A zip code recognition system using the localized arc pattern method

A system which is composed of two stages, the preprocessing stage and the recognition stage, is proposed. It is assumed that written images of a three-figure zip code sometimes protrude from the given frame without clear separation between figures. In the first stage, an image is separated into three figures by using a cutting operation and a unifying operation which recovers defects of written figures. In the second stage, every figure is judged as a numeral or otherwise rejected by using the localized arc pattern method. By an application of the proposed system to the IPTP database, 90.6% was recognized correctly, 0.7% was recognized erroneously, and 8.7% was rejected as unknown erroneously, and 8.7% was rejected as unknown. By choosing optional parameters properly, the system yielded the value 0.157 of the precision index defined by IPTP