Devanagari Handwritten Character Recognition

This paper aims to classify images of handwritten characters in the Devanagari script. We implement a ResNet architecture with 85 convolution layers to classify images on the publicly available Devanagari Handwritten Character Dataset (DHCD), that holds 92,000 images divided into 46 different classes. Our network implements the bottleneck variant of the residual module and executes the pre-activation method where the activation function and batch normalization are placed before the convolutions. This model outperforms previous works done to date on DHCD by recording an accuracy of 99.72%.