monitoring the Personal Software Process Automatic Celebrity Weight Estimation Key Factors of Usability of Science and Technology Faculties’ Website: Marketing Purpose A Novel Hierarchical Edge Computing Solution Based on Deep Learning for Distributed Image Recognition in IoT Systems Extracting Components from Thai-Official Documents using Image Processing and Machine Learning Techniques xx
2019 4th International Conference on Information Technology (InCIT2019) Bhutanese License Plate Recognition Using Hu’s Moments and Centroid Difference Yonten Jamtsho Panomkhawn Riyamongkol Rattapoom Waranusast Department of Electrical and Computer Department of Electrical and Computer Department of Electrical and Computer Engineering Engineering Engineering Naresuan University Naresuan University Naresuan University Phitsanulok, Thailand Phitsanulok, Thailand Phitsanulok, Thailand yontenj61@email.nu.ac.th panomkhawnr@nu.ac.th rattapoomw@nu.ac.th Abstract— The purpose of this study was to develop an Fig. 1. Number plate formats in Bhutan automatic number plate recognition system for Bhutanese license plates. This paper presents a license plate localization The typical ANPR has three stages: license plate based on state-of-the-art YOLO (You-Only-Look-Once) object localization, character segmentation and character detector. Once the region of interest was extracted, the recognition. Each step is vital to have high accuracy in the connected component analysis method was proposed with some recognition rate. Failing to detect license plate will lead to additional method to segment the characters and handle failure in the subsequent steps [2]. Due to advancement in the different types of license plates. The features from the field of computer vision, many new approaches were segmented characters were extracted using Hu’s moments, and introduced to reduce the processing time and eliminate false the Centroid Difference (CD) method was proposed to handle positives [2]. characters having the same geometrical shape. The extracted features were trained and tested on WEKA using random forest The rest of the paper is organized as follows: Section II classifier algorithm and the obtained accuracy was 94.6%. discusses the related work and Section III explains the proposed algorithm. In Section IV, the experimental results Keywords— ANPR, Hu’s moments, centroid difference, are presented, followed by conclusions in Section V. random forest classifier, WEKA II. RELATED WORK I. INTRODUCTION A. License Plate Localization With the rapid development of human industrialization and urbanization, the number of vehicles on the roads has Since most of the license plates have different background multiplied drastically all over the world. Similarly, in Bhutan, and foreground color, the color information is used to extract with the increasing number of cars, it has become crucial to the license plates from the image. In [3], the input pixels of implement the Automatic Number Plate Recognition (ANPR) the image are classified using the HLS model into 13 technology to automate the traditional way of monitoring the categories. The HLS model is then used to locate the plate traffic which might help to reduce the human resources and region, and the verification is done by comparing its width to cost. ANPR is a mass surveillance method of reading vehicle height ratio. Ashtari et al. [4] suggest a method using the color license plates for recognition without the use of human features to locate the Persian license plate by defining a resources. ANPR technology has a wide range of benefits salient and standard template. The blue color rectangle such as automatic toll fee collection, traffic law reinforcement, parking space management and private spaces access control [1]. In Bhutan, there are eight types of license plates with different size and color, but the study focuses on three types of license plates as they are used widely (see Fig. 1). The license plate number starts with two alphabets, which is the registration code (BP-Bhutan Private, BT-Bhutan Taxi, BG-Bhutan Government) followed by region code and finally end with four numerals. The four numerals are prefixed by alphabet if the numbers are exhausted. These license plates either have a red background with a white or yellow foreground or yellow background with a black character. So, to meet the requirement of the license plate format, this study proposes the latest state-of-the-art method for plate localization; a combination of connected component analysis and white pixel counting with inversion method for character segmentation; feature extraction using Hu Moments and CD method for character recognition. 1
2019 4th International Conference on Information Technology (InCIT2019) situated on the side of the Persian license plate is selected as Fig. 2. Overview of the proposed system a suitable feature. Apart from the color information, the boundary information is another method that detects a recognition, the features from the individual character are rectangular shape with a known aspect ratio. The plate area extracted using the Hu’s moments [14] with Centroid contains rich pixel densities due to the presence of several Difference and then passed to the random forest classifier for characters [5]. In [5-6], the Sobel Vertical operator is classification. proposed to detect the vertical edges of the license plate region because the plate region contains more vertical edges B. Localization due to the high contrast between the alphabets and the plate background. The accurate license plate localization directly affects the efficiency of the segmentation and the recognition phase. The In the last few years, deep learning technique is used problem of license plate localization is the same as the object widely in the field of computer vision. Yang & Chen detection problem. The traditional approaches like edge proposed a robust method for license plate localization using information, sliding window technique and morphological convolutional neural network [7]. The accuracy obtained operation show good results when the license plates are in from their study was 99.2%. good quality. Due to the variation of the license plate from B. Character Segmentation The detected license plate is then segmented to extract the individual characters. In some countries, they have a different font and plate color to differentiate between the types of vehicles. Due to variation in the plate features, the projection profile methods are proposed in [8-9]. In [8], the number of back pixels was counted in each column of the vertical projection and based on the transition from a crest to its trough; the character is segmented. Whereas, in Vietnamese license plates, the combination of vertical and horizontal projection together is applied to segment the characters with predefined constraints [9]. The Connected Component Analysis (CCA) [10] is one of the most used algorithms for segmentation with various thresholding methods. The combined use of connected component labelling algorithm with a predefined aspect ratio and pixel counting method achieved an accuracy of 96.5% [10]. C. Character Recognition After the character segmentation, the last stage in the ANPR technology is the recognition phase. The efficiency of character recognition entirely depends on the quality of the segmentation of each character. Among different proposed approaches, template matching is one of the simple and straight forward method [1]. The extracted character, after normalization, is matched with the templates stored in the database using the hamming distance [8] and Jaccard value [11]. The template matching method is insufficient if an extracted character is distorted. Therefore, the machine learning technique is introduced to extract unique features from the segmented characters. In [12], some of the methods such as zoning and regional feature extraction, are introduced. The proposed feature extraction method achieved an accuracy of 84% [12]. III. THE PROPOSED SYSTEM A. System Overview The typical overview of the proposed system is shown in Fig. 2. The input of a system is an image acquired by the digital camera and, the expected output is the recognized license plate characters. The input image is fed to the trained YOLOv2 (You-Only-Look-Once) [13] model for the license plate extraction. Then the combination of image preprocessing with the connected component analysis isolates the characters for the recognition. For the character 2
2019 4th International Conference on Information Technology (InCIT2019) country to country, we propose a YOLO start-of-the-art yellow background and black foreground (BT). Therefore, to technique for the localization which can handle different solve the problem of these license plates (Fig. 3a), firstly the types of license plates found in Bhutan. YOLO takes care of 35% of the upper region was cropped to remove the different formats of license plates found in Bhutan Bhutanese scripts (Fig. 3b). Then a pixel counting method irrespective of color information and the quality of the image. counts the number of white pixels in the image (Fig. 3c), and Unlike other object detection methods, YOLO can predict after that, it was checked with the threshold value. Base on bounding boxes and the class probabilities from a full image the threshold value condition, the white pixels were inverted in one evaluation [15]. In YOLO, the input image is divided to black pixels and vice versa (Fig. 3d). After getting the in s x s grid cell where each grid cell is responsible for required binary image (background-0, foreground-1), the predicting the object. During the detection process, each grid CCA method was applied to the binary image to segment the cell predicts 5 components: (x, y, w, h, confidence) for the characters (Fig. 3e). bounding boxes. The (x, y) coordinates represent the center of the bounding box that encloses the object and (w, h) Fig. 3. a) Input image, b) Crop image, c) Thresholding, d) Inverted represents the dimensions of the box with relative to image image, e) Segmented characters size. The cell which contains the center of the object is responsible for predicting the bounding boxes. The D. Character Recognition coordinates are normalized to 0 and 1. Since YOLO predicts The character recognition is the last phase in the proposed multiple bounding boxes per grid cell, those bounding boxes with highest IOU with ground truth were used to filter out the system. The features from the individual character were unwanted bounding boxes. Before training the YOLO model, extracted to create a character classifier. The success of the we had changed number of filters in the last convolutional character classification entirely depends on the quality of the layer to match the number of classes (C). Since YOLOv2 uses features used in the classifier. Since characters represent the concept of anchor boxes (A) to predict the bounding different shapes of different forms, the combination of Hu’s boxes (A=5) having 4 offsets with one confidence score, so Moments along with the CD method was proposed to the formula to find the filters is given by: differentiate number (6, 9) and (A, 4). Hu’s moments contain six orthogonal invariants and one skew orthogonal invariant filters=(C+5) * A (1) based on the algebraic invariants [14] which are invariant to rotation, translation and scaling. Hu’s moments are used to Since we intend to detect only one class, the number of describe, characterize and quantify the shape of an object. So filters was set to 30. In the study, 500 datasets were annotated far, the Hu’s moments were used in recognition of and trained using the modified version of Darknet [13] based handwritten based characters [18]. In their study, the overall on Alexey’s implementation [16] since some of the unwanted accuracy achieved was around 63%. For the classifier of the logs are removed keeping only the final iteration. The original proposed method, 1000 characters were used to generate the Darknet framework generates more logs, which creates a long features. Table I shows the frequency of most used characters queue during the training. Training a YOLO model using and numbers in the Bhutanese license plate. Darknet required some prerequisites like GPU, Nvidia CUDA, cuDNN and OpenCV. Due to hardware constraints in TABLE I. TOTAL NUMBER OF CHARACTERS AND FIGURES USED IN THE the study, Google Colaboratory was used to train the model. CLASSIFIER It provides Tesla K80 GPU with 12GB RAM and has all the preconfigured libraries required by the Darknet. Character Count Character Count Character Count 0 66 5 60 A 66 C. Character Segmentation 1 64 6 80 B 65 2 63 7 62 G 64 The detected license plate from the previous step was 3 68 8 64 P 66 preprocessed and, the actual license plate characters were 4 65 9 85 T 62 extracted. In the proposed method, the upper part (35%) of the license plate was cropped to eliminate the Bhutanese script since it does not play any significant role in the identification of the vehicles. Apart from this, it also helps to remove bolts from the license plates. Then the cropped plate was converted to grayscale from the RGB format. On the grayscale image, an iterative bilateral filter was used for noise removal whereby the edges are preserved during noise reduction [17]. Then the brightness of the plate was adjusted using Adaptive Histogram Equalization (AHE) [17] since it shows better contrast than Histogram equalization. In the final stage, the image is binarized, and the CCA with a predefined aspect ratio was used for the segmentation. Fig. 2 shows the sequence of steps for segmentation. 1) White Pixel Counting with Inversion Method The mentioned approach works well for a license plate having a red background with a yellow (BG) or white foreground (BP), but this method cannot handle plates with a 3
