Handwritten Devanagari Digits Recognition Using Residual Neural Network

Abstract

Handwritten digit recognition is a highly evolved research domain of pattern recognition. It is used to classify pre-segmented handwritten digits. The Devanagari script is one of the writing systems of various Indian languages including Sanskrit and Hindi. In this paper, an efficient Handwritten Devanagari numeral digit recognition using ResNet is proposed. Deep learning is a recent research trend in this field. Architectures like Residual neural Networks (ResNet) are being used. ResNet is an architecture that is computationally expensive and normally used to provide high accuracy in classification problems. The structural design of the network consists of sacks of two convolutional (Conv2D) layers with Batch Normalization and an activation function called Relu. We evaluated our scheme on 16000 handwritten samples of Devanagari numerals from the UCI machine learning database and from the experiment we have achieved 99.40% recognition rate.