A Hierarchical Handwritten Word Segmentation

Abstract

This paper presents a technique to hierarchically segment handwritten words. Handwritten words are broken down into either concave or convex segments called p-strokes (partial strokes). P-strokes are recombined to form Windows of Observation (WO), where the number of p-strokes defines the window size. Once a word is segmented into a sequence of p-strokes, window of observation of different sizes are scanned stroke-by-stroke across the word in a fully overlapped manner, producing a stream of potential characters.

Unsupervised clustering of window size 1 to 5 resulted in a codebook of 461 representative windows of observation. Secondly, supervised clustering using 25,000 handwritten alphacharacters resulted in a set of 2,525 prototypes. The segmentation and the clustering techniques are described in this paper.