Bangla Handwritten Character Recognition Using Convolutional Neural Network

Abstract

Recognizing handwritten characters is more challenging than recognizing printed characters. The size and shape of a handwritten character written by different people are not the same. The function of character recognition is complicated by the numerous variants in writing styles. Bangla handwritten character recognition had already been the focus of many researchers. Due to its specific characteristics of feature extraction and classification, the convolutional neural network (CNN) has recently shown notable progress in the fields of image-based recognition, video analytics, and natural language processing. As a result, this research provides a deep convolutional neural network (DCNN)-based mechanism for recognizing Bengali handwritten characters. In the field of pattern recognition, one of the most efficient ways to achieve higher accuracy or a lower error rate is to use a deep, optimized architecture that can process a large amount of data. Therefore, this paper has used DCNN for feature extraction and classification. This paper compares their accuracies by applying the DCNN model over three datasets: BanglaLekha-Isolated dataset, CMATERdb dataset, and our created dataset. By applying DCNN in researchers created dataset, accuracy has achieved up to 93.07%. This paper has also worked on simple CNN, DCNN, more advanced VGG-16 models for classification. And finally, the accuracies obtained from all the datasets have been compared.