UB software speeds mail service

Funded by the U.S. Postal Service, the new Handwritten Address Interpretation (HWAI) technology is expected to enable the Postal Service to realize substantial savings in its processing costs because it reduces the need for manual keying.

It also demonstrates that researchers at UB's Center of Excellence for Document Analysis and Recognition (CEDAR), who developed and refined the technology during the past decade, are nearing a solution to one of the most difficult problems in artificial intelligence-reading human handwriting.

The Postal Service says initial test results have been outstanding.

"This project represents a major step forward, not only for the Postal Service, but for technology in general," said Edward Kuebert, manager of image and telecommunications technology at the Postal Service. "It will do the seemingly impossible-help postal machines read handwritten mail."

Handwritten addresses account for approximately 15-20 percent of the entire mail stream, but today's equipment reads less than 2 percent of all handwritten mail, according to the Post-al Service.

CEDAR's software will improve the automated reading of handwritten addresses by an order of magnitude. Upgrades to the system in the next six months will boost the success rate even further.

It automatically will read and process daily several million pieces of mail with handwritten addresses, with virtually no human intervention.

The HWAI software has been integrated into the automated mail-processing system using a Remote Computer Reader.

This enhancement has been installed in 34 U.S. Postal Service mail-processing and distribution centers across the U.S. It is expected to be installed in a total of 250 processing and distribution centers in the U.S. during 1997.

The HWAI technology approaches its task in steps. First, it locates the address block on the mail image. Next, the address is divided into lines and then into parts of lines, such as city, state and ZIP code. Then the software tries to read the numbers in the address, starting with the ZIP code and proceeding to the street number and even post office box number, if there is one. This interpreted address is matched against a database of valid, deliverable addresses for the actual solution.

"The goal of all this reading is for the computer to determine the correct address of the mail piece, which is then imprinted with a bar code to be read by subsequent automatic mail sorters either at the same post office or other post offices as it travels to its destination," said Sargur Srihari, UB professor of computer science, executive director of CEDAR and principal investigator.

He explained that because cursive handwriting tends to run together, it usually cannot be read character by character, the way machine-printed addresses are read by automatic mail sorters.

Instead, handwriting interpretation programs attempt to take the whole image of a word and then conduct a pattern-matching process, matching the image with a list of words that look like it.

The problem, he explained, is that if the list of possible matches the program comes up with is very long, it becomes unworkable.

"CEDAR's principal discovery was in making the list small enough so that word recognition would be successful," Srihari said.

The program constrains the list of possible words that match by first recognizing the ZIP code and then the street number on the address, explained Venu Govindaraju, