Combining Multiple Feature Extraction and Classification Methods to Study Performance of Handwritten Sanskrit Character Recognition

Abstract

The recognition of Sanskrit handwriting has been found to be one of the most challenging research topics. The Sanskrit language is written using the Devanagari Script. In this paper, we implemented novel algorithm to recognize handwritten characters using four different feature extraction methods, resizing the image, Canny Edge detection, Hybrid Discrete Wavelet-Discrete Cosine Transform (DWT-DCT), Histogram of oriented Gradients (HOG). Classification is done by using support vector machine with cubic kenrel and neural network with ReLU activation function. When features are extracted by HOG, the SVM provides classification accuracy of 97.10% while the neural network provides 91.40%. SVM has been found to provide superior classification than neural networks for all feature extraction strategies.