Character image thinning is the key processing technique to convert a character image into a set of strokes. Conventional pixelwise thinning algorithms are based on a formal definition of skeleton, but ignore the fact that the image is a result of writing with a fixed width pen.
In this paper, a new thinning algorithm is proposed utilizing the width information based on stroke relation graph (SRG). SRG, where the arcs represent possible strokes and the nodes represent crossing points of strokes, is used to express stroke relations in stroke junction, or a non-regular region. Since each subgraph of SRG can be considered as a candidate for the thinning result in the region, thinning in a non-regular region can be transformed into the graph search problem of finding the "best" subgraph. Contour shape similarity, stroke width and smoothness of stroke connection are used to rank the subgraphs.
Experiments were conducted with 1078 Hangul characters written by 11 persons. The results showed that the proposed thinning algorithm yields better skeletons than conventional pixelwise thinning, particularly for the images of slightly touching strokes.