Improved Symbol Segmentation for TELUGU Optical Character Recognition

Abstract

In this paper, we propose two approaches to improving symbol or glyph segmentation in a Telugu OCR system. One of the critical aspects having an impact on the overall performance of a Telugu OCR system is the ability to segment or divide a scanned document image into recognizable units. In Telugu, these units are usually connected components and are called glyphs. When a document is degraded, most connected component based algorithms for segmentation fail. They give malformed glyphs that (a) are partial and are a result of breaks in the character due to uneven distribution of ink on the page or noise; and (b) are a combination of two or more glyphs because of smudging in print or noise. The former are labelled broken and the latter, merged characters. Two new techniques are proposed to handle such characters. The first idea is based on conventional machine learning approach where a Two Class SVM is used in segmenting word into valid glyps in two stages. The second idea is based on the spatial arrangement of the detected connected components. It is based on the intuition that valid characters exhibit certain clear patterns in their spatial arrangement of the bounding boxes. If rules are defined to capture such arrangements, we can design an algorithm to improve symbol segmentation. Testing is done on the Telugu corpus of about 5000 pages from nearly 30 books. Some of these books are of poor quality and provide very good test cases to our proposed approaches. The results show significant improvements over developed Telugu OCR (Drishti System) on poor-quality books that contain many ill-formed glyphs.