An investigation into Telugu font and character recognition

Abstract

Optical Character Recognition (OCR) for Indian scripts and for handwriting is an active area of research. OCR systems for Indian and many other oriental languages are not yet able to successfully recognize printed documents of varying scripts, quality, font styls and sizes. To address some of these issues, we have designed and developed a framework to simultaneously recognize fonts and characters from documents printed in Telugu. Telugu script is structurally complicated due to the large number of character shapes that can be formed. Specifically, we developed a system based on Rough Sets theory for recognizing Telugu fonts and characters. Using the theory of rough sets, we determine the most important features of characters from the point of view of classification. The present work achieved 94.37%, and 90.17% accuracies for character and font recognitions respectively. These are the raw recognition accuracies without doing any post processing. The present model emphasizes the role of theory of rough sets in feature selection and dimension reduction in Telugu font and character recognition. A comparative study with related earlier works and our contributions to the field of Telugu character recognition are presented. The developed approach falls into the category of a priori classification, where domain knowledge of the content is not used. No explicit local feature analysis is used in this methodology. Applicability of the proposed approach is tested on documents printed in English, which is an entirely different script from Telugu. Thus, a general framework for script-independent character and font recognitions is suggestive from the present methodology. Encouraging recognition accuracies in both the scripts (Telugu and English) are achieved with a scope for further improvement.