CNN-Based Bangla Handwritten Character Recognition: Exploring Ekush Dataset for Performance Enhancement | DIU | Thesis Presentation
Download as PPTX, PDF0 likes63 views
"This thesis presentation explores the use of Convolutional Neural Networks (CNNs) for Bangla handwritten character recognition. It focuses on the Ekush dataset and discusses techniques for enhancing performance, challenges faced, and the results achieved in improving recognition accuracy." Paper Link: https://internship.daffodilvarsity.edu.bd/?app=home&cmd=reportsViewPDFskin&project_number=26095 Video: https://www.youtube.com/watch?v=mK0WJVAMMqs
CNN-Based Bangla Handwritten Character Recognition: Exploring Ekush Dataset for Performance Enhancement | DIU | Thesis Presentation
- 1. Final Year Defense Presented by Supervised by CNN-Based Bangla Handwritten Character Recognition: Exploring Ekush Dataset for Performance Enhancement Marcel David Baroi ID: 201-15-3421 Department of CSE Daffodil International University Tasnim Tabassum Lecturer Department of CSE Daffodil International University
- 2. Final Year Defense Overview Exploring the Ekush dataset to recognize Bengali handwritten characters Three CNN models, along with 13 ML models, are explored for comparative analysis. 2
- 3. Final Year Defense 3 Motivation Contribute to the BHR's (Bangla Handwritten Recognition System) improvement, as it is an enormous task that is almost impossible for one person to complete alone. In an attempt to make a tiny contribution, I have attempted to apply various models to an already-existing dataset, since I found it extremely challenging to gather sufficient data for a very successful final study of my own.
- 4. Final Year Defense 4 Objective To explore the Ekush dataset, a dataset of Bangla handwritten characters, and analyze its suitability for training and evaluating CNN models for character recognition. To implement CNN architectures specifically suited for recognizing Bangla handwritten characters, taking into account the unique characteristics of the Bangla script, such as complex ligatures and conjuncts. To achieve high accuracy and performance in recognizing Bangla handwritten characters using the developed CNN model. To compare the performance of the proposed CNN model with existing approaches for Bangla handwritten character recognition. To contribute to the development of robust and efficient Bangla handwritten character recognition systems for various applications, such as document image analysis, online education, and assistive technologies.
- 5. Final Year Defense 5 Related works (1) Year Dataset Recognition Used Models Height Accuracy and Model (2017) CMATERdb Numeral CNN 98.78% CNN (2018) NumtaDB Numeral LeNet-5, ResNet-18, Proposed Method, Ensemble (K-Fold) 96.79% Proposed Method (2022) Ekush Digit KNeighborsClassifier, SVC, NuSVC, AdaBoostClassifier, GradientBoostingClassifier, CNN 96.707% CNN
- 6. Final Year Defense 6 Related works (2) Year Dataset Recognition Used Models Height Accuracy and Model (2005) database of 6000 samples Digit Dempster-Shafer (DS) technique 95.1% (2018) CMATERdb digits, alphabets, and special character CNN, VGG16,NiN, ResNet, FractalNet, DenseNet 99.13% DenseNet (2018) Austrian Research Institute for Artificial Intelligence, Austria Numeral MPL, SVM, J48, Random Forest, Naive Bayes, Bayes Ne, Random Tree, 90.37% MLP
- 7. Final Year Defense 7 Related works (3) Year Dataset Recognition Used Models Height Accuracy and Model (2017) collected more than 16,000 Bengali sentiment classification word2vec model 75.5% word2vec (2012) Numeral GA, SA and HC 97% GA (2009) ISI Bangla numerals, CENPARMI Farsi numerals, and IFHCDB Farsi numerals Numeral MQDF, PNC, CFPC, SVM, DLQDF 99.73% DLQDF
- 8. Final Year Defense 8 Contribution Analyzed the character section of the Ekush dataset Experimented with more than 16 different models Concluded the best possible algorithm for the dataset
- 10. Final Year Defense 10 Dataset Description Type Number of classes Amount of train data for each class Amount of test data for each class Vowel 11 1000 600 Total Vowel 11 11000 6600 Consonant 39 1000 600 Total Consonant 39 39000 23400
- 11. Final Year Defense 11 Custom CNN
- 14. Final Year Defense 14 Experimental Result Random Forest KNN SVM Logistic Regression AdaBoost Naive Bayes Extra Trees Bagging Classifier Decision Tree Stochastic Gradient Descent Nearest Centroid Linear SVM M ultinom ial Naive Bayes Custom CNN DenseNet LeNet 0 20 40 60 80 100 120 Model Accuracy Vowel Consonant
