Bengali Handwritten Character Classification Using Transfer Learning on Deep Convolutional Network

Abstract

Bengali is the sixth most popular spoken language in the world. Computerized detection of handwritten Bengali (Bangla Lekha) character is very difficult due to the diversity and veracity of characters. In this paper, we have proposed a modified state-of-the-art deep learning to tackle the problem of Bengali handwritten character recognition. This method used the lesser number of iterations to train than other comparable methods. The transfer learning on Resnet-50 deep convolutional neural network model is used on pretrained ImageNet dataset. One cycle policy is modified with varying the input image sizes to ensure faster training. Proposed method executed on BanglaLekha-Isolated dataset for evaluation that consists of 84 classes (50 Basic, 10 Numerals and 24 Compound Characters). We have achieved 97.12% accuracy in just 47 epochs. Proposed method gives very good results in terms of epoch and accuracy compare to other recent methods by considering number of classes. Without ensembling, proposed solution achieves state-of-the-art result and shows the effectiveness of ResNet-50 for classification of Bangla HCR.