Implementation of Residual Network (ResNet) for Devanagari Handwritten Character Recognition

Abstract

Deeper convolutional neural network architecture has successfully given good results in many applications, but the problem with them is sometimes increasing the number of layers that result in saturation or degradation of the accuracy. This is because of the problem called vanishing gradient. Residual network which was a major breakthrough in the field of neural network helped to overcome this problem and allow to have a deeper network. The current state-of-the-art architecture for handwritten Devanagari character recognition is the convolutional neural network, and attempts were made to make the network deeper; but it also resulted in lower accuracy. Here, we have made an another attempt to use deeper network for this problem but this time using residual network. We have performed the experiments using ResNet 34 and ResNet 50 architecture and compared their result with the current state-of-the-art architecture of convolutional neural networks of two different depths: four layers and eight layers. ‘Devanagari handwritten character dataset’ (DHCD) has been used for obtaining and evaluating the results of the above-mentioned architectures. The experiment proves that the proposed ResNet implementation provides significantly higher accuracy than the current state-of-the-art architecture for the Devanagari handwritten character recognition problem.