An off-line approach to extract natural strokes from handwritten Chinese characters

This paper presents a new attempt to extract natural strokes from handwritten Chinese character. The extraction of this feature set has hardly been attempted sufficiently. One approach for this has been proposed in which duplicated employment of building blocks in multiple strokes is allowed. Unlike that approach, our technique would prohibit the employment of the same building block in forming distinct output strokes, so as to facilitate the analysis of the problems involved in the extraction of these strokes. Stroke model, which describes the structural characteristics of a stroke class, is constructed for every class. Special treatments, with substantial reference to the structural knowledge of the target stroke set, are employed to handle the deformations introduced during the thinning process. By the experiments, it is verified that decision regions of the stroke classes, based on several parameters, exhibit overlapping in both the single-writer and multiple-writer cases. The idea of “personalization” is introduced to partially improve this situation. Finally, by restricting the building blocks to be employed by one output stroke only, the extractor is able to produce a measure called “valid stroke count” for each input character which is of 93% accuracy out of the total expected stroke count. This measure can serve as an important guide in narrowing the search space in the character matching phase