Devanagari Character Classification Using Heuristic Approach

OCR as a field of research that has large scope for exploration as the solutions drawn so far are still open for improvements. Its application has risen in variety of domains and the need in popular global languages apart from the prominent Latin script has much to investigate. Focus in this OCR work is on Devanagari script as it can influence multiple scripts like Sanskrit, Hindi, and Marathi. Also to uncover other OCR approaches apart from the traditional connected segments, support vector machines, recurrent neural networks, convolutional neural networks with LSTM, etc. Many offline Devanagari OCR techniques exist in recent years in principle but with lot of constraints. Hence there are challenges to extract textual information from scanned documents like large character set, linguistic-based complexities, critical conjuncts, typical geometric patterns of character, zone-based shapes, use of shirorekha to address. Aim is to use template-based matching principle to extract text localization features for recognition. Compared to SIFT approach which yields 77.3% accuracy on an average of three typefonts considered, the proposed approach is able to yield 94.7% on the base set of Devanagari script considered.