Segmentation of Handwritten Numeral Strings by Introduction of Structural Model

We have been developing a segmentation method for touching pairs of numerals. In this study, we extend the method to the segmentation of numeral strings that contain the touching patterns of multiple numerals. The mechanisms of the separation of overlapping strokes and the removal of ligature are introduced into our “segmentation graph.” We also attempt to combine local features and structural features to reduce segmentation errors. As the first step, we construct simple structural models for all categories and verify the recognition results using the technique of subgraph isomorphism. Then, we apply this technique to the detection of a touching pattern. We show that even the simple models can improve the detection accuracy. As a result, the recognition accuracy for numeral strings is also improved. This result indicates that segmentation errors that originate from noncharacters can be reduced by the introduction of structural models into our segmentation method.