- 15. Final Year Defense 15 Experimental Result Model Accuracy (%) Vowel Consonant Random Forest 83.64 71.60 KNN 82.00 69.41 SVM 88.14 78.77 Logistic Regression 75.86 59.42 AdaBoost 54.23 28.48 Naive Bayes 66.73 43.53 Extra Trees 85.05 72.15 Bagging Classifier 75.18 56.29 Decision Tree 60.59 41.99 Stochastic Gradient Descent 68.64 42.85 Nearest Centroid 71.11 50.00 Linear SVM 87.75 40.60 Multinomial Naive Bayes 71.00 48.12 CustomCNN 92.96 87.72 DenseNet 97.93 95.99 LeNet 91.63 85.35
- 16. Final Year Defense 16 Conclusion During my exploration of the Ekush dataset, I trained more than sixteen models, all of which were expertly constructed and modified by me. Of them, DenseNet was the clear winner, achieving remarkable accuracy in both vowel and consonant recognition (97.93% for vowels and 95.99% for consonants). It truly demonstrated the mastery of neural networks! However, SVM—the brave seasoned pro in machine learning—remained the top performer in the traditional model class, with accuracy of 88.14% and 78.77%, respectively. But a strange pattern became apparent: all other models, in spite of their best attempts, appeared to prefer the dataset's vowel section, finding it more difficult to handle the consonant world's subtle complexities. This intriguing observation invites more research and opens up possibilities.
- 17. Final Year Defense 17 Future Scope Data Augmentation and Exploration Architecture and Algorithm Exploration Algorithm-Specific Explorations Addressing Challenges and Real-World Applications
- 18. Final Year Defense 18 Reference Basu, Subhadip, et al. "Handwritten Bangla digit recognition using classifier combination through DS technique." Pattern Recognition and Machine Intelligence: First International Conference, PReMI 2005, Kolkata, India, December 20-22, 2005. Proceedings 1. Springer Berlin Heidelberg, 2005. Alom, Md Zahangir, et al. "Handwritten bangla character recognition using the state-of-the-art deep convolutional neural networks." Computational intelligence and neuroscience 2018 (2018). DeFries, R. S., and Jonathan Cheung-Wai Chan. "Multiple criteria for evaluating machine learning algorithms for land cover classification from satellite data." Remote Sensing of Environment 74.3 (2000): 503-515. Kamavisdar, Pooja, Sonam Saluja, and Sonu Agrawal. "A survey on image classification approaches and techniques." International Journal of Advanced Research in Computer and Communication Engineering 2.1 (2013): 1005-1009. Sharma, Arun, and Vidushi Sharma. "An Empirical Study of Supervised Learning Techniques on Multispectral Dataset." Shamim, S. M., et al. "Handwritten digit recognition using machine learning algorithms." Global Journal Of Computer Science And Technology 18.1 (2018): 17-23. Al-Amin, Md, Md Saiful Islam, and Shapan Das Uzzal. "Sentiment analysis of Bengali comments with Word2Vec and sentiment information of words." 2017 international conference on electrical, computer and communication engineering (ECCE). IEEE, 2017.
- 19. Final Year Defense 19 Reference Das, Nibaran, et al. "A genetic algorithm based region sampling for selection of local features in handwritten digit recognition application." Applied Soft Computing 12.5 (2012): 1592-1606. Surinta, Olarik, et al. "Recognition of handwritten characters using local gradient feature descriptors." Engineering Applications of Artificial Intelligence 45 (2015): 405-414. Das, Nibaran, et al. "Recognition of handwritten Bangla basic characters and digits using convex hull based feature set." arXiv preprint arXiv:1410.0478 (2014). Liu, Cheng-Lin, and Ching Y. Suen. "A new benchmark on the recognition of handwritten Bangla and Farsi numeral characters." Pattern Recognition 42.12 (2009): 3287-3295. Alom, Md Zahangir, et al. "Handwritten bangla digit recognition using deep learning." arXiv preprint arXiv:1705.02680 (2017). Noor, Rouhan, Kazi Mejbaul Islam, and Md Jakaria Rahimi. "Handwritten Bangla numeral recognition using ensembling of convolutional neural network." 2018 21st international conference of computer and information technology (ICCIT). IEEE, 2018. Shawon, Md Tanvir Rouf, Raihan Tanvir, and Md Golam Rabiul Alam. "Bengali Handwritten Digit Recognition using CNN with Explainable AI." 2022 4th International Conference on Sustainable Technologies for Industry 4.0 (STI). IEEE, 2022.
- 20. Final Year Defense 20 THANK YOU