A correspondence based approach to segmentation of cursive words

Abstract

A novel contour-based approach to off-line segmentation of cursive words is proposed. The approach avoids expensive processing such as slant correction, thinning, filtering, and regional pixel labeling. Correspondences between the maximum and minimum excursions of the handwriting image are established and categorised. Valid established correspondences additionally provide three types of useful singular features about the segmented letter shapes. These singular features of the handwriting are first defined and isolated as: Loops, ascenders and descenders. A consistent analysis of established valid correspondences gives the segmentation. Correspondences also provide stroke width estimates, character width, and slant estimates that are useful to a recognition system. Results of the proposed approach on images, with variations in slant and types of ligatures are also presented.