Structure Extraction from Decorated Characters by Graph Spectral Decomposition and Component Selection Criterion

For recognition of decorated characters through existing optical character recognition system (OCR), structure extraction of the decorated characters is required. Omachi et al. proposed a structure extraction method for these characters. However, characters including both thin lines and thick lines simultaneously tend to fail the extraction of structure. To overcome the performance in Omachi's method, we have proposed another method based on graph spectral decomposition. In the previous work, we showed applicability of the proposed method to wider range of decorated characters. However, selection criterion among decomposed components for obtaining good results have not been clarified. In this paper, we propose a new criterion to select components in the graph spectral decomposition framework for structure extraction. Experimental results show the validity of the criterion.