Thai OCR error correction using token passing algorithm

Spell checking can be used to improve OCR result, which is quite time consuming for Thai language. Since there is no explicit word boundary, the spell checking has to go through all possible ambiguity characters and word boundaries. This paper proposes a token passing algorithm, which is often used in speech recognition. To this problem, the output of OCR consists of strings of at most 5 most probable characters. The tokens are generated for each letter and are passed to the next 5 characters. Each time the token is passed, the dictionary is used to check for the correct spelling. The wrong spelling tokens are discarded. Word boundary is inserted to the token to correct boundary ambiguity. At the end of the passing processes, all possible correct spelling words are found as long as the correct characters are in the list of 5 characters. Many possible word sequences could be constructed from this process. One best word sequence could be selected using bi-gram, trigram and/or lowest token probability. This method can also handle some insertion and deletion error.