Geometric Model and Projection Based Algorithms for Tilt Correction and Extraction of Acsenders / Descenders for Cursive Word Recognition

Abstract

Cursive word recognition requires tilt correction before extraction of features such as ascenders and descenders. Skew and slant are the two types of tilts found in cursive word images. This paper presents new algorithms for skew and slant correction using geometric model and image projections. A new algorithm for extraction of ascenders/descenders based on fitting top line and base line references using horizontal histogram projection of the image is also presented. The algorithms are tested on a large sample of 'area' name images drawn from postal addresses. The tilt correction algorithms have been tested on a large sample of images and the results are encouraging. The ascenders/descenders are correctly extracted from 87.86% of test images