Handwritten Bangla Numeral and Basic Character Recognition Using Deep Convolutional Neural Network

Abstract

In this paper, the problem of recognizing handwritten Bangla characters is addressed. Handwritten character recognition is one of the most practiced tasks in computer vision. Over the past few years Convolutional Neural Network has produced the best results in case of English handwritten character recognition. Although Bangla the official language of Bangladesh and several Indian states with over 200 million native speakers Bangla handwritten character recognition is quite far behind. We present a 9 layer sequential Convolutional Neural Network model to recognize 60 (10 numerals+ 50 basic characters) Bangla handwritten characters. BanglaLekha-Isolated dataset is used as train-validation set. A new dataset of 6000 images is created for cross validation. Our proposed model trained to recognize 60 characters achieves state-of-the-art 99.44% accuracy on BanglaLekha-Isolated dataset and 95.16% accuracy on prepared test set. Experiments on recognizing Bangla numerals separately also show state-of-the-art performance.