Unveiling the Comparative Insight of Bangla Handwritten Digit Recognition using Machine Learning and Deep Learning Models

Abstract

In the realm of Artificial Intelligence, despite significant advancement in the computer vision and pattern recognition field, recognition of Bangla handwritten digit recognition has still been undergoing some unique challenges, including the distortion in the images, the variation in writing styles, the different shapes and sizes of digits, and the varying levels of noise. There is a further room for more improvement. In this paper, we not only have delved into the comparative analysis of several machine learning and deep learning models to portray the macro and micro level insights, but also have achieved more accuracy in some models. If we can merge this machine learning and deep learning with OCR, which is largely getting attention in these days, it could be a huge contribution to this field. We have used the Ekush Bangla handwritten numeral digit dataset of size 30687. Four machine learning algorithms, Artificial Neural Network and three architectures of CNN namely ResNet, Xception, LeNet have been used to depict the analysis. We have explored the pattern matching approach and achieved 91% accuracy for the Support Vector Machine, 80.44% accuracy for Decision Tree, 91.22% for Random Forest & 90.38% for KNN. And in the Deep Learning approach, through the Convolutional Neural Network algorithm’s architectures we have achieved 99.48%, 98.63% and 97.23% respectively in the Xception, ResNet, and LeNet. Finally in ANN we have got accuracy of 90.50%