Classification-based segmentation of ZIP codes

In this paper a system for the segmentation of unconstrained handwritten ZIP codes is presented. A binarization algorithm is described that utilizes the bimodal nature of the input 256 gray level histogram. ZIP code images are scaled to a standard size for further processing. Objects that are not comparable to the size of numerals are classified as noise and removed. An algorithm that segments broken or touching characters is presented. Several trial segmentations of each ZIP code are produced. A character classifier is used to determine which trial segmentation is most likely to be correct