A CNN Architecture-Based Deep-Learning Model for Recognition of Handwritten Devanagari Character with High Accuracy Rate

Abstract

Handwritten character recognition is increasingly important in a variety of automation areas,
for example, authentication of bank signatures, identification of ZIP codes on letter
addresses, and forensic evidence, etc. Handwritten character recognition is the process where
a machine detects and recognizes characters from a text image and converts that processed
data into a code that is understood by the machine. This is a fundamental but challenging task
in the field of pattern recognition. In this paper, we have used a new public image dataset for
the Devanagari script character: the Devanagari character dataset (DCD). This considered
dataset contains 92 thousand images of 46 different classes of characters in the Devanagari
script, fragmented from handwritten documents. The paper also explores the challenges faced
in the identification of Devanagari characters. This paper proposes a deep learning based
convolutional neural network (CNN) architecture for the recognition of those handwritten
characters in an unrestricted environment, with datasets. Deep convolutional neural networks
have shown better results than traditional shallow networks in many recognition tasks. While
keeping distance from the routine approach of character recognition by Deep CNN, we focus
on the use of dropout and dataset augmentation approaches to improve test accuracy. By
applying these techniques to Deep CNN, we were able to increase the test accuracy to about
0.98 percent. The proposed architecture achieved the highest test accuracy of 98.13% on the
considered dataset. The results indicate that the proposed model may be a strong candidate
for handwritten character recognition and automatic handwritten Devanagari script character
recognition applications