2019 4th International Conference on Information Technology (InCIT2019) 1) Centroid Difference Method 1) Character Recognition The Hu’s moments for number (6, 9) and (A, 4) were Hu’s moments alone are not enough to get better classifier which gives high accuracy since some characters have a almost similar due to the presence of same geometrical shape. similar shape such as (B, 8), (6, 9), (A, 4) and (T, 7). These Therefore, a Centroid Difference method was proposed to characters give the same features for Hu’s moments. generate more distinguishing features for those characters Therefore, the CD method, as shown in Fig. 4, was proposed along with Hu’s moments. The 10-fold cross-validation to differentiate (6, 9) and (A, 4). technique was used to train and test the features in WEKA [19]. The classification accuracy of the model using random Fig. 4. Two centroids of number 6 and 9 forest classifier [20] was compared with the features extracted from the Hu moments against the combined use of In the CD method, two centroids were computed from the Hu’s moments and the CD method. The classification matrix given character. C(x, y) represents the central coordinates of obtained from the trained model is given in Table III and the small hole inside number “6” and “9”, and I(x1, y1) Table IV. denotes the centroids of the image. After finding the two centroids for each character, the y coordinate of the small hole TABLE III. RESULTS OF THE CLASSIFICATION BASED ON HU’S MOMENTS was subtracted from the corresponding image coordinate, as given by Equation 2: 0 1 2 3 4 5 6 7 8 9ABGPT 0 65 0 0 0 1 0 0 0 0 0 0 0 0 0 0 CDy= y1- y (2) 1 0 58 0 0 3 0 0 2 0 0 0 0 0 0 1 2 0 0 57 0 0 1 0 0 0 4 0 0 1 0 0 After subtraction of the corresponding y-coordinates, the 3 0 0 0 68 0 0 0 0 0 0 0 0 0 0 0 resultant y-coordinate value was used as a feature along with 4 0 2 0 0 57 0 0 0 0 0 4 0 0 2 0 the Hu’s moments. It was found that the difference in the 5 0 0 1 0 0 55 0 0 0 0 0 2 2 0 0 centroids of number “6” and “9” was negative and positive, 6 0 0 2 0 0 0 46 0 0 31 0 0 1 0 0 respectively. 7 0 1 0 0 0 0 0 61 0 0 0 0 0 0 0 8 0 0 0 0 0 2 0 0 58 0 0 4 0 0 0 The CD method was also used in extracting features from 9 0 0 2 1 0 0 29 0 0 53 0 0 0 0 0 number “4” and character “A” since the two centroids can be A 0 0 0 0 6 0 0 0 0 0 55 0 0 3 2 easily computed. The same steps were applied to (A, 4) and B 0 0 0 0 0 0 0 0 11 0 0 51 3 0 0 it was found that the difference in the centroid of “4” was less G 0 0 0 0 0 0 1 0 0 0 0 0 63 0 0 than “A”. P 0 1 0 0 3 0 0 0 0 0 0 0 0 62 0 T 0 1 0 0 0 0 0 3 0 0 0 0 0 0 58 IV. EXPERIMENTAL RESULTS TABLE IV. RESULTS OF THE CLASSIFICATION BASED ON HU’S MOMENTS The proposed YOLO approach for license plate AND CENTROID DIFFERENCE METHOD localization works well for different types of Bhutanese license plates. The model was trained for 10,000 iterations, 0 1 2 3 4 5 6 7 8 9ABGPT where each iteration took about 3.8 seconds. The overall test 0 65 0 0 0 1 0 0 0 0 0 0 0 0 0 0 accuracy was 99% with the average training loss of 0.004731 1 0 58 0 0 2 0 0 2 0 0 0 0 0 1 1 for 64 batch size with 8 subdivisions. The model generated 2 0 0 61 0 0 2 0 0 0 0 0 0 0 0 0 from the training phase was able to localize the license plates 3 0 0 2 66 0 0 0 0 0 0 0 0 0 0 0 irrespective of size, colors and different illuminations. The 4 0 3 0 0 60 0 0 0 0 0 1 0 0 1 0 region of interest was cropped and then send to the 5 0 0 1 0 0 55 0 0 0 0 0 2 2 0 0 segmentation module to isolate the characters from the plate. 6 0 0 0 0 0 0 80 0 0 0 0 0 1 0 0 To check the accuracy of the character segmentation 7 0 1 0 0 0 0 0 61 0 0 0 0 0 0 0 algorithm; 100 cropped license plates were tested, and the 8 0 0 0 0 0 2 0 0 57 0 0 5 0 0 0 accuracy was 89%, as shown in Table II. 9 0 0 1 1 0 0 0 0 0 83 0 0 0 0 0 A 0 0 0 0 3 0 0 0 0 0 62 0 0 1 0 TABLE II. SUCCESS RATE (%) OF CCA B 0 0 1 0 0 0 0 0 8 0 0 54 2 0 0 G 0 0 1 0 0 0 0 0 0 0 0 0 63 0 0 Total Success Failure Success P 0 2 0 0 1 0 0 0 0 0 0 0 0 62 1 Images 89 11 Rate (%) T 0 1 0 0 0 0 0 2 0 0 0 0 0 0 59 100 89% In Table III, the number “6” was incorrectly classified 31 times as “9”, whereas the number “9” was predicted 29 times as number “6”. However, in Table IV, the number of misclassifications was reduced to 0 with the introduction of the proposed CD method along with Hu’s moments. In the case of (A, 4), the CD method had insignificant effect in reducing the misclassification. The overall recognition rate in Table III and Table IV were 86.7% and 94.6% respectively. Therefore, the proposed CD method has significant 4
2019 4th International Conference on Information Technology (InCIT2019) implications on distinguishing number “6” and “9”. The V. CONCLUSIONS algorithms were implemented using Python programming language with the OpenCV library [21]. Since there were The purpose of the study was to develop an ANPR using imbalanced character datasets used in the classifier, the Bhutanese license plates. For the license plate localization, the Precision, Recall and F1-Score (Refer Table V) for features YOLO approach was proposed since it takes care of different extracted using the Hu’s moments and the CD gives the formats of Bhutanese plates. The model which is trained on overall performance of each character. the degraded images (low resolution) were able to generalize the localization of license plates on the high-resolution dataset In Table V, the Precision and Recall rate for (6, 9) and (A, and has higher robustness. In the character segmentation, 4) is above 90%, which shows that there is low false-positive CCA techniques and the white pixel counting with inversion rates and false-negative rates. The proposed CD method has method were introduced. The proposed character a significant impact on differentiating those characters having segmentation methods with a preprocessing method were able the same geometrical shape, i.e., (6, 9) and (A, 4). Fig. 5 to handle different formats of license plates. In the character shows some of the recognized license plate characters by the recognition phase, Hu’s moments combine with Centroid proposed system. Difference method were proposed for the feature extraction. The proposed Centroid Difference method was able to TABLE V. PRECISION, RECALL AND F1-SCORE FOR EACH CLASS distinguish the features of number “6” and “9” significantly along with Hu’s moments. The random forest classifier was Class Precision Recall F1-Score chosen as the machine learning algorithm to train and test the (%) (%) (%) extracted features, and the classification accuracy was 94.6%. 0 100 98.5 99.2 The proposed algorithm could not achieve 100% accuracy due 1 to deformed and fewer character datasets used in the classifier. 2 89.2 90.6 89.9 3 91 96.8 93.8 VI. FUTURE WORK 4 98.5 97.1 97.8 5 90.9 92.3 91.6 In this study, the new method was proposed to handle 6 characters (6, 9) and (A, 4) but, in future, more method needs 7 91.7 91.7 91.7 to be developed to differentiate other numbers and characters 8 100 100 100 having same geometrical shape. Along with this, the number 9 93.8 98.4 96.1 of training datasets needs to be increased for more robustness A 87.7 89.1 88.4 and high accuracy. B G 100 97.6 98.8 ACKNOWLEDGMENT P 98.4 93.9 96.1 T 88.5 83.1 85.7 The authors would like to thank Naresuan University for Weighted 94.0 98.4 96.2 supporting with the resources and advising staffs in this Avg 95.4 93.9 94.7 project. Also, Yonten; the scholarship student would like to 96.7 95.2 95.9 extend his most profound gratitude to His Majesty the King of 94.6 94.6 94.6 Bhutan and Naresuan University for granting him a master’s degree scholarship. Fig. 5. Recognized license plates; a) Different illumination, b) Night view REFERENCES [1] S. Du, M. Ibrahim, M. Shehata, and W. Badawy, “Automatic License Plate Recognition (ALPR): A State-of-the-Art Review,” IEEE Trans. Circuits Syst. Video Technol., vol. 23, no. 2, pp. 311–325, Feb. 2013. [2] R. Laroca et al., “A Robust Real-Time Automatic License Plate Recognition Based on the YOLO Detector,” 2018 Int. Jt. Conf. Neural Netw. IJCNN, pp. 1–10, Jul. 2018. [3] X. Shi, W. Zhao, and Y. Shen, “Automatic License Plate Recognition System Based on Color Image Processing,” in ICCSA, 2005. [4] A. H. Ashtari, M. J. Nordin, and M. Fathy, “An Iranian License Plate Recognition System Based on Color Features,” IEEE Trans. Intell. Transp. Syst., vol. 15, no. 4, pp. 1690–1705, Aug. 2014. [5] D. Zheng, Y. Zhao, and J. Wang, “An efficient method of license plate location,” Pattern Recognit. Lett., vol. 26, no. 15, pp. 2431–2438, Nov. 2005. [6] P. Rattanathammawat and T. H. Chalidabhongse, “A Car Plate Detector using Edge Information,” in 2006 International Symposium on Communications and Information Technologies, 2006, pp. 1039– 1043. [7] H. Li, R. Yang, and X. Chen, “License plate detection using convolutional neural network,” in 2017 3rd IEEE International Conference on Computer and Communications (ICCC), 2017, pp. 1736–1740. [8] M. Sarfraz, M. J. Ahmed, and S. A. Ghazi, “Saudi Arabian license plate recognition system,” in 2003 International Conference on Geometric Modeling and Graphics, 2003. Proceedings, 2003, pp. 36–41. [9] T. D. Duan, T. L. H. Du, T. V. Phuoc, and N. V. Hoang, “Building an Automatic Vehicle License-Plate Recognition System,” in International Conference in Computer Science, 2005. [10] N. Vishwanath, S. Somasundaram, M. R. R. Ravi, and N. K. Nallaperumal, “Connected component analysis for Indian license plate 5
2019 4th International Conference on Information Technology (InCIT2019) infra-red and color image character segmentation,” in 2012 IEEE International Conference on Computational Intelligence and Computing Research, 2012, pp. 1–4. [11] E. R. Lee, P. K. Kim, and H. J. Kim, “Automatic recognition of a car license plate using color image processing,” in Proceedings of 1st International Conference on Image Processing, 1994, vol. 2, pp. 301– 305 vol.2. [12] R. P. Patel, N. M. Patel, and K. Brahmbhatt, “Automatic Licenses Plate Recognition,” Int. J. Comput. Sci. Mob. Comput., vol. 2, no. 4, p. 10, 2013. [13] J. Redmon and A. Farhadi, “YOLO9000: Better, Faster, Stronger,” ArXiv Prepr. ArXiv161208242, Dec. 2016. [14] M.-K. Hu, “Visual pattern recognition by moment invariants,” IRE Trans. Inf. Theory, vol. 8, no. 2, pp. 179–187, Feb. 1962. [15] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You Only Look Once: Unified, Real-Time Object Detection,” ArXiv150602640 Cs, Jun. 2015. [16] Alexey, Windows and Linux version of Darknet Yolo v3 & v2 Neural Networks for object detection (Tensor Cores are used): AlexeyAB/darknet. 2016. [17] S. Kaur and S. Kaur, “An Efficient Approach for Automatic Number Plate Recognition System under Image Processing,” Int. J. Adv. Res. Comput. Sci., p. 8, 2014. [18] S. Zekovich and M. Tuba, “Hu Moments Based Handwritten Digits Recognition Algorithm,” p. 6. [19] E. Frank, M. A. Hall, and I. H. Witten, The WEKA Workbench. Online Appendix for “Data Mining: Practical Machine Learning Tools and Techniques,” 4th ed. Morgan Kaufmann, 2016. [20] L. Breiman, “Random Forests,” Mach. Learn., vol. 45, no. 1, pp. 5–32, Oct. 2001. [21] G. Bradski, “The OpenCV Library,” Dr. Dobb’s Journal of Software Tools, 2000. 6
2019 4th International Conference on Information Technology (InCIT2019) Application of Large Neighborhood Search and Differential Evolution for Solving Vehicle Routing Problem in Post Office Delivery for the post office, Chiang Rai branch Nuttachat Wisittipanit Materials for Energy and Environment Research Group, Department of Material Engineering, School of Science Mae Fah Luang University Chiang Rai, Thailand nuttachat.wis@mfu.ac.th Abstract—Transportation cost accounts for approximately assigned addresses. The post office in focus is the Chiang half of the total operational expense for a logistics entity such Rai post office branch, situated on Uttarakit Road, Tumbon that it should possess decent solutions to manage Viang, Chiang Rai Province, Thailand. This particular post transportation to reduce the cost. A post office is the logistics office branch is responsible for both receiving and establishment that has to manage a group of delivery vehicles distributing parcels in the areas of Tumbon Viang, Rob- where its current routing solution might not be effective, Viang, Mae Korn, Sansai, Mae Yao and Rimkok. It has the relying on skills of delivery drivers which could lead to total of sixteen delivery vehicles for both normal and EMS unnecessary consumption of fuel; thus, determining the (Express Mail Service) operations; in those sixteen vehicles, optimal vehicle routing to deliver parcels, known as Vehicle there are four trucks and fourteen motorcycles. Each Routing Problem (VRP), could help save its cost greatly. This operational day, as parcels enter the post office, they are study used Evolutionary-based algorithms, Differential handled and distributed to customers at addresses in various Evolution (DE) combined with Large Neighborhood Search areas in a timely manner; the parcels are assigned to drivers (LNS) to find optimal routing for delivery vehicles belonging of delivery vehicles according to their responsible zones, to the Chiang Rai post office, Thailand. The approach was with some parcels requiring an expedited delivery process. applied to the routing of two vehicles responsible for one Once the post office designates and loads packages to the delivery zone in urban Chiang Rai area. The results showed delivery vehicles, the drivers would use their familiarity and that such optimization method had the average total distance instinct for the routing processes for the delivery; the per day per vehicle of 19.43 km, ~32.7% lower than that of exercise of mastery for delivery in the areas is one of the current routing employed by the post office, given that 29 days major problems of the post office operation. The instinctive of data for vehicle routing were tracked. The proposed routing method might not be efficient enough, causing slow combined method could be used as an effective decision-maker delivery and high fuel consumption; additionally, if new in the routing process for the post office. drivers were assigned to an area not familiar to them, the delivery could face a major delay or could not match Keywords—Large Neighborhood Search, Differential assigned addresses. To tackle such issue, a systematic Evolution, Vehicle Routing Problem, Optimization, Logistics approach or automated system can be employed. I. INTRODUCTION The VRP is an NP-hardness (Non-deterministic Polynomial-time hardness) problem [4] that can be solved Management of vehicle routing in an organization that is by metaheuristics methods such as Large Neighborhood involved in transportation operation is a daunting task, Search (LNS) and Differential Evolution (DE) to determine especially if the survival of such entity largely depend on the optimization in routing process. In this research, the DE how it handles the routing complexity. For industries related algorithm, combined with LNS, was used to find the daily to logistics, the large portion of operational cost is from the routing optimization of two delivery vehicles, i.e. finding the transportation activity [1] such as fuel expense and driver routing solution that minimized the total distance of both wages. Therefore, a logistics establishment should vehicles which is essentially a single-objective optimization determine the optimized ways to deal with the routing problem. The research tracked the routing for approximately problem effectively i.e. having an exceptional solution to the one month given that the vehicles belong to the Chiang Rai transportation management such that it could drastically post office. The results showed that such systematic minimize its overall operation cost, saving both time and approaches reduced the average total distances per day and expenses as well as satisfying its customers. per vehicle by ~32.7%, compared to the traditional methods. The proposed method could be considered as a major This research is focused on solving the Vehicle Routing assistant in the routing process for the delivery drivers which Problem (VRP) [2,3] of a post office, which is a major ultimately saves cost and time for their organization. logistics entity managing a large fleet of vehicles, composed of various types and capacities, to deliver parcels to certain This research is supported by Mae Fah Luang University, fiscal year 2018 7
