Handwritten Bangla Character Recognition Using Deep Convolutional Neural Network: Comprehensive Analysis on Three Complete Datasets

Bangla handwritten character recognition is a difficult job compared to other languages due to the morphological complexity of adjacent characters and a wide variety of curvatures in writing styles people have. Another reason for that is the unique presence of compound characters. Most of the recent research works conducted in this field standardize Deep Convolutional Neural Network (DCNN) models for delivering the most effective outcomes. This paper proposes a DCNN model to classify all the character classes from three popular databases known as BanglaLekha Isolated, Ekush, and CMATERdb. As for BanglaLekha Isolated, our model achieves 93.446% accuracy on the 50 alphabets category and an overall 91.45% considering the whole dataset. The other two datasets, Ekush and CMATERdb result in 95.05% and 94.17% respectively, where the second one holds 171 classes of compound characters alone and performs 93.259% correctness, which is so far the best for this specific category in this dataset.