Handwritten Vedic Sanskrit Text Recognition Using Deep Learning and Convolutional Neural Networks

Abstract

Recognizing Vedic Sanskrit text is essential for accessing classical Indo-Aryan language, predominantly utilized in the Vedas. Currently, there is limited awareness about the Vedas, making this field a highly demanding and challenging area in pattern recognition. To accelerate progress in optical character recognition (OCR), deep learning methods are indispensable. This article presents a novel approach to Vedic Sanskrit text recognition, incorporating deep convolutional architectures with their respective interpretations. We introduce three modified 4-fold CNN architectures and the AlexNet model. Our system comprises a handwritten dataset containing 140 distinct Vedic Sanskrit words, with approximately 500 images per word, totaling around 70,000 images. The dataset is partitioned for training and testing in an 80:20 ratio. Training is conducted using 20% of the samples, and the resulting model is applied to the deep convolutional network with varied sets of neurons in their hidden layers. Our proposed method demonstrates robust support for accurate Vedic Sanskrit word classification. The recognition rate achieved in our research is 97.42%, with an average recognition time of 0.3640 milliseconds, surpassing existing CNN-based approaches.