2019 4th International Conference on Information Technology (InCIT2019) A. Differential Evolution Both of the tracked delivery vehicles were responsible Differential Evolution (DE) algorithm is an for zone number 1, according to Table 2; and the vehicles were labelled (or named) as Vehicle A and B. Vehicle A and optimization-oriented algorithm, based on evolutionary B made parcel delivery for sub-zones of zone 4, namely sub- approaches [5] which include the elements of mutation and zone U and V, respectively. From 5th February to 9th March cross-over over a continuously explored space. DE 2018, the average number of parcel delivery and travel algorithm is initiated with a randomly generated population distance per day (total of 29 days, excluding Sunday) of each of size N of D-dimensional vectors; each value in a D- delivery vehicle is shown in Table 1. dimensional vector is a continuous real number in the range of 0 to 1. Then, the population evolves into a new one by the The delivery routes of both vehicles were tracked by cycles of mutation, crossover and selection (which mimics GPS-enabled smartphones; an application was installed to the process of life evolution) until some criteria are satisfied. retrieve the routing information and total distance traveled B. Large Neighborhood Search as well as times and location coordinates; the application recorded a coordinate whenever the vehicle stops for a brief Large Neighborhood Search (LNS) is another period (more than 3 seconds). The stored coordinates were optimization method that tries to find the best solution for then mapped onto the map application (Google Earth Pro®) the optimization problem [6]; this method is suitable for a to show the exact physical locations; the visual presentation large problem in which the exact solution is extremely hard also helped verify if the coordinates were correct. Next, the to determine. The solution from LNS algorithm might not be coordinate data were filtered and arranged such that some the best; however, many applications employing LNS undesired coordinates, e.g. stops at traffic lights and other showed that its solutions were adequately effective. For an non-registered addresses, were removed. Then, a python optimization process, LNS reiteratively transforms a current script determined the shortest distances between each pair of solution into a new one, called neighborhood solution coordinates via the googlemaps module (Google Map API), because the new solution is relatively similar to the old one. and automatically constructed a distance matrix that has ������ × Generally, the new solution is obtained by a small ������ dimension where ������ equals to the number of travel modification to the old solution; examples of modification coordinates including the main post office. Noting that this are as follows, type of matrix is not a symmetric one because the shortest distance from point ������ to point ������ might not be the same as (1) Destroy and Repair: Some parts of a solution are that from point ������ to point ������ considering some routes might destroyed and then a new solution is constructed by enforce a one-way traffic. repairing the old solution, adding the destroyed parts back in different arrangement. TABLE 1. AVERAGE NUMBER OF DELIVERY AND TRAVEL DISTANCE PER DAY OF BOTH DELIVERY VEHICLES FOR THE TOTAL OF 29 DAYS (5TH (2) Swap: Some parts of a solution are swapped to create a new solution. FEBRUARY TO FRIDAY, 9TH MARCH, 2018) (3) Reverse: Some parts of a solution are reversed to Vehicle Average delivery Average travel distance create a new solution. number (KM) 1 52.14 12.85 In certain circumstances, LNS can increase the size of 2 modifications to escape local minima, reaching for the 45.66 14.18 global optimum faster; however, this has some risks that the modifications might be too big and the algorithm has a high For each day of the 29 days surveyed, two distance probability to move far away from the global optimum [7]. matrices were generated using the retrieved coordinates from vehicle ������ and ������ , one matrix for each vehicle rote. II. MATERIALS AND METHODS Thus, the total of 58 matrices were created. These matrices served as essential input data for the vehicle routing problem A. Dataset using large neighborhood search and differential evolution. Residential and commercial addresses, where the Chiang B. Application of DE and LNS to VRP Rai post office is responsible for postal delivery, can be The application of DE|LNS first starts with DE where the grouped into 6 zones. Information on the delivery operation of the post office was surveyed for 29 days (approximately distance matrix of size (������ + 1) × (������ + 1), given ������ being one month, excluding Sunday) from Monday, 5th February the number of delivered customers, is used in the to Friday, 9th March, 2018, following 2 delivery vehicles computation. In the optimization process of DE, initially, a (motorcycle-type) using GPS-enabled 4G smartphones with defined set of randomly generated vectors (values ranging a route-tracking application. These two vehicles have two from 0-1), each has the size of (������ + ������) where ������ equals to the large pouches strapped on and can carry relatively large number of vehicles and ������ equals to the number of delivered number of parcels. Each vehicle was responsible for a customers, is created. The number of generated vectors is defined area of operation in the region of Tumbon Viang and ������������������������������������������������������ in the DE parameter. The DE fitness function Rob-Viang, Amphoe Muang Chiang Rai, Chiang Rai evaluates each vector that generates one solution to the VRP province. The Chiang Rai post office is responsible for both where the solution refers to the delivery routes of all the regular delivery and EMS (Express Mail Service) and the vehicles; next, DE determines the global best vector which average numbers of addresses, delivered per day for six basically is the vector in the set that produces the best zones in Tumbon Viang and Rob-Viang areas from 5th solution; such solution is the vehicle routes having the least February to 9th March 2018 were as follows, zone 1: 97.79, total delivery distance. Then, each vector goes through the zone 2: 114.54, zone 3: 225.12, zone 4: 259.68, zone 5: mutation and crossover process, obtaining a new set of trial 158.34 and zone 6: 196.86. vectors which are again evaluated by the fitness function; DE chooses the best solution between a current vector and its trial vector. In the case that the current vector is worse 8
2019 4th International Conference on Information Technology (InCIT2019) than its trial vector, the current vector becomes the trial 0 173 4 vector. Then, DE updates the global best solution. The 21 process iterates until it reaches the maxiter input parameter 193 V: 1 of DE and obtains a final global best solution along with its vector. Post DE requires several essential parameters to be set, other 279 than swarmsize and maxiter which can affect the mutation and crossover process as well as the output. To illustrate the 71 whole process of DE|LNS optimization operation, let use a 165 simple case where there are 2 vehicles trying to find optimal routes to deliver parcels to 5 customers from the post office. 1 V: 0 First, a distance matrix between the post office and 80 3 customers is determined, shown in Fig. 1. The sample matrix in this case is symmetric; however, a real one might be Fig. 4. Delivery routes from the post office of vehicle 0 and 1 asymmetric depending on the locations where the delivery process takes place. distance of (71 + 80 + 165) + (79 + 21 + 173 + 193) = 782 units. In addition, the solution from Fig. 4 gives delivery routes for each vehicle i.e. vehicle 0’s route: customer 1 -> 3 and vehicle 1’s route: customer 2 -> 4 -> 0. Once the final solution from DE algorithm is obtained i.e. DE finishes the iterations according to ������������������������������������������ parameter, the solution is transferred to the optimization operation by LNS algorithm. One of the basic operations of LNS is known as Destroy and Repair operation where, randomly, input routes are partly destroyed based on ������������������������������������������������������������ parameter and then LNS repairs or reconstructs them inserting one address added at a time to the route, and selects a new route that meets certain criteria. In this research, the criteria refer to a situation when total distance of the new route is minimal. Fig. 1. A distance matrix for illustration purpose having 5 customers To clarify the LNS optimization process, let use the input and 1 post office. Note that this distance matrix is a symmetric one. routes according to Fig. 4 with two vehicles. Suppose that in the Destroy process, LNS destroys customer 1 and 0 which When DE starts, it generates a group of randomly valued belong to routes of vehicle 0 and 1, respectively, as shown vectors with the size of (2 + 5) = 7 (2 vehicles, 5 customers). in Fig. 5. Each generated vector from DE is split into two sections, one for vehicles and the other for customers as shown in Fig. 2. After destroying customer nodes in the routes (or To be destroyed To be destroyed CM: 1 CM: 0 71 193 173 POST 80 POST CM: 4 165 0.3 0.5 0.3 0.7 0.1 0.2 0.4 79 21 CM: 3 No. V No. CM Vehicle 0 CM: 2 Vehicle 1 Fig. 2. Split vector generated from DE After Removal POST POST CM: 4 Then, for each section, indexes are assigned to each item CM: 3 starting from 0 according to items’ values, ranked from the lowest to highest. The customer section is further split into CM: 2 sub-sections in which the number of them depends on the number of vehicles; the sub-sections have equal numbers of Fig. 5. Delivery routes from the post office of vehicle 0 and 1 items in case the number of customers is even or difference of one item in case the number of customers is odd. In removing them from the routes), LNS keeps them in the pool addition, the sub-sections are randomly assigned to the which is now consisted of customer 0 and 1. Then, LNS vehicles. The described process is shown in Fig. 3 where begins the Repair process which test customers in the pool sub-section 0 is assigned to vehicle 0 and like-wise, sub- as follows: section 1 is assigned to vehicle 1. (1) The process tests the insertions of customer 1 in all 10 possible ways in the routes; then, it selects the insertion that causes the minimum of total distance. 0 124 013 In the case of customer 1 (index 1), the insertion 0 0.3 0.5 0.3 0.7 0.1 0.2 0.4 is chosen since it leads to the minimal total distance of 459. No. V No. CM (2) After selected position is chosen for customer 1, the next operation of Repair process also includes Fig. 3. Sub-sections assigned to each vehicle that customer in the route. Next, in the case of Using information from Fig. 3, the fitness function then customer 0, the insertion 5 is chosen since it leads appoints customers to each vehicle according to the assigned indexes, resulting in the delivery routes as shown in Figure to the minimal total distance of 617 units according 3-5. Such result is only one solution among herds of to Figure 3-8. solutions produced by DE. This particular solution has total The LNS operation keeps repeating the Destroy and Repair process until the number of iterations reach maxiterLNS. During LNS run, if the solution is better than 9
2019 4th International Conference on Information Technology (InCIT2019) that of DE or previous iteration of LNS, the process updates relatively low quality of optimized solution, the process that solution to be the current one. There is one essential might be quite fast. Nevertheless, it all depends on the nature parameter bigRep that governs the DE|LNS iteration of a problem; for a case that run-time duration is not an process, this parameter is set to 4. Once all the iterations of important factor, such as finding a protein folding from DE and LNS, according to the set parameters, are complete, amino-acid sequence, both mentioned parameters can be the final solution is considered the best one of that particular exceptionally large. On the other hand, for this particular run. The total distance and specific routes of all the vehicles case, the parcel delivery of the post office is a day-to-day are then reported. operation; therefore, any information provided for the decision of vehicle routing must be processed in a timely III. NUMERICAL EXPERIMENTS AND RESULTS manner. As such, the parameters “ swarmsize” and In this research, the DE|LNS method was applied to the “maxiter” as well as “maxiterLNS” were adjusted so that the routing of two delivery vehicles responsible for zone 1 of the duration it takes to run the DE|LNS optimization program is Chiang Rai post office, from 5th February – 9th March 2018 acceptable i.e. approximately within 30 minutes after (29 days). The exact delivery coordinates (latitudes and gathering all the coordinates for delivery. longitudes) were extracted from the GPS-enabled 4G smartphones having a route-tracking application; those The optimization program ran the optimization process coordinates were then used by a Python script that on routing data for the total of 29 days. After the program determined the shortest travel distances of all coordinate runs for DE|LNS optimization process over 29 days were pairs using the googlemaps module provided by Google Inc finished, the average total distance between DE|LNS routing (Google Maps API). The script automatically constructed a and the post office routing as well as the percent difference distance matrix (unit: km) having ������ × ������ dimension where ������ between them were calculated. Table 3 provides the equals to the number of travel coordinates including the comparison of average total distance between DE|LNS main post office; for instances, if there are 45 customers or routing and the post office routing over 29 days along with addresses where vehicles must deliver parcels to, the the percent difference. The average numbers of delivered distance matrix would have (45 + 1) × (45 + 1) or 46 × 46 addresses for vehicle 1 and 2, also with percent difference dimension. The matrix might not be symmetrical depending shown. The average total number of delivered address and on the coordinates. That means, for this particular research, the average program run duration are also shown. the total of 29 distance matrices were obtained (from 29 days). Each matrix acted as a distance look-up table (pairs TABLE 3. COMPARISON OF AVERAGE TOTAL DISTANCE BETWEEN DE|LNS of coordinates) for the optimization program written in ROUTING AND THE POST OFFICE ROUTING OVER 29 DAYS (WITH PERCENT Python. DIFFERENCE). THE AVERAGE NUMBERS OF DELIVERED ADDRESSES FOR A. Results VEHICLE 1 AND 2 (WITH PERCENT DIFFERENCE); THE AVERAGE TOTAL For the routing of each day, the DE|LNS method computationally and automatically split the routes between NUMBER OF DELIVERED ADDRESS AND THE AVERAGE PROGRAM RUN 2 vehicles and certain routing for each vehicle was also determined, where the method tried to find the optimal DURATION ARE ALSO SHOWN. objective which is the minimum travel distance. Table 2 shows the parameter values of DE and LNS algorithms set Average Total Distance DE|LNS Routing 19.43 to run the program. These parameter values were the same (KM) Post Office 27.03 for 29 program runs across 29 distance matrices. Percent Difference (%) 32.70 Average Number of Vehicle 1 52.14 Delivered Addresses Vehicle 2 45.66 Percent Difference (%) 13.25 Average Total Number of Delivered Addresses 97.79 TABLE 2. ESSENTIAL PARAMETERS FOR DE|LNS AND THEIR VALUES Average Program Run Duration (minutes) 16 Algorithm Name Value IV. DISCUSSION AND CONCLUSION DE swarmsize 100 This research employed DE combined with LNS LNS 100 (DE|LNS) to solve VRP of the post office’s delivery maxiter 0 operation with 2 vehicles over 29 days (5th February – 9th lb_value 1 March 2018), focusing on only one delivery zone i.e. zone 1 ub_value 2.5 of Chiang Rai post office. Routing data from those vehicles FMax 1.5 were filtered, cleaned and then transformed into GPS 0.8 coordinates where they were used to generate distance FMin 50 matrices by Google Maps API. DE|LNS utilized such CRx 50% matrices for the optimization process of the vehicle routing maxiterLNS where the final optimized solutions had minimum total lnsdestroy delivery distances and also presented certain routing for each vehicle. The numbers of customer addresses assigned to One of the disadvantages of using metaheuristics method both vehicles might be slightly or vastly different, depending such as DE or LNS is that the time it takes to run an on the optimum routing. optimization program might be painfully slow if some The results showed that the average total distance of essential parameters, namely, “swarmsize” and “maxiter” DE|LNS routing, over 29 days, was 32.7% less than that of are too large. Fundamentally, there is a trade-off between the post office routing, according to Table 3. The difference speed and quality of optimized solution, meaning that if one desires to have the best solution possible (objective value is very close absolute minimum/maximum or at exact minimum/maximum), the optimization process would take an extremely long time; however, if one can accept a 10
