HANDWRITTEN CHARACTER RECOGNITION OF DEVANAGARI SCRIPT

Abstract

Optical Character Recognition [1] development has been gaining popularity in recent years, for Devanagari script along with other Indic Scripts. The script serves as a base for over 100 languages around the world including few popularly used scripts: - Hindi, Marathi, Sanskrit etc. Development of robust OCR systems for Devanagari script will allow us to preserve old manuscripts by converting the physical files to digital formats. This will make the process of storage, retrieval and transfer very convenient. This project proposes the use of Convolutional Neural Networks as feature extractor for extraction of features from handwritten Devanagari characters. For classification, classifiers employed are SVM (Linear, Polynomial and RBF), KNN, RF, DT and MLP. Use of CNN model for feature extraction eliminates the need of handcrafted features by traditional pattern recognition methods. The experiments with these seven techniques have been done on the DHCD dataset proposed in year 2015. Use of CNN proved to be very effective for Devanagari characters recognition as all the models achieved recognition accuracy of over 93% and total training time including feature extraction and classification did not exceed a total of 12.16 minutes