Discrete wavelet transform and convolutional neural network based handwritten Sanskrit character recognition

Abstract

Sanskrit is one of the ancient languages from which the majority of present Indian languages are developed. Although the national mission for manuscripts (NMM) is digitizing handwritten Sanskrit manuscripts, there are still a lot of papers that need to be digitized. Recognition of handwritten script is a challenging task due to individual differences in writing styles and how those variations alter over time. The Sanskrit language is written in Devanagari script. A novel approach using discrete wavelet transform (DWT) and convolutional natural network (CNN) is proposed in this paper. Devanagari handwritten character dataset which includes 2000 handwritten images of 36 classes (2000*36=72000) is used in this research. Fine-tuned GoogLeNet model implemented here gave optimum values of epochs and learning rate of 15 and 0.01 respectively. Classification accuracy obtained by proposed DWT – CNN model is 98.97% with a loss of 0.098. Fine-tuned GoogLeNet model achieves 99.68% accuracy with a 0.0635 loss. Results obtained are also compared with existing approaches and found superior.