2019 4th International Conference on Information Technology (InCIT2019) in total average distance between two routing methods was relatively large, suggesting that the optimization method yielded much better vehicle routing than the post office method. The percent difference between average number of delivered addresses for vehicle 1 and 2 was quote low at 13.25%. Another critical factor was the time it took for the optimization process to run; the average program run duration stood at around 16 minutes which was acceptable in the real-world operation for the post-office where every parcel has to come and go quite fast. That means, in a real scenario, once all the addresses are obtained, it took the optimization program roughly a quarter of an hour to generate the optimized routing for vehicles. There would be some issues if this method were going to be used for a real routing application; those issues refer to how the addresses on parcel can be converted into GPS coordinates and unexpectedly increased run-time of the optimization program when the number of delivered addressed is large. And number of delivery vehicles can affect program run-time. To make the entire process truly practical, first, decent data management of addresses must be arranged and image processing components combined with artificial intelligence (AI) software must be installed to automatically map customer addresses into GPS coordinates such that the optimization program could readily transform the coordinates into distance matrices which are crucial elements to the entire process. In addition, certain parameters must be programmatically adjusted, depending on number of customer addresses to deliver and number of vehicles used for delivery, to make the program run-time within acceptable range of the post office operation. ACKNOWLEDGMENT The author thanks Mae Fah Luang University who supported this study and also thanks Chiang Rai Post Office who kindly provided the data essential for this research. REFERENCES [1] E. Bardi, J. Coyle, and R. Novack, Management of Transportation., Thomson South-Western, 2006, ISBN: 0-324-31443-4. [2] A. G. Canen, and N. D. Pizzolato, “The vehicle routing problem,” Logistics Information Management, vol. 7(1), pp. 11-13, 1994. [3] N. Christofides, A. Mingozzi, and P. Toth, The Vehicle Routing Problem., Chichester, UK: Wiley, 1979, pp. 315–338. [4] R. E. Ladner, “On the structure of polynomial time reducibility,” J. ACM, vol. 22, pp. 151–171, 1975. [5] R. Storn, and K. Price, “Differential Evolution – A Simple and Efficient Heuristic for global Optimization over Continuous Spaces,” Journal of Global Optimization, vol. 11, pp. 341, https://doi.org/10.1023/A:1008202821328, 1995. [6] R. K. Ahuja, J. B. Orlin, and D. Sharma, “Very large-scale neighborhood search,” International Transactions in Operational Research, vol. 7, pp. 301–317, doi:10.1111/j.1475- 3995.2000.tb00201.x, 2000. [7] T. Brueggemann, and J. Hurink, “Two very large-scale neighborhoods for single machine scheduling,” OR Spectrum, vol. 29, pp. 513-533, 2007. 11
2019 4th International Conference on Information Technology (InCIT2019) Anonymity Supporting Tool for Community of Inquiry-Based Platform in Encouraging People to Share Knowledge: A Case Study of Pusilkom Universitas Indonesia Bintang Annisa Bagustari Annisa Monicha Sari Faculty of Computer Science Faculty of Computer Science Universitas Indonesia Universitas Indonesia Depok, Indonesia Depok, Indonesia bintang.annisa81@ui.ac.id annisa.monicha@ui.ac.id Erni Juraida Dana Indra Sensuse Faculty of Computer Science Faculty of Computer Science Universitas Indonesia Universitas Indonesia Depok, Indonesia Depok, Indonesia dana@cs.ui.ac.id erni.juraida71@ui.ac.id Especially in tacit knowledge, there are common Abstract—Creating ideas and having an innovative tendencies in an organization where experts who have environment in an organization are crucial regarding gain a unique knowledge will cause the accumulation of competitive advantage. People would utilize community of knowledge, monopolize, and have the highest reputation in inquiry (CoI) to boost innovation through tacit knowledge their organization [6]. Further- more, knowledge will have sharing processes which they can learn collaboratively. Social a top selling point when there is no competition to cope presence, cognitive presence, and teaching presence are with it, and individual in the organization will use that fact considered as dimensional aspects in CoI that could help an to raise their power which has exclusive control of key effective and efficient collaboration learning. However, the issue organization knowledge [7]. Another issue when the new of personal identity is deemed to be a barrier due to knowledge member wants to share their knowledge in the organization sharing. The hidden contributor in a forum discussion can be are, they feel don’t have the right to share knowledge, feel seen as a way to avoid bureaucracy, conflicts, or the fear of intimidated [8], and fears of accepting criticism or ribbing critics. According to the case, the presence aspects within CoI others [9]. Moreover, the social barriers to sharing could potentially provide people to have an anonymous identity. knowledge are lack of legitimate language, conflict This paper proposes anonymity tool of CoI-based platform and avoidance, bureaucracy and hierarchy, and incoherent evaluates the impacts regarding the element of social presence paradigm [6]. and cognitive presence. We analyze the tool by developing CoI- based prototype and assess participants’ perspectives using an Lesson learned system and expertise locator system are exploratory qualitative study in Pusilkom Universitas two type of explicit knowledge sharing system Indonesia. Findings of this research stated that anonymity has considerably explain in KM literature, and for sup- porting high influence on the aspect of social presence in sharing while tacit knowledge usually used virtually in com- munity of anonymity has low influence on the cognitive aspect of CoI- practice [4]. Community of inquiry (CoI) is developed based knowledge sharing in the organization. from community of practice construct, but new construct introduced in CoI such as agency, objective directed to Keywords—tacit knowledge, knowledge sharing, action, antilogy, and tensions [10]. CoI blend community, collaboration, anonymity, community of inquiry the social aspect, and inquiry to create online or blended learning environment [11]. Previous research had using I. INTRODUCTION CoI framework as guidance to measure effectiveness of learning in online course [12], adding new skill oral Knowledge sharing activity cannot be separated in the indicator (fluency, grammatical, coherence, pronunciation, organization or company to create data, information, and etc.) to online English course using CoI framework [13], knowledge which is essential to the organization. Knowledge but lack of context to encourage people to share their sharing define as activity in providing information or knowledge, while tacit knowledge is hard to form because knowledge to help people in the organization to solve the individual barrier that explains before. According to [14], issues, create new ideas and applying the procedure or policy the issue of online learning through collaborative [1]. An organization will reach its full profit when knowledge technology using CoI should be able to provide to people sharing effectively implement, so create organization with different domain knowledge. competitive advantage. Past research has been explaining the effectiveness ofknowledge sharing activity in different views, According to the case mentioned before, presence aspect focusing on is- sues in transferring tacit knowledge and the in CoI can also utilize anonymity due to the social barrier in complexity of knowledge between internal of the organization an organization. While the anonymity can hide sharers’ [2], and the issues about finding the knowledge itself [3]. identity as can also be applied to the seekers, we can Knowledge management systems are developed to guide people to share their knowledge, both tacit and explicit [4]. Explicit is the knowledge that formally transmitted while tacit knowledge is hard to communicate and formed [5]. 12
2019 4th International Conference on Information Technology (InCIT2019) consider this problem attached to the dimensional aspects construct meaning through continuous communication. Social within CoI. We put these research questions (RQ) as the presence defines as the ability of an individual to link the basis of this research: 1) How does anonymity in CoI-based interaction with each other, create relationship and character platform influence cognitive presence to encourage sharing into the community. Teaching presence defines as learning between the users? process development, the design of learning course, and facilitating of the cognitive and social dimension in the The goal of this paper is to evaluate anonymity as learning activity [18], [19] (see Table 1 for Elements of CoI). support to encourage people in sharing knowledge with a C. Collaborative Learning & Pusilkom Universitas COI-based platform. Using exploratory study, we want to prove whether anonymity becomes an essential thing to Indonesia consider in the real world’s problems as a case study to the In the educational aspect, collaborative learning (CL) COI framework. We also explore literature in the past become an approach between two or more people to learn research comprehensively to understand issues and together [21]. CL direct to enhancement learner’s challenges regarding the participation of knowledge productivity, self-esteem, psychological health, skill, and sharing especially for the tacit knowledge, the their competence [22]. Many different models of implementation of COI, and related strategies of collaborative learning exist, but they all include multiple collaborative knowledge sharing within organizations. As people working to reach the common goal of a deeper a result, collaborative learning with the anonymous understanding of a specific topic, concept, process, or skill. community of inquiry prototype in this study will be Such as In Pusilkom UI, CLs activities mostly happen in proposed which aims to encourage people in sharing their Scrum Events such as daily scrum, sprint review and sprint knowledge. retrospective for project/product development division. It helps the project team to reach a better and same II. LITERATURE REVIEW understanding about the project they work together [23]. A. Knowledge Sharing & Tacit Knowledge Pusilkom itself, so-called Center for Computer Science of Universitas Indonesia, was established in 1972. This facility Knowledge sharing not only about obtain or distribute aims to improve the performance and competence of every knowledge from one area to another but is an activity of member of Universitas Indonesia in the field of computer exchanging and processing [15]. Knowledge sharing increase science. Since its first establishment, Pusilkom UI has the ability of individual in the organization to gain knowledge participated in numerous projects to realize a better future for location with in unformed social network and make them computer science in the academic, government, and realize to create new idea [16]. Michael Polanyi differentiates business/industry settings. In 1986, this center was elected as human knowledge into tacit knowledge and explicit the Inter-University Center for the field of computer science knowledge. Tacit knowledge is knowledge which hard to in the research project development funded by the World communicated, formed and has a personal quality, while Bank. This shows that Pusilkom UI has gained some explicit knowledge or called codified knowledge is the recognition in the field of computer science. In 2005 when UI knowledge that formally transmitted and tends to systematic was still part of BHMN, Pusilkom UI was repositioned as an language [5]. Another research describe tacit knowledge is inherently difficult to understand, the first thing to do in order Fig. 1: Prototype Model in this Research to capture, store and disseminate it is to make it explicit. Externalization as a process to change tacit knowledge academic business unit under the Faculty of Computer become explicit (i.e., by sharing metaphors and analogies Science (FASILKOM). As an academic business unit with the throughout social interactions) [17]. support from highly competent members, Pusilkom UI focuses on providing services in the field of information B. Community of Inquiry (CoI) technology to the industry/business/government in Indonesia [24]. CoI describes as a learning activity which involves three interdependent dimensions, cognitive presence, social presence and teaching presence within the com- munity [18]. Cognitive presence defines as the ability of an individual to TABLE I Elements of CoI [20] Elements Categories Indicators Social Presence Open Risk-free Expression Cognitive Communication Encourage Presence Group Cohesion Collaboration Affective Expression Emotions Triggering Event Sense of Puzzlement Exploration Information Integration Exchange Resolution Connecting Ideas Apply New Ideas Teaching Presence Design & Organization Setting Curriculum & Facilitating Methods Discourse Sharing Personal Meaning Focusing Discussion 13
2019 4th International Conference on Information Technology (InCIT2019) III. METHODOLOGY B. Participants’ Perspectives Exploration We evaluate the prototype using a case study of Pusilkom The methodology used in this study are exploratory qualitative research using prototype software development. Universitas Indonesia. We conduct the evaluation on 30th High fidelity prototype develops in this study with Adobe XD November 2018. There are nine participants from Pusilkom tool. After the prototype freeze, we evaluate the prototype Universitas Indonesia are inter- viewed based on the within a case study Pusilkom Universitas Indonesia. prototype in this study, eight of them are the employee and one a leader in Pusilkom Universitas Indonesia. The duration A. CoI-based Anonymous Prototype of the interview approximates about 10-15 minutes the interviewed conduct in two section, the exploration section, The prototype is commonly used for software and assignment criteria. Firstly, in the exploration section, the development processes technique. Different types of users were given the time to explore the application to gain prototypes are implemented to obtain different objective [25]. insight into the prototype the second, users were given time to According to research [26], prototype software development complete some assignment to control the evaluation beyond defines as an archetype which is patterned as an authentic the objective of this research. model; an important alternative approach shows by an individual to the next type of systems; or a standard of a Fig. 4: Example of teaching presence prototype interface typical example. Past research had discussed the effectiveness of the use of prototype as software development technique is, prototype enable project divided into a smaller development process, make process became easier, cost-effective, greater user participation in development process [25]. In developing this prototype, the study modelled a mapping strategy for each category in CoI as shown as Figure 1. The social presence is implemented based on the principles of social cohesion and open communication which described in Figure 2. The aspect of cognitive presence involved triggered event strategy (Figure 3) while teaching presence can be seen in Figure 4 as direct instructions. In this study, we develop a high-fidelity prototype with mapping three CoI elements to identify the con- tents of the system. The prototype builds using Adobe XD tools and then reviewed by using CoI framework. After the review assessed enough based on the objective in this research, then we freeze the design. Fig. 5: Non-Active Anonymity in the prototype Fig. 2: Example of social presence prototype interface Fig. 3: Example of cognitive presence prototype interface Fig. 6: Active Anonymity in the prototype We recorded the evaluation with screen recording tools. Then we translate the recorded and analyze the user tendency about anonymity-based Community of Inquiry (CoI) framework. Finally, the result of the evaluation can discuss 14
