Word Extraction Method by Generating Multiple Character Hypotheses

摘要

It is necessary to extract precisely words of headers and data for recognizing logical structure of form images. However, word extraction often fails because of layout analysis or character recognition error, which leads correct character hypotheses not to be generated. We propose a word extraction method which generates multiple character hypotheses and extracts their combinations which correspond with the character orders of words. Firstly character hypotheses which overlap with each other are generated by combinatorial recognition of connected components and their combinations which correspond with words are extracted by clique extraction from a graph. And then, character hypotheses are generated by recognition with limited target and their combinations which correspond with words areextracted by matching between lattices based on local optimum, in which variety of recognition results and regular expression of words are considered. We confirmed the effect of our method by the experiment for form images.