A two phase trained Convolutional Neural Network for Handwritten Bangla Compound Character Recognition

Abstract

Recognizing Bangla compound characters is a challenging problem due to its high curly nature. In this paper, we propose a convolutional neural network (CNN) architecture to recognize handwritten Bangla compound characters. The learning of proposed architecture is done in two phase. In the first phase, a CNN is trained in an unsupervised way to minimize the reconstruction loss. Afterward, these weights are used to initialize the starting layers of second CNN to reduce the recognition loss through supervised learning. The effectiveness of the proposed model is validated on compound character dataset CMATERdb 3.1.3.3, which consists of 171 different character classes. It achieves recognition results of 93.90% and 97.37 % in top 1 and top 2 choices. The recognition performance outperforms state-of-the-art method for handwritten Bangla compound characters by a margin of 3.57%.