Hypothesis Preservation Approach to Scene Text Recognition with Weighted Finite-State Transducer

This paper shows that the use of Weighted Finite-State Transducer (WFST) significantly eliminates large-scale ambiguity in scene text recognition, especially for Japanese Kanji characters. The proposed method consists of two WFSTs called WFST-OCR and WFST-Lexicon. WFST-OCR handles the multiple hypotheses caused by erroneous text location, character segmentation and character recognition processes. The following WFST-Lexicon and its convolution of WFST-OCR resolve the hypotheses. The WFSTs integrate the conventional OCR and post-processing processes into one process. The benefit from the proposed method is that all the ambiguities are held as WFST data, and solved in one integrated step, the system outputs texts that are statistically consistent with regard to segmentation possibilities and the given language model. An experimental system demonstrates practical performance in spite of the hypothesis complexity inherent in the ICDAR test set and Kanji character texts.