Reading and Understanding Handwritten Addresses l

A methodology for understanding handwritten addresses is proposed. Address understanding is the process of using multiple information sources to assign a fiveor nine-digit ZIP Code to an address block consisting of several text lines. This method uses many diverse pattern recognition and image processing algorithms. To fully process the address image, we must perform thresholding, remove underlining, separate lines of text, segment text lines into words, determine the syntax of the image, and recognize digits, characters, and words. Our approach emphasizes using contextual information that is available in the address to determine a ZIP Code for a mail-piece. While other research efforts concentrate on simply locating arid recognizing the ZIP Code, our system allows the use of recognition information from the city name and state name to develop a better understanding of the address. This recognition information combined with USPS directory information assists in determining the ZIP Code. Tests of the present system on 508 address images result in a 75.2% accept rate with a 1.6% error rate. This paper describes the algorithms used and suggests improvements for future systems.