2019 4th International Conference on Information Technology (InCIT2019) and answer the RQ of this study (the prototype model in this 2. Assignment Criteria Section research shows in Figure 1). The second step is the assignment criteria (Table 3). IV. FINDINGS AND DISCUSSION In this evaluation, users were given 10 minutes to complete some instruction and answer about the There are two features we identify as the platform contents specified questions. based on CoI-framework with anonymity tool. We associate cognitive presence with Articles feature, where the end-users B. Findings can share/post the article about certain topic with anonymity or without anonymity, while the social presence was mapped In this research, our objective is to propose how into Discussion feature, where the end-users can initiate the anonymous based community of inquiry effective to discussion about certain topic, comment other discussion or encourage people in sharing knowledge. Based on the like the discussion with or without anonymity. Also, we evaluation prototype we get the result based on the evaluation propose the direct instruction as an aspect of teaching presence as shown in Table 4. We checked each CoI aspect in which which the end-users have an option to use the anonymity- the users are examined preferred to enable anonymity tool or support or not (list of feature shows in Table 2 and the design not. of Prototype shows in Figure 2 and Figure 3). • Anonymity gives high influence on social presence in A. Participants’ Perspective Exploration Community of Inquiry framework. Most of the participant chose to be anonym when commenting There are two evaluation techniques we conducted in this about some topic discussion. study, exploration and assignment criteria: • Anonymity gives low influence on the cognitive 1. Exploration Section presence in Community of Inquiry framework. Most In this evaluation, users were given time about 5 of the participant chose to be non-anonym (use minutes to explore the prototype to gain the insight original identity). Furthermore, most of the participant what the process or activity can be done in the prefers to read or waiting other to share knowledge prototype. based on the article. TABLE II Features propose in this research • The participants argue directed by the instruction whether using the anonymity or not, but further improvement like the positions of instruction needs, TABLE IV: Findings in this research TABLE III: The Assignment Criteria in this research so that users more pay attention with its instruction. • The participant also argues the rewards features is a After the users understand what overall the prototype, then we stop this section and continue to good idea to encourage people to share knowledge, the section assignment criteria. comment, like or triggering the article, but the participant wants the reward more to produce real value (i.e. recognition from the company/organization). C. Discussion In this study, we evaluate to explore our experiment based on the RQ. We get the result that anonymity gives high influence social presence in CoI framework. Most of the participant chose to be anonymous when they want to comment on some topic discussion. The participant said, ” ...with anonymous, I became easy to communicate my opinion about some topic or issue” ”... I’m worried my opinion will refuse or blamed by other, so I choose anonym” 15
2019 4th International Conference on Information Technology (InCIT2019) Based on the evaluation, participant more open when they When the knowledge is crucial to the company, they will want to comment on some topic because they feel easier when consider the trust aspect, while the simple of knowledge they to be anonym. This research consistent with past research will not consider it (i.e. a question about nearest road to the about individuals’ barrier, where they feel to not have the right office) [28]. to share knowledge and feel intimidated [8]. Moreover, they feel afraid when they want to comment because didn’t want Moreover, after the participant discovers knowledge, other blamed him/her. Related with past research while people he/she prefer to begin the discussion or triggering the fear of accepting criticism or ribbing others [9]. discussion in the face-to-face meeting. The participant said: Participant also said that, ” ...when I read some topic or article, I prefer to begin the discussion face-to-face than using technology like ”...I am not confident to share knowledge when certain application.” forum dominates by our leader” Participant feel that technology is not a comfortable place In an online environment, several communities of to begin the discussion because they can discuss with their inquiry provide participants with a degree of anonymity, college in face to face meeting. These findings consistent in which is thought to foster a sense of safety and trust. This is past research, while they have ability attending face-to-face also aligned with their personal attributes which are hidden learning [29], so they prefer not using application, this not in online environments and become not afraid of being consider both the anonymity non-anonymity. judged or treated differently because of their gender, race, class or physical appearance or ones name, the language V. CONCLUSION AND LIMITATION one uses, ones values, perspectives, experiences, the examples one gives, etc. [27]. Moreover, these findings The objective of this research is to evaluate a tool of useful to overcome past research about social barrier where anonymity of community of inquiry-based platform in people choose to avoid the conflict and consider hierarchy or encouraging people to share knowledge. We use exploratory bureaucracy [6]. qualitative research method to gain the purpose in this research. CoI framework is used in this study as an approach The second RQ finds that anonymity gives low influence to identify the content to the proposed prototype. Based on the on the cognitive presence in the Community of Inquiry test in this study which has involved the participant of framework. Most of the participants chose to be non- employee in Pusilkom Universitas Indonesia, anonymity gives anonymous (use original identity). Furthermore, most of the high influence on the social presence of Community of Inquiry participants prefer to read or waiting other to share framework and low influence on the cognitive presence of knowledge based on the article. The participant also said, Community of Inquiry framework. ”...I was waiting other to share knowledge than Several issues could be some limitation in this study which initiate to share my knowledge”. the case should consider the culture and size of the Most of the participant prefers to choose to be non- organization when choosing the project test. Pusilkom anonymous because if he/she wants to share or begin a new Universitas Indonesia is an organization which focuses on topic discussion, they will consider the credibility the project, where they probably will discuss a topic which related discussion. But another participant chooses to be anonym the specific project (for- mal topics), so they prefer to discuss when discussing about free topics (i.e., other company issues how to finish their project. Future research with different size which have low priority or about sensitive issues). The and culture the organization may produce a different result. participant said, Moreover, the legalities should be considered by the organization in using the anonymity tool because they might ” ...I only comment on the discussion when he/she use cause conflicts in the organization (i.e. become a platform to their original account to begin a new topic.” blame or impose individual in the organization). ” ...I will wait for my leader to share or mention some REFERENCES article so that I want to comment on it.” [1] S. Wang and R. A. Noe, Knowledge sharing: A review and directions for ” ...when the topic is sensitive, I prefer to choose to be future research, Human Resource Management Rev., vol. 20, no. 2, pp. anonymous.” 115131, 2010. The participant would wait for their leader first for [2] B. Kogut and U. Zander, 1992 OrgSci Kogut Zander, Organ. Sci., vol. triggering the discussion. This finding consistent with past 3, no. 3, pp. 383397, 1992. research, about social barrier where people choose to avoid the conflict and consider hierarchy or bureaucracy [6]. The [3] D. G. Ancona and D. F. Caldwell, Bridging the Boundary: External participant also wants to share when they trust about the Activity and Performance in Organizational Teams, Adm. Sci. Q., vol. knowledge which related to the findings of this research. 37, no. 4, p. 634, 1992. Otherwise, participants want to share with anonym depends on the topic without considering credibility or trust about [4] I. Becerra Fernandez and R. Sabherwal, Knowledge Management knowledge. This consistent with past research, where a Systems and Processes. 2010. different type of knowledge requires a different form of trust. [5] I. Nonaka, A Dynamic Theory of Organizational Knowledge Creation, Organ. Sci., vol. 5, no. 1, pp. 1437, 1994. [6] G. Disterer, Individual and social barriers to knowledge transfer, Proc. Hawaii Int. Conf. Syst. Sci., p. 220, 2001. 16
2019 4th International Conference on Information Technology (InCIT2019) [7] T. H. Davenport and L. Prusak, Working Knowledge-How Organizations Manage What They Know, vol. 2000. 2000. [8] K. Yip, Exploring Barriers to Knowledge Sharing: A Case Study of a Virtual Community of Practice in a Swedish Multinational Corporation., pp. 164, 2011. [9] A. Ardichvili, M. Maurer, W. Li, T. Wentling, and R. Stuedemann, Cultural influences on knowledge sharing through online communities of practice, J. Knowledge Management., 2006. [10] S. Goodchild, Theorizing community of practice and community of inquiry in the context of teaching-learning mathematics at university, Res. Math. Educ., vol. 16, no. 2, pp. 177181, 2014. [11] A. T. Tolu, Creating Effective Communities of Inquiry in Online Courses, Procedia - Social. Behavior Science., vol. 70, pp. 10491055, 2013. [12] C. R. Jones, Examination of Online Community College Stu- dent: Community of Inquiry Theoretical Model, McKendree, 2017. [13] L. E. Herrera Daz and D. Gonzlez Miy, Developing the Oral Skill in Online English Courses Framed by the Community of Inquiry, PROFILE Issues Teach. Prof. Dev., vol. 19, no. 1, p. 73, 2017. [14] Y. C. Hsu and Y. M. Shiue, Exploring the community of inquiry in the interdisciplinary project-based learning through collaborative technology, Proc. 2017 IEEE Int. Conf. Appl. Syst. Innov. Appl. Syst. Innov. Mod. Technol. ICASI 2017, pp. 17851788, 2017. [15] A. Willem and M. Buelens, Knowledge sharing in public sector organizations: The effect of organizational characteristics on interdepartmental knowledge sharing, J. Public Adm. Res. Theory, vol. 17, no. 4, pp. 581606, 2007. [16] I. Reychav and J. Weisberg, International Journal of Information Management Going beyond technology: Knowledge sharing as a tool for enhancing customer-oriented attitudes, vol. 29, pp. 353361, 2009. [17] D. Stenmark, Leveraging tacit organizational knowledge, J. Manag. Inf. Syst., vol. 17, no. 3, pp. 924, 2000. [18] D. R. Garrison, Critical Inquiry (Garrison Anderson Archer 2000), vol. 2, pp. 119, 2000. [19] D. R. Garrison, T. Anderson, and W. Archer, The first decade of the community of inquiry framework: A retrospective, Internet High. Educ., vol. 13, no. 12, pp. 59, 2010. [20] D. R. Garrison and J. B. Arbaugh, Researching the community of inquiry framework: Review, issues, and future directions, Internet High. Educ., vol. 10, no. 3, pp. 157172, 2007. [21] P. Dillenbourg, P. Dillenbourg, P. Dillenbourg, and P. Dillenbourg, what do you mean by collaborative learning? pp. 119, 2007. [22] M. Laal and S. M. Ghodsi, Benefits of collaborative learning, Procedia - Soc. Behav. Sci., vol. 31, no. May 2014, pp. 486490, 2012. [23] P. I. K. U. Indonesia, Pusilkom UI Profile. [Online]. Available: http://www.pusilkom.ui.ac.id. [Accessed: 01-Dec-2018]. [24] F. of C. S. U. Indonesia, Pusilkom Universitas Indonesia- Services. [25] M. Carr, Prototyping and Software Development Approaches, Prototyp. Softw. Dev. Approaches, no. 3, pp. 116, 2004. [26] J. D. Naumann and A. M. Jenkins, Prototyping: The New Paradigm for Systems Development, MIS Q., vol. 6, no. 3, p. 29, 1982. [27] K. Ferdig, R. E. & Trammell, Content Delivery in the Blogosphere. [28] B. D. Z. Levin, R. Cross, L. C. Abrams, and E. L. Lesser, Trust and knowledge sharing.pdf. [29] A. Jimoyiannis and P. Tsiotakis, beyond students’ perceptions: investigating learning presence in an educational blogging com- munity, J. Appl. Res. High. Educ., vol. 9, no. 1, pp. 129146, 2017. 17
2019 4th International Conference on Information Technology (InCIT2019) Sugar Cane Grading from Photos Using Convolutional Neural Networks Phuong Pham Sally E. Goldin Data Engineering Department of Computer Engineering Société Générale King Mongkut’s University of Technology Thonburi Paris, France Bangkok, Thailand tmppham@icloud.com sally_e.goldin@kmutt.ac.th Abstract— Sugar manufacturing companies need up-to-date This paper describes our approach and results, as well as the information about sugar cane condition in order to forecast obstacles we faced in this effort. sugar yield. However, the wide spatial distribution of cane fields in Thailand makes exhaustive surveys of sugar cane Over the past half-decade, CNNs have been successfully condition impractical. In this paper, we report on initial efforts applied to a wide range of image recognition and to apply convolutional neural networks (CNN) to classify sugar classification problems, for instance to classify food items cane condition based on ground level photographs of sugar [7], car models [8], plankton species [6], galaxy morphology cane fields at different stages of the growing season. Our long [3] and natural scenes [5]. Recently, researchers have used term goal is to create a system where farmers can submit CNN for recognizing plant types [13], plant diseases [4][14], mobile phone photos of their fields that will be evaluated for and general crop health [1]. Overall accuracy for these cane quality using a machine learning (ML) model. Results agriculturally-related studies has ranged from 55% to 99%, suggest that this approach has promise, but that we need a with performance strongly related to the size of the input larger set of exemplars to create a model that can provide data set and the degree of visual variation among the input classification performance accurate enough for operational use. images. Tens of thousands of input images are typically required for successful classification. Keywords—convolutional neural network, machine learning, sugar cane, image classification, agricultural information system II. METHODOLOGY A. Data Set I. INTRODUCTION The data set consists of 2281 ground level digital photographs, gathered by sugar company personnel during Sugarcane is an important commercial crop in Thailand. their field surveys. The images were categorized by these In order to forecast sugarcane yield, sugar companies need personnel into three cane quality classes: good (1141), detailed information on the cane condition (cane health) at medium (980), and poor (160). Within each cane quality various stages in the growth cycle. Visible features such as class, the images are grouped by growing season. Each color, leaf density and cane height are some indicators of photo represents a different cane field. Table I summarizes cane condition. Since properties of a cane field such as soil the distribution of the data set. characteristics and drainage as well as microclimate and cultivation practices influence cane condition, the cane TABLE I. DATA SET DISTRIBUTION quality and corresponding visual features vary from one field to another. Season There are more than 100 million hectares of sugar cane Class Early Mid Late All fields in Thailand [17]. Most fields are cultivated by small- Good 381 1141 scale individual farmers under contract to sugar companies. Average 289 576 184 980 These companies cannot do exhaustive field surveys to get Poor 75 160 the information that they need. Currently they visit a sample All 745 423 268 2281 of 10-20% of the fields annually, taking photos and measurements in order evaluate cane condition. However, 54 31 this practice does not provide sufficient data to support reliable yield forecasts. 