An attempt on the extraction of strokes from Hiragana characters based on semantic approach and hierarchical structure

We propose a new approach to segmenting and extracting strokes from handwritten character images based on the hierarchical structure. Previously, we proposed the integration of local processing by using the hierarchical structure, and demonstrated the extraction of lines from an image as its application. We now expand the proposed technique of line extraction and extract contours of the recognition object in a piecewise manner. The merging of lines is conducted in three steps. The first step is ordinal, i.e., based on points proximity. The second and the third steps are based, respectively, on the rule of good continuity and the semantic approach. The data are "Hiragana", cursive syllabary used in Japan. The ETL8 database is used. As a result, we confirmed the stroke extraction of the character samples which is known to be quite difficult. Experimental results are shown and analyzed, demonstrating the high performance of our method.