Candidate search and elimination approach for Telugu OCR

Abstract

In this paper we propose an OCR system for Telugu based on the candidate search and elimination technique. The initial candidates for recognition are found by applying a zoning method on input glyphs. We propose cavities as a structural approach suited specifically for Telugu script, where cavity vectors are used to prune the candidates found by zoning. A final template matching stage using controlled non linear normalization is performed to conclude the search process. The search can be concluded when at any stage ever an unique candidate is found. A recognition accuracy of 97-98% was achieved on real images scanned from Telugu literature.