1053 483 This research investigated the utility of machine learning Since the images were taken by different people at to address to this problem. Working in co-operation with a different times, they vary in size and appearance (average large sugar cane company in Thailand, we created a software dimension 1600x1200 pixels). Although this variability is test bed to analyze sugar cane health from photographs of expected to reduce learning performance, it is realistic since the fields taken at ground level. Specifically, we trained the model is intended to classify field photos taken by many convolutional neural networks (CNN) to discriminate photos different farmers. Fig. 1 shows some representative images based on cane health. The ultimate goal is to build an from the poor, medium and good categories during mid- operational system that will allow individual farmers to season. The images look quite similar, highlighting the fact upload geotagged mobile phone photographs of their fields that this is a challenging classification task. which would then be automatically classified to get more detailed information about cane conditions over a wide area. 18
2019 4th International Conference on Information Technology (InCIT2019) size of the input data set. This is a strategy where original images are transformed by cropping, histogram modification, rotation, flipping or other operations, in order to produce additional inputs. In this study we used cropping, extracting square subsets from the bottom central, bottom left and/or bottom right areas of each image. Cropping focused on the bottom because, in most cases, the top of the image contained sky. Depending on the experiment, we used original images, cropped images, or both, for training and for testing. A 2:1 hold-out strategy was used to select training and testing data sets. Training sets ranged from 540 to 6072 input images. Fig. 1 Images of poor, medium and good quality cane (left to right) at 7-9 Input Image Size. To speed up training, input images for months into the growing season. a CNN are usually downsampled to a consistent square size before they are submitted to the network. Following the B. Software Tools and CNN Configuration literature, the original and cropped images were resampled Based on preliminary testing, we chose the Torch to 32x32 pixels in most experiments. However, one framework [16] for building and running our models. Torch experiment investigated the effects of increasing the was easier to install, provided better performance, and had downsampled input image size to 128x128. better documentation than Tensorflow [15] or Caffe [2]. All training and testing reported here used a server provided by Segregation by Growing Season. Initial experiments used the KMUTT High Performance Computing group, which images from early, mid and late growing season periods hosts an NVIDIA K10 Tesla GPU. together. Further experiments explored the effects of training and testing each growing season separately. A CNN is a multilayer neural network consisting of one Network Parameters. CNN machine learning or more convolutional layers, as well as (possibly) pooling performance is sometimes sensitive to network parameters layers, Rectified Linear Unit (ReLU) layers, and fully such as the number and type of layers, the kernel size and connected layers. Convolutional layers are used to extract stride, and so on. This work examined two potentially important features from the input images by convolving important variables, the number of convolutional layers and them with small filter kernels. The features learned by a the number of iterations before stopping the training process. convolutional layer are usually summarized by a pooling Theoretically, either of these variables could improve layer. A ReLU layer zeroes any negative output values classification accuracy, the former by extracting more and produced by the pooling layer preceding it. The learned more useful features, the latter by providing more training features are eventually passed into fully connected layers, \"experience\" to the network. where each input image is mapped to a suitable output class [9]. III. RESULTS We ran many experiments involving the above variables, This research adopted the popular Lenet 1 CNN trying to maximize classification accuracy. Since this configuration [10]. The characteristics of this network research focused on solving a practical problem rather than configuration, which was developed for the MNIST data on an exhaustive exploration of possibilities, it did not classification problem (handwritten digit recognition), are necessarily adopt a complete orthogonal design in testing summarized in Table II. This basic structure was used for all these variables. Overall, the work explored dozens of experiments, while varying a few of the parameters in some different conditions. This short paper summarizes the most cases. For instance, one experiment increased the number of important conclusions from the more extensive research convolutional layers to see if that would improve reported in [12]. performance. A. Preliminary Classification Results In all cases, a simple percent correct metric (true TABLE II. LENET 1 NETWORK CONFIGURATION positives divided by total test cases) was used to evaluate the performance of the models. Results based on the more Layer Type Kernel Size/ Output complex F metric followed the same patterns. The first Stride Channels experiment trained and tested the Lenet 1 with original Convolution 1 5x5/1 16 sugarcane images resampled to 32x32 pixels, using a Max Pooling 1 2x2/2 16 learning rate of 0.001 and 15 iterations, with a 2:1 hold-out ReLU 1 (NA) 16 strategy for selecting training data. The resulting overall Convolution 2 5x5/1 20 accuracy was 53.86% with 43.38% of the true positives in Max Pooling 2 2x2/2 20 class “good”, 10.48% in class “medium” and 0% in class ReLU 2 (NA) 20 “poor”. Although this accuracy was disappointing, the model Fully Connected 1 Flatten clearly had learned some information, since the results were Fully Connected 2 120 neurons Flatten notably higher than chance level. Fully Connected 3 84 neurons Flatten We judged that the failure for class \"poor\" was due to the 3 output classes very small number of training images available for this category (160). Therefore, all following experiments omitted C. Experimental Variables images labeled \"poor\" in both training and testing. Training and Testing Images. Given the small number of Eliminating the \"poor\" images raised the overall percent images available, we used data augmentation to increase the correct from 53.86% to between 55.28% and 60.92%, depending on which images (original, cropped or both) were used for training and testing. 19
2019 4th International Conference on Information Technology (InCIT2019) The learning rate in subsequent experiments was also Segregation by Growing Season. Models trained and increased to 0.005, since we discovered while exploring the tested with mid-season images had some tendency to situation with \"poor\" images that this consistently produced provide more accurate results, but this was not true in all slightly higher accuracies. cases. B. Factors That Improved Accuracy D. Comparison with Human Judges Although the initial classification showed better than As we reviewed our results, we found that none of our chance performance, the overall model accuracy was not experiments resulted in percent correct classification better sufficient to support an operational system. We considered than 67.9%. (This accuracy was achieved in one experiment various hypotheses regarding the unimpressive performance, reported in [12], which used only mid-season images, square then designed experiments to evaluate these hypotheses, aspect ratios for the input images, and 50 iterations.) Given varying different aspects of the input data, training data their visual similarity, one could question whether the and/or network configuration, as summarized in section II.C. labeled images actually contained sufficient information for This section describes the main factors that had a positive a reliable classification. The labels were assigned by the impact on performance. human staff who physically surveyed the cane fields as well Original versus Cropped Images. One robust result that as taking photos. Perhaps they classified the fields using appeared in many experiments was that accuracies tended to additional information not captured in the photographs. be higher when the training data included both original and To test this hypothesis, we conducted an experiment with cropped (subset) images, while the test data was original human judges, using a web application presented in the Thai images. This condition produced an overall accuracy of language. Our goal was to assess how accurately people 66.78% for mid-season images only, 64.93% for mid- and could classify the same photos supplied to the CNN. Ninety late-season images. photos, 30 from each class, were selected from the full data Cropped Image Size. In the initial experiments, the size set. Each time a new user entered the application, the of cropped subsets was set to 200 x 200 pixels. To program randomly chose 10 images from each class, investigate if using larger subsets would include more displayed them randomly, and asked the user to label each information and result in better accuracy, the cropped size to one as good, medium or poor, with no time limit and no 400 x 400. This change required excluding some of the ability to go back and reclassify earlier images. original images which were too small or had the wrong aspect ratio to support the larger cropped size, reducing the Table III summarizes the results from this experiment. overall data set (medium and good classes only) to 1809 images. TABLE III. HUMAN SUBJECTS’ PERFORMANCE Test runs with 400 x 400 cropped size generally produced higher accuracies than with 200 x 200 cropped Subject Experience (years) Accuracy size. The best performance for the 400 x 400 case was for (% correct) mid-season images (66.78%). The biggest improvement seen 12 with the larger cropped size was for mid-season and late- 22 50.0 season images (57.44% versus 62.87%). 31 73.3 41 66.7 C. Factors That Did Not Improve Accuracy 54 50.0 The experiments above produced modest but consistent 61 43.3 improvements in accuracy. Some variables, on the other 63.3 hand, did not have a consistent positive effect. Experts Average 57.8 Input Image Size. The original images average 70 56.7 1600x1200 pixels. These images were resampled to 32x32 80 43.3 pixels before input to the CNN. To test if this extreme down- 50.0 sampling might cause enough information loss to negatively Non-experts Average 55.8 affect the accuracy, we trained and tested models with input Overall images of both size 32x32 and 128 x128. In contrast to our expectations, the prediction accuracies when using input size Subjects 1 through 6 were employees of the sugar 128x128 were not very different from, and always lower company who are familiar with sugarcane growing. Subjects than, the accuracies when using input size 32x32. 7 and 8 were undergraduate students with no knowledge Number of Iterations. Increasing the number of iterations about sugarcane. This table supports a number of interesting (presentations of the training set) to 50 did not produce any conclusions. The average accuracy of human classification consistent improvement in accuracy. In many cases, the with three classes is 55.8%, which is 22.8% above the accuracy decreased when comparing 15 to 50 iterations, expected random performance. This average result is close possibly due to overfitting. to the best result from the CNN classifier with three classes Increased Number of Convolutional Layers. Some (56%). On the other hand, the best result from human studies in the literature (e.g. [9]) have used many more classification is 73.3%, which is 9.4% higher than the best layers than in the configuration adopted in this work. To result of the CNN classifier. This suggests that the dataset examine whether more convolutional layers might improve might contain additional useful information that the model classification performance, we created a network could learn. Performance did not particularly correlate with configuration with 4 convolutional layers (plus sugar cane experience. Finally, examining per-class accuracy accompanying max pooling and ReLU layers) to compare (not shown), made it clear that results for the “poor” class with the 2 layer configuration used in most experiments. In were equivalent to “good” and “medium”, supporting the general, the 4 layer model provided poorer performance than theory that the CNN model's inferior performance on “poor” the 2 layer model. images was due to insufficient training data. IV. CONCLUSIONS The results of this research indicate that deep convolutional neural networks are capable of classifying sugarcane health with some limitations. The best CNN 20
