A backprogagation neutral network in an address block classifiction system
MetadataShow full item record
The U.S. Postal Service (USPS) is investing heavily in research and development of automated mail handling systems. A major component in these systems is the use of Optical Character Recognition (OCR) to read the destination address and ZIP Code, and then bar code the mail piece. High speed sorting equipment can then sort the mail using the bar code. Current USPS OCR/automated mail handling systems only process letter mail (no automated address-reading systems exist for nonletter mail).
Moreover, these OCR systems only capture and read a restricted field-of-view image. Letters can be rejected by these OCR systems because of nonstandard address location (outside the field-of-view), skewed address lines, or handwritten addresses. Current research is working toward building OCR systems capable of processing all forms of mail which include letters, flats, and irregular parcel and pieces (IPPs). These systems must scan an entire mail image for the destination address block which can assume any orientation. For nonletter mail, such as magazines, this is an exceedingly difficult task, since the entire face of up to 11 by 14 inches must be searched, and the address block must be chosen from all the other extraneous nonaddress information.
This paper details an experimental address block location system developed at MITRE. The" system uses a backpropagation neural network trained to discriminate the frequency characteristics of address blocks from other candidates. The current system is trained on magazine flat mail.
- Masters Theses