An efficient method for improving classification accuracy of handwritten Bangla compound characters using DCNN with dropout and ELU

Abstract

Handwritten character recognition is an essential part of optical character recognition domain. Bangla handwritten compound character recognition is a complex task that is challenging due to extensive size of and sheer diversity within the alphabet. The current work proposes a novel method of recognition of compound characters in Bangla language using deep convolutional neural networks (DCNN) and efficient greedy layer-wise training approach. Introduction of dropout technology mitigates data overfitting and Exponential Linear Unit (ELU) is introduced to tackle the vanishing gradient problem during training. ELU is a special rectified linear unit which provides sustainability against the vanishing as well as exploding gradients. Furthermore, dropout influences network elements to learn diverse representation of data, which contributes to generalization of model. The model is tested on CMATERdb 3.1.3.3 data set of compound characters, and the performance is found to outperform existing state-of-the-art methods of Bangla handwritten complex character recognition.