2019 4th International Conference on Information Technology (InCIT2019) classifier performance with two classes is 67.9%, which is recognition”. Sensors (2017) Vol. 17. https://doi:10.3390/s17092022. 18% above expected random performance. The best CNN Accessed 18 September 2018. classifier performance with three classes is 56%, which is 23% above random performance. These results are not [5] M. George, M. Dixit, G. Zogg, and N. Vasconcelos, “Semantic accurate enough to be applicable to the real world problem clustering for robust fine-grained scene recognition”. Proceedings of of sugarcane health assessment. However, they show that it the European Conference on Computer Vision (ECCV), (2016). is possible to classify sugarcane images into different health https://arxiv.org/abs/1607.07614. Accessed 18 September 2018. categories with CNN. [6] I. Goodfellow, Y. Bengio, and A. Courville. Deep Learning. There is no previous best result in the literature on the Cambridge: MIT Press, 2015. same dataset to compare with ours. Nevertheless, there are some previous studies on fine-grained image classification. [7] H. Kagaya,, K. Aizawa and M. Ogawa, “Food detection and Table III compares our results with those from several recognition using convolutional neural network”. (2014) Proceedings studies in the literature. In fact our model is nearly as of the 22nd ACM International Conference on Multimedia. accurate as these comparison models, despite the very small https://doi.org/10.1145/2647868.2654970. Accessed 18 September number of input images. 2018. In the experiment with human judges, one of the eight [8] J. Krause, J. Gebru, J. Deng, L. J. Li, and F.F. Li, “Learning features subjects was able to categorize sugarcane images into three and parts for fine-grained recognition”. Proceedings of the 22nd different health categories with the accuracy of 73%. This International Conference on Pattern Recognition. IEEE. (2014) indicates that there could be more information contained in https://doi.org/10.1109/ICPR.2014.15. Accessed 18 September 2018. the images than what the current CNN models have learned. [9] A. Krizhevsky, I. Sutskever, and G.E. Hinton, “ImageNet classification with deep convolutional neural networks”. Proceedings of the 25th International Conference on Neural Information Processing Systems - Volume 1. ACM (2012) [10] Y. LeCun, C. Cortes, and C.J.C. Burges, “The MNIST database of handwritten digits. http://yann.lecun.com/exdb/mnist/ Accessed 18 September 2018. TABLE IV. COMPARISON OF ACCURACIES IN FINE-GRAINED CNN [11] A. van den Oord, I. Korshunova, J. Burms, J. Degrave, L. Pigou, and CLASSIFICATION P. Buteneers “Classifying plankton with deep neural networks”. National Data Science Bowl (2015) Task Domain Accuracy # Input Images http://benanne.github.io/2015/03/17/plankton.html. Accessed 18 Indoor scenes [5] 68.2 % 6,700 September 2018. Car models [8] 70.5% 16,185 Sugar cane (all seasons) 63.0% 2,013 [12] T.M.P. Pham, “Sugar cane grading from photo using machine Sugar cane (mid season) 67.9% 811 learning”. Unpublished senior thesis, Department of Computer Engieering, King Mongkut’s University of Technology Thonburi (2017). [13] S. Razavi, and H. Yalcin, “Using convolutional neural networks for plant classification”. Proceedings of the 25th Signal Processing and Communications Applications Conference. IEEE (2017). https://doi.org/10.1109/SIU.2017.7960654. Accessed 18 September 2018. One issue we face involves criteria for evaluating and [14] S. Sladojevic, M. Arsenovic, A. Anderla, D. Culibrk, and D. comparing the performance of different models. We can Stefanovic, “Deep neural networks based recognition of plant diseases observe from the experiments that the models’ accuracies by leaf image classification”. Computational Intelligence and often vary only slightly from one another. It is difficult to Neuroscience. (2016) Vol. 2016, Article ID 3289801, determine how big a difference in accuracy should be http://dx.doi.org/10.1155/2016/3289801. Accessed 18 September considered to be \"real\" and important. 2018. Our biggest challenge in this research was the small and [15] “Tensorflow: An open source machine learning framework for imbalanced dataset. We had only 2281 images, with only everyone”. https://www.tensorflow.org/ Accessed 18 September 2018. 160 images in the \"poor\" class. For the future work on this problem, we suggest that more initial training images, [16] “Torch: A powerful computing framework for LuaJit. http://torch.ch/ ideally about 10,000 per class, should be obtained in order to Accessed 18 September 2018. build an accurate and effective classifier for this problem. [17] P. Weerathaworn, “Sugar industry in Thailand”. National Research Council of Thailand. Presentation at Agri Benchmark Cash Crop Conference, Goiânia, Brazil (2015) http://www.agribenchmark.org/cash-crop/conferences-and- events/2015-brazil.html Accessed 20 September 2018. ACKNOWLEDGMENT We gratefully acknowledge the assistance of Mitr Phol Sugar Corporation in providing data and advice for this project. REFERENCES [1] P. Bhatt, S. Sarangi, S. Pappula., “Comparison of CNN models for application in crop health assessment with participatory sensing”. Proceedings of the IEEE Global Humanitarian Technology Conference. IEEE (2017). https://doi.org/10.1109/GHTC.2017.8239295. Accessed 18 September 2018. [2] Caffe Deep Learning Framework. http://caffe.berkeleyvision.org/ Accessed 18 September 2018. [3] S. Dieleman, K.W. Willett, and J. Dambre, J, “Rotation-invariant convolutional neural networks for galaxy morphology prediction”. Monthly Notices of the Royal Astronomical Society. Vol. 450 (2), 1441-1459 (2015) https://doi.org/10.1093/mnras/stv632. Accessed 18 September 2018. [4] A. Fuentes, S. Yoon, S.C. Kim, and D.S. Park, “A robust deep- learning-based detector for real-time tomato plant diseases and pests 21
2019 4th International Conference on Information Technology (InCIT2019) Recommender System Based on User Evaluations and Cosmetic Ingredients Yoko Nakajima Hirotoshi Honma Haruka Aoshima Department of Creative Engineering, Department of Creative Engineering, Department of Computer Science and Engineering, National Institute of Technology, National Institute of Technology, Toyohashi University of Technology, Kushiro College, Kushiro College, Kushiro, Japan, Kushiro, Japan, Toyohashi, Japan, h.aoshima.tut@gmail.com yoko@kushiro-ct.ac.jp honma@kushiro-ct.ac.jp Tomoyoshi Akiba Shigeru Masuyama† Department of Computer Department of Computer Science and Engineering, Science and Engineering, Toyohashi University of Technology, Toyohashi University of Technology, Toyohashi, Japan Toyohashi, Japan akiba@cs.tut.ac.jp masuyama@tut.jp Abstract—This study considers the compatibility between mend the right product to the right person at the right time. users and basic skin care products based primarily on the Moreover, Moe and Aung [2] present a new approach for products’ ingredients. We have developed a product recom- building ontologies using Taxonomic conversational case- mender system that is expected to recommend products that based reasoning to apply cross-domain recommendation provide the desired cosmetic effect for different user groups based on facial skin problems and related cosmetics. They depending on age and skin type. From the cosmetics review showed their system is user-friendly and more accurate site, based on an analysis of user evaluations, we extracted than the other related works and gives more personalized the names of ingredients that are thought to have the best recommendations and makes more profits for commercial effect and developed a method to recommend products that sites. include these ingredients as their main ingredients. We pro- pose the ingredient frequency-inverse product frequency (IF- We focused on the fact that we purchased cosmetics IPF) method to derive ingredients characterizing strong-effect based on the comments and evaluations of others. For product group. We have defined the scale “the recommended example, when a user searches for skin lotions on a product satisfaction level” to evaluate the effectiveness of @cosme website, the user reads reviews posted by users our recommendation service. As a result, our system can who have similar attributes, such as age and skin type, recommend products with a high degree of serendipity and and can also identify items that are similar to the skin hidden attraction, among others. lotion that the user uses currently. For example, a user with dry skin will search for lotion with a high moisturizing Index Terms—natural language processing, cosmetics rec- effect, and a user who is interested in skin whitening will ommender system, user review information, cosmetic ingredi- select products that have been reviewed positively for skin ent information, knowledge acquisition whitening. However, different basic cosmetics may have different effects; thus, finding products that are compatible I. INTRODUCTION with a given user is difficult, even when using review information from @cosme. Conventionally, review sites’ In recent years, Web services have become increasingly recommendation services are primarily based on collabo- popular, and consumers have frequently referred to user rative filtering [3], [4]. They rely on recommender systems product reviews while making purchases. Kakaku.com1, based on user similarity and browsing page. Recommender Minna-no-Cinema2, and @cosme3 are popular product re- systems that use collaborative filtering are predominantly view sites that focus on home electronics, movies, and based on a black box approach that does not reveal the cosmetics, respectively. The number of cosmetic review internal mechanisms of the recommendation procedure. sites, users, and products have been increasing every year. Consequently, the reasons for a particular recommendation are unknown [5]. A recommendation function is also imple- Many cosmetic recommendation systems have been de- mented in the @cosme website; however, the reason why veloped so far. Wang et al. [1] constructed a personalized a particular product is recommended is not disclosed. recommender system which incorporates content-based, collaborative filtering, and data mining techniques to have This study considers the compatibility between users and an effective command of the relationship between cus- basic skin care products based primarily on the products’ tomers and products. They showed the system can recom- ingredients. We have developed a product recommender system that is expected to recommend products that pro- †Currently with Department of management, School of Management, Tokyo University of Science, Tokyo, Japan. 1Kakaku.com: http://kakaku.com 2Minna-no-Cinema: https://www.jtnews.jp 3@cosme: http://www.cosme.net/ 22
2019 4th International Conference on Information Technology (InCIT2019) vide the desired cosmetic effect for different user groups B. Bihada–Mania Website depending on age and skin type. From the cosmetics review site, based on an analysis of user evaluations, we extracted Manufacturers are not obliged to disclose product ingre- the names of ingredients that are thought to have the best dients on their website. Users must purchase the product effect and developed a method to recommend products and review the product packaging to determine the in- that include these ingredients as their main ingredients. gredient information. Therefore, acquiring comprehensive In this system, basic skin care lotions were used as the ingredient information for all the products is difficult. target products. Lotions are one of the most commonly used cosmetic products. They are applied directly to the The Bihada–Mania website responds to users who want skin; therefore, the product quality and compatibility with information regarding the ingredients of products. Bihada– skin type are important. Several recommendation methods Mania is the largest site in Japan that publishes cosmetic have utilized user review information [6]–[9]; however, to ingredient information and toxicity reports. As of December the best of our knowledge, focusing on the ingredients in 2018, ingredient information had been posted for 31,000 cosmetics has not been considered previously. products. Bihada–Mania, as shown in Table I, presents both ingredient and toxicity information. II. @COSME AND BIHADA-MANIA SITES TABLE I INGREDIENT AND TOXICITY INFORMATION In this section, we will introduce the two sites that have been considered in this research. Ingredient Toxicity information A. @cosme Website Glycyrrhizinate Dipotassium(GD) Flavoring, Anti-Inflammatory Water Solvent The @cosme website is the largest scale cosmetic web- Alcohol Solubilizer, Astringents Butylene Glycol(BG) Moisturizing, Solvent, Viscosity reducing site in Japan. As of June 2018, the site listed 33,000 Glutathione Flavoring, Whitening Oleic Acid Flavoring, Washing brands and 310,000 products and had 5,000,000 users and Lysine HCl Skin conditioning Aesxulus Hippocastanum Antioxidant, Astringents approximately 14,000,000 reviews. The number of reviews Stearyl Glycyrrhetinate Flavoring, Skin conditioning Orange Fruit Extract Moisturizing, Astringents posted have been increasing every year. Typically, a user Tocopheryl Acetate Antioxidant, Promoting blood circulation Ascorbyl Phosphate Na Whitening, Antioxidant Hydrogenated Castor Oil Surfactants reviews the purchase process of a product and its usage. Users provide personal information, such as age and skin quality, when they register. III. PROPOSED METHOD Tags can be used to evaluate products. Users can select We believe that level of compatibility between users and basic skin care products is strongly dependent on the from a range of effect tags and can add several tags. The ingredients in those products. Based on that premise, we at- tempted to develop a recommender system that can provide type and the number of selectable tags differ depending on the desired cosmetic effect, e.g., “moisture,” “whitening,” “pores,” and “acne,” for each user group differentiated by the product category. The following fifteen types of effects age and skin type. Based on an analysis of user evaluations and cosmetic ingredients, we extracted the names of ingre- tags can be used while posting body lotion reviews. dients presumed to best provide a desired cosmetic effect from a cosmetic review site and recommended products that The cosmetic effect tag contained these ingredients as their primary ingredients. A graphical summarization of the proposed method is Moisture, Pores, Acne, Whitening, Mild stimulation, presented in Fig. 2. Anti-Aging, UV cut effect, Skin elasticity, Refreshing, Horny care, Oily, Tightening, Cleansing, Cost perfor- mance, and Organic cosmetics. Along with assigning tags, users can post reviews that do not exceed 2,000 characters. An image of a review posted on @cosme is shown in Fig. 1. Fig. 1. User review posted on @cosme site Fig. 2. A graphical summarization of the proposed method The process flow for recommending lotion with a high level of “moisture” as a cosmetic effect will be described. Input: user age, skin type, cosmetic effect (“moisture”) 23
