Word-level recognition of small sets of hand-written words

A system is presented for off-line word-level recognition of small sets of hand-written words. The system performs well on a variety of tasks such as recognizing numerals hand-written in English (one through ten), Hebrew characters, and Pascal's reserved words. Only small changes are needed in the parameters of the system when switching from one application to another. The heart of the proposed system is the use of dynamic programming for matching between a word in question and a prestored model of a word. The matching algorithm which uses line segments as its input is similar to Lee and Chen's algorithm for matching Chinese characters. We vary the segmentation algorithm showing that regardless of the segmentation algorithm chosen, the average recognition rate usually exceeds 95%. Different thinning algorithms are also shown not to reduce the stated recognition rate and a rejection scheme further enhances it.