Handwritten Vedic Sanskrit Text Recognition Using Deep Learning

Abstract

Vedic Sanskrit Text recognition is an origin to gain the information about classical language - Indo-Aryan language, basically used in Vedas. In current situation very less people are aware about Vedas; hence it is one of the most demanding and challenging research domains in pattern recognition. In order to rapid development in OCR, deep learning method is a vital technology. This article provides novel approach for Vedic Sanskrit Text recognition with its meaning using deep convolution architecture with its meanings. We have proposed three different 4-fold modified CNN architectures and Alexnet model. This system has a handwritten dataset which included 140 different Vedic Sanskrit words. Each word had approximately around 500 images each, the whole dataset had around 70000 images. Dataset is divided for training and testing with ratio 80:20. Dataset is trained using 20 percent samples and the same input is applied to the deep convolution network with several set of neurons in their hidden layers. Proposed method is highly supported for the correct Vedic Sanskrit word classification. The recognition rate obtained for our research was 97.42% in 0.3640 ms average recognition time, superior to which existing approaches with CNN.