Substructure Shape Analysis for Kanji Character Recognition

A method towards analytical recognition of Chinese characters is described. Basic character components (substructures) of any size are recognized anywhere on a Kanji string, even if they touch other components. The algorithm performs skeleton extraction, skeleton grouping, indexing of structural features on a previously generated look-up table, structural verification of hypothesis using model graphs, and geometrical verification by an array of neural nets, each one specialized on the geometry of each model. The system retrieves 98 % of substructures with 91% precision rate.