ZIP codes and street numbers are important information in handwritten addresses. They can be used to help identify a high level of sort of an address both directly and indirectly. For example, they can be used to obtain 9 digit ZIP level sort even if only a 5 digit ZIP is present. This is accomplished by using multiple hypotheses for the street number and ZIP to generate small lexicons of possible street names and then performing matching of entire addresses to USPS databases.

One difficulty in this approach is correctly locating the street numbers. It is not known if a street number is present in a given address or, if one is present, what line it is on. Ambiguities between alphabetic characters and numerals make it difficult to determine if a street number is present on a given line or not. Furthermore, the variable number of digits in a street number and the same character-numeral ambiguity make it difficult to determine where a street number ends and the street name begins (since we cannot read the street names until after we have used the street numbers to generate lexicons).

We compare the use of a fuzzy logic based system with a neural network system for locating street numbers under these conditions. A segmentation-based, lexicon free, numeric field recognition system is used together with a variety of image processing measurements and character recognition scores to generate features for the fuzzy and neural detection algorithms. All results will be presented on a set of address images from the USPS mail. A discussion of the potential applicability of the techniques to other problems such as check processing may be included.

Back to Bay Area OCR home page