Handwritten word recognition with an OCR-based segmenter

This paper presents lexicon-free handwritten word recognition with an OCR-based segmenter. The input image is segmented into small pieces. A fast character classifier is then used to regroup the segments into potentially valid characters. Character hypotheses are then sent through a highly accurate tri-corroboration-based character classifier. The contiguous sets of characters will then result in a ranked list of possible words. Results on 12,968 numeric fields extracted from the Bureau of Census 95 database yielded a correct field rate of 92.7% with a 1.4% reject rate. However on 6,785 alpha fields the correct field rate performance was 51.7% with a 20.1% reject rate. The largest cause of error is the selection of the wrong character length path hypothesis. In fact when selecting the proper length hypothesis on numeric fields a correct rate of 99.1% with a 0.1% reject rate is achieved. Future improvements currently underway are discussed