@cscjournals

Design and Development of a 2D-Convolution CNN model for Recognition of Handwritten Gurumukhi and Devanagari Characters

, and . International Journal of Image Processing (IJIP), 12 (3): 72-82 (October 2018)

Abstract

Owing to the innumerable appearances due to different writers, their writing styles, technical environment differences and noise, the handwritten character recognition has always been one of the most challenging task in pattern recognition. The emergence of deep learning has provided a new direction to break the limits of decades old traditional methods. There exist many scripts in the world which are being used by millions of people. Handwritten character recognition studies of several of these scripts are found in the literature. Different hand-crafted feature sets have been used in these recognition studies. Feature based approaches derive important properties from the test patterns and employ them in a more sophisticated classification model. Feature extraction using Zernike moment and Polar harmonic transformation techniques was also performed and a moderate classification accuracy was also achieved. The problems faced while using these techniques led us to use CNN based recognition approach which is capable of learning the feature vector from the training character image samples in an unsupervised manner in the sense that no hand-crafting is employed to determine the feature vector. This paper presents a deep learning paradigm using a Convolution Neural Network (CNN) which is implemented for handwritten Gurumukhi and devanagari character recognition (HGDCR). In the present experiment, the training of a 34-layer CNN for a 35 class self-generated handwritten Gurumukhi and 60 class (50 alphabet and 10 digits) handwritten Devanagari character dataset was performed on a GPU (Graphic Processing Unit) machine. The experiment resulted with an average recognition accuracy of more than 92% in case of Handwritten Gurumukhi Character dataset and 97.25% in case of Handwritten Devanagari Character dataset. It was also concluded that the training and classification through our network design performed about 10 times faster than on a moderately fast CPU. The advantage of this framework is proved by the experimental results.

Links and resources

Tags