Devanagari Handwritten Character Recognition using CNN as Feature Extractor

This paper 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, MLP and XGB. Use of CNN model for feature extraction eliminates the need of handcrafted features by traditional pattern recognition methods. XGB Classifier, which is a very recent technique, has also been explored (which has not been done previously to the best of our knowledge). The experiments with these eight 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 92% and total training time including feature extraction and classification did not exceed a total of 12.16 minutes.