Bangla handwritten digit recognition using autoencoder and deep convolutional neural network

Abstract

Handwritten digit recognition is a typical image classification problem. Convolutional neural networks, also known as ConvNets, are powerful classification models for such tasks. As different languages have different styles and shapes of their numeral digits, accuracy rates of the models vary from each other and from language to language. However, unsupervised pre-training in such situation has shown improved accuracy for classification tasks, though no such work has been found for Bangla digit recognition. This paper presents the use of unsupervised pre-training using autoencoder with deep ConvNet in order to recognize handwritten Bangla digits, i.e., 0-9. The datasets that are used in this paper are CMATERDB 3.1.1 and a dataset published by the Indian Statistical Institute (ISI). This paper studies four different combinations of these two datasets-two experiments are done against their own training and testing images, other two experiments are done cross validating the datasets. In one of these four experiments, the proposed approach achieves 99.50% accuracy, which is so far the best for recognizing handwritten Bangla digits. The ConvNet model is trained with 19,313 images of ISI handwritten character dataset and tested with images of CMATERDB dataset.