Evaluation of Neural Based Feature Extraction Methods for Printed Telugu OCR System

摘要

The Telugu is one of the oldest and most popular languages of India, especially in South India. The reported works on development of optical character recognition (OCR) systems for Telugu script is little. Moreover, Telugu is a complex script in which the characters are made up of one or more connected components resulting in a huge number of possible combinations, running into hundreds of thousands. In any OCR system, feature extraction is one of the most important phases. There are several methods that are suitable for different language scripts. These methods are broadly classified into template base, structural, statistical, neural network based and SVM based. In this paper we describe various feature extraction methods and evaluate by applying to Telugu script. In this process we have identified diagonal based, geometrical based and distance metric based feature extraction methods and also proposed a Pixel based feature extraction method. All these methods are implemented and evaluated with 364 Telugu characters using multilayer neural network as a classifier. The recognition accuracies of geometrical, diagonal, pixelmap and distance metric based feature extraction methods are 98.6%, 100%, 98.31% and 99.32% respectively. From the experiment it is understood that diagonal based method most suitable for Telugu script than other feature extraction methods.