2019 4th International Conference on Information Technology (InCIT2019) Output: recommended product group with high level of we divided evaluation comments for all products into those “moisture” with and without the cosmetic effect tag “moisture.” Herein, for each comment group, we developed a cosmetic effect Step 1: Define a user group (user attributes) of similar word set from words that appear at high frequency in age and skin type. the comments with the cosmetic effect tag attached. If a cosmetic effect word from the set is included in an Step 2: Extract reviews from the @cosme site for all cos- evaluation comment, the “moisture” cosmetic effect of that metic products that fit each of the user attributes. product can be considered strong. In addition, several words were added manually to the cosmetic effect word set despite Step 3: Acquire products with a high evaluation as ef- not appearing with great frequency because they clearly fective ”moisturizers” with a strong-effect product suggest a strong “moisture” cosmetic effect. group X. The following demonstrates an example of the use of Step 4: Extract ingredient groups that characterize the cosmetic effect word set in this experiment to delineate a strong-effect product group X. “moisture” cosmetic effect. Step 5: Output a recommended product group that has “moisture” cosmetic effect word set the ingredient groups acquired in Step 4 as its primary ingredients. (moisture), (be moistened), (be filled), As described previously, lotions can be selected based on (fresh), (freshness), (freshly), the expectation of having a “moisturizing” effect; however, in addition, it is also possible to discern lotions that have (moisturizing), (penetration), (moist), expected cosmetic effects for “whitening,” “pores,” and “acne” independently, as well. In the following subsections, (plump), (smooth), (familiar with we will provide detailed descriptions and definitions for each step in the process flow. the skin), (absorb), (primordial cell), A. Step 1 (Defining User Attributes) (not dry), (basic skin strength), On @cosme, users provide information about their age (collagen), (soft skin, smooth skin), and skin type when they register as members. Herein, we used this information to define user attributes. As user (Elastin), (nature), (moisturize skin), attributes, we identified six skin types, i.e., normal, oily, mixed, dry, sensitive, and atopic, and four age groups, i.e., (beauty essence), (Polyglutamic acid), teens, 20s, 30s, and 40s4, which were combined to make a total of 24 classes. (nano), (Proteoglycan), B. Step 2 (Usage Data) (wrap), (firmness), (rejuvenation, rejuvenate), We limited the product category for recommendations to (fresh, active), (penetrate), lotions for basic skin care, which is the most commonly used type of lotion. Since lotion is applied directly to the (charge), (slightly), (retention), skin, compatibility with skin type is important. Currently, information posted on @cosme and Bihada–Mania includes (lift up), (soft), (hyaluronic acid), 15,484 and 2,392 lotions, respectively. In this study, we used user comments from @cosme and ingredient infor- (active oxygen), (nurture), (Moisten), mation from Bihada–Mania. Herein, we utilized product data for 2,041 lotions that were listed on both sites. In this (deep in skin), (nasolabial), step, the reviews of these 2,041 lotions are classified based on each user attribute. (ceramide), (stratum corneum, horny layer), C. Step 3 (Strong-effect Product Group) C(concentration vitamin C), (amino acid), We extracted products evaluated as having a strong ,, , ,, “moisture” cosmetic effect from reviews of each lotion categorized by user attribute and users with the same user , , ,, , attributes. Herein, the set of extracted products is defined as the strong-effect product group X. To determine the (onomatopoeia) “moisture” cosmetic effect of each lotion, we used the cosmetic effect tag users attributed to a product when reviewing the product on @cosme and user evaluation comments. Strong-effect product group X is a set of products that appear to have a strong cosmetic effect corresponding to To evaluate cosmetic effect using evaluation comments, each user attribute. A strong-effect product group for a a cosmetic effect word set was prepared using words that “moisture” cosmetic effect for user attributes [20s, normal may indicate a strong “moisturizing” cosmetic effect. First, skin] is constructed as follows. For all reviews of lotions from all those with same user attributes, the total of the 4Extremely few data are available for age group of over 50s and we “moisture” cosmetic effect tag attribution and cosmetic delete them from consideration. effect word set usage rates in the user comments were obtained, and the top 10% of the products became strong- effect product group X. Herein, if the total number of products in the top 10% for these values is less than five, the top five products become the strong-effect product group. Table II lists the number of lotion products corresponding to reviews from each user attribute and the number of products in the strong-effect product group generated in Step 3 for the “moisture” cosmetic effect. In this example, for user attributes [20s, normal skin], there are 592 target products. Among these products, the top 10% includes 60 products that comprise the strong-effect product group. However, for user attributes [40s, atopic skin], there are only 19 target products; thus, the products with the top five evaluation values comprise the strong-effect product group. D. Step 4 (IF-IPF Method) Here, we determine the ingredients that characterize the lotions of strong-effect product group X for each of the user attributes constructed in Step 3. Typically, the lotion ingredients particular to a strong-effect product group X 24
2019 4th International Conference on Information Technology (InCIT2019) TABLE II differed greatly for the teens and 40s groups, even for the NUMBER OF PRODUCTS IN STRONG-EFFECT PRODUCT GROUP FOR sensitive skin type. Therefore, it can be confirmed that a highly effective ingredient varies significantly based on the “MOISTURE” EFFECTT user attributes and whether it is distinctive. Age Normal Oily Mixed Dry Sensitive Atopic E. Step 5 (Derivation of Recommended Products) teens 15/148 17/163 25/249 15/148 15/144 5/43 We determine recommended products that use the highly 20’s 60/592 44/436 97/969 78/771 69/687 23/223 effective ingredient corresponding to each user attribute 30’s 56/551 25/247 100/992 86/859 67/660 14/129 acquired in Step 4. Herein, for all of the lotion products, 40’s 25/242 46/457 45/446 24/237 we obtained the IF-IPF values for their ingredients in 5/44 5/19 advance. Among all the lotion products, the recommended products became those with one or more strongly effective are thought to be the ingredients most prevalent in the ingredients corresponding to the same user attributes where products of X. However, over 90% of the ingredients they were shown to be ingredients with the top five IF-IPF contained in lotion are water and BG (Butylene Glycol), so values. these cannot be said to be characteristic of the products in X. Therefore, we propose the ingredient frequency-inverse Table IV lists the number of recommended products with product frequency (IF-IPF) method to derive ingredients high moisture effect for each user attribute. characterizing strong-effect product group X. The IF-IPF method inspired by the term frequency-inverse document IV. VERIFICATION EXPERIMENT frequency (TF-IDF) method, one of the methods used to process the natural language. The IF-IPF method is The proposed system was evaluated experimentally, and described as follows. we determined recommended products targeting a “mois- ture” cosmetic effect for each user attribute to verify IF (Ingredient Frequency) accuracy. Ingredient frequency (IF) is the proportion of ingredient A. Verification Method i contained in the lotions of strong-effect product group X. There is a rule that requires cosmetic ingredients be In the recommender system, the degree to which a rec- displayed in descending order of their amount in the ommended product satisfies the user is the most important product; thus, IF, as defined by Equation (1), lists them index. The optimal verification method to evaluate the in order of the weight of the ingredients. recommendation services herein is the online estimation method. It takes a great deal of time and expense to IF (i, X) = ∑ np − αp,i (1) assess the predictive accuracy by allowing collaborators np corresponding to each set of user attributes to use the p∈X recommended products for a set duration. Instead, we defined product satisfaction to determine whether a rec- np : The number of ingredients in product p in skin ommendation service is effective. lotion group X. For the entirety of the reviews that user attributes A αp,i : Rank of ingredient i listed in product p. submitted for a certain product P, the ratio of the assigned effect tag “moisture” for product P by user attributes A was IPF (Inverse Product Frequency) defined as product satisfaction. Furthermore, the average IPF is the inverse of the total number of products that level of satisfaction for the recommended products for user attributes B was defined as the recommended product sat- contain ingredient i. In many products, the main ingredi- isfaction level. A higher recommended product satisfaction ents, i.e., water and BG, have a very low calculated IPF level indicates that the system demonstrates higher recom- value. IPF is defined in Equation (2). mendation accuracy. For comparison, the average product satisfaction value for products that were not recommended N (2) (nonrecommended product satisfaction level) was used to IP F (i) = log determine accuracy. pf (i) B. Experimental Results N : The number of skin lotion products. Table V lists the levels of recommended product sat- pf (i) :The number of products including ingredient i. isfaction for the proposed recommended products corre- sponding to a “moisture” cosmetic effect. Although there is IF-IPF Value some variance depending on the user attributes, for many The IF-IPF value is derived as a product of IF and IPF results, recommended product satisfaction exceeded that of nonrecommended product satisfaction. For atopic skin as given by Equation (3). category, although there is a significant difference, this is considered a result of support for products that contain anti- IF -IP F (i, X) = IF (i, X) × IP F (i) (3) inflammatory ingredients. In addition, compared to other user groups, it can be implied from the data that users with In Table III, with respect to the user attributes for all atopic skin prefer products that reduce skin inflammation generations of normal and sensitive skin, the IF-IPF value and have a “moisture” effect. On the contrary, mixed or indicates that the calculated ingredient is among the top five ingredients. For each user attribute, the five derived ingredients can be considered highly effective ingredients. From the same age groups, i.e., 20s with normal skin and 20s with sensitive skin, the aesthetic ingredient also differed greatly. In addition, the effective aesthetic/beauty ingredient 25
2019 4th International Conference on Information Technology (InCIT2019) TABLE III INGREDIENT CALCULATED AS IN TOP FIVE RELATIVE TO USER ATTRIBUTES FOR ALL GENERATIONS OF NORMAL AND SENSITIVE SKIN Age Normal 2.73 Sensitive 3.77 teens 2.64 2.80 20’s Raffinose 2.45 Tranexamic acid 2.48 30’s DPG 2.37 Erythritol 2.42 40’s PEG-20 Sorbitan Isostearate 2.23 Methyl Gluceth 2.34 Niacinamide 10.01 Dipotassium Glycyrrhizate 10.95 Glycereth-26 8.26 Diphenylsiloxy Phenyl Trimethicone 10.12 Betaine 8.19 Sodium Hyaluronate 8.31 PEG 7.39 Pentylene Glycol 8.09 Sodium Hyaluronatea 7.21 DPG 8.08 Pentylene Glycol 9.65 Hydrolyzed Collagen 11.70 DPG 7.83 Betaine 10.70 PCA-Na 7.49 DPG 10.14 DPG 7.43 Diglycerin 9.55 Betaine 7.40 Sodium Hyaluronate 8.84 PG 5.64 Pentylene Glycol 3.75 Dipotassium Glycyrrhizate 3.99 Hydrolyzed Collagen 3.68 Sodium Hyaluronate(2) 3.86 DPG 3.45 Pinus Pinaster Bark Extract 3.76 Morus Alba Root Extract 3.44 Dipotassium Glycyrrhizate 3.60 Ubiquinone 3.39 Sodium Acetylated Hyaluronate Pentylene Glycol Hydrolyzed Collagen Polyglutamic Acid TABLE IV recommended products used as the base for the strong- NUMBER OF RECOMMENDED PRODUCTS WITH HIGH MOISTURE effect product group. Furthermore, when there is a small number of users providing evaluations, it can be considered EFFECT that a review from a particular user might have an great impact on the entire system. Age Normal Oily Mixed Dry Sensitive Atopic For lotion evaluations, it is important to appropriately teens 19 13 43 28 73 56 select the desired cosmetic effect, and it has been found 20’s 23 27 22 42 37 26 that, for accuracy of recommended product evaluations, a 30’s 27 13 22 23 35 11 certain number of people with given user attributes are 40’s 18 46 21 22 23 40 needed. dry skin category demonstrates only a slight difference in V. CONCLUSION product satisfaction level between recommended products and nonrecommended products. It is expected that many Herein, it was postulated that basic skin care and user users with different skin types registered under the mixed compatibility are based on the ingredients in cosmetics, skin category. As a result, it was expected that there would an analysis of user evaluations and cosmetic ingredients not be a great divide in the considered appropriateness of was conducted using the @cosme cosmetic review site, the products. and ingredient names expected to have strong cosmetic effect were extracted. In addition, a lotion recommendation A similar experiment was conducted for cosmetic effects system was implemented based on lotions that contained other than the “moisture” effect, i.e.,“whitening,” “pores,” featured ingredients from the list of extracted ingredients. and “acne.” The results are shown in Tables VI, VII, and Users were classified according to age and skin type into 24 VIII, respectively. classes of user attribute combinations, and recommended products were derived using four cosmetic effects, i.e., As a result, our system was able to recommend products “moisture,” “whitening,” “pores,” and “acne,” for each set with high satisfaction level for effects other than “moisture.” of user attributes. As can be seen, the overall satisfaction level for each are low compared to the results for the “moisture” effect. Thus, Based on the levels of recommended product satisfaction, since lotion is a product purchased for its “moisture” effect, it was found that the accuracy of the evaluations was the number of users evaluating it highly for the “pores” generally good. Relative to the products recommended by and “acne” cosmetic effect tags is low. In fact, “pores” and this system, there were many products not included in the “acne” skin type users are expected to rate products other strong-effect product group. These products include many than lotion highly for other purposes (cream and cleansing, that are unpopular, have few users, or lack reviews due to among others). recently going on sale; thus, it is considered that this system can recommend products with a high degree of serendipity In addition, depending on the user attributes, there are and hidden attraction, among others. cases where the nonrecommended product satisfaction level exceeds that of the recommended product satisfaction level. In our experiments, basic skin care and user compatibility This appears to be caused by a small number of people with were assumed to be based on cosmetic ingredients. As a a set of user attributes and a small number of reviewed result, it is possible that compatibility may be improved products. For example, [teens, atopic skin] and [40s, oily through the absence of an ingredient; thus, in the future, skin] represent only 5% of product user reviews from it will be necessary to verify the two together. In addition, the [30s, mixed skin] group. For these user attributes, there is concern regarding the reliability of the extracted 26
Search
Read the Text Version
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
- 15
- 16
- 17
- 18
- 19
- 20
- 21
- 22
- 23
- 24
- 25
- 26
- 27
- 28
- 29
- 30
- 31
- 32
- 33
- 34
- 35
- 36
- 37
- 38
- 39
- 40
- 41
- 42
- 43
- 44
- 45
- 46
- 47
- 48
- 49
- 50
- 51
- 52
- 53
- 54
- 55
- 56
- 57
- 58
- 59
- 60
- 61
- 62
- 63
- 64
- 65
- 66
- 67
- 68
- 69
- 70
- 71
- 72
- 73
- 74
- 75
- 76
- 77
- 78
- 79
- 80
- 81
- 82
- 83
- 84
- 85
- 86
- 87
- 88
- 89
- 90
- 91
- 92
- 93
- 94
- 95
- 96
- 97
- 98
- 99
- 100
- 101
- 102
- 103
- 104
- 105
- 106
- 107
- 108
- 109
- 110
- 111
- 112
- 113
- 114
- 115
- 116
- 117
- 118
- 119
- 120
- 121
- 122
- 123
- 124
- 125
- 126
- 127
- 128
- 129
- 130
- 131
- 132
- 133
- 134
- 135
- 136
- 137
- 138
- 139
- 140
- 141
- 142
- 143
- 144
- 145
- 146
- 147
- 148
- 149
- 150
- 151
- 152
- 153
- 154
- 155
- 156
- 157
- 158
- 159
- 160
- 161
- 162
- 163
- 164
- 165
- 166
- 167
- 168
- 169
- 170
- 171
- 172
- 173
- 174
- 175
- 176
- 177
- 178
- 179
- 180
- 181
- 182
- 183
- 184
- 185
- 186
- 187
- 188
- 189
- 190
- 191
- 192
- 193
- 194
- 195
- 196
- 197
- 198
- 199
- 200
- 201
- 202
- 203
- 204
- 205
- 206
- 207
- 208
- 209
- 210
- 211
- 212
- 213
- 214
- 215
- 216
- 217
- 218
- 219
- 220
- 221
- 222
- 223
- 224
- 225
- 226
- 227
- 228
- 229
- 230
- 231
- 232
- 233
- 234
- 235
- 236
- 237
- 238
- 239
- 240
- 241
- 242
- 243
- 244
- 245
- 246
- 247
- 248
- 249
- 250
- 251
- 252
- 253
- 254
- 255
- 256
- 257
- 258
- 259
- 260
- 261
- 262
- 263
- 264
- 265
- 266
- 267
- 268
- 269
- 270
- 271
- 272
- 273
- 274
- 275
- 276
- 277
- 278
- 279
- 280
- 281
- 282
- 283
- 284
- 285
- 286
- 287
- 288
- 289
- 290
- 291
- 292
- 293
- 294
- 295
- 296
- 297
- 298
- 299
- 300
- 301
- 302
- 303
- 304
- 305
- 306
- 307
- 308
- 309
- 310
- 311
- 312
- 313
- 314
- 315
- 316
- 317
- 318
- 319
- 320
- 321
- 322
- 323
- 324
- 325
- 326
- 327
- 328
- 329
- 330
- 331
- 332
- 333
- 334
- 335
- 336
- 337
- 338
