Handwritten address interpretation using word recognition with and without lexicon

This paper describes an address interpretation system developed for determining the nine-digit ZIP code in mail pieces containing handwritten addresses. The system consists of 1) skew and slant correction, 2) line segmentation, 3) recognition of ZIP code field using context-free strategies, 4) recognition of street number field, 5) recognition of the PO Box field (if valid), and 6) evaluation of the recognition performance using all “bha” address images. In the evaluation test, the error rate of street line recognition was 1.45% with 40.14% accept rate and 1.62% with 40.71% accept rate when the National ZIP+4-database was used for street name lexicon generation or consistency check of ZIP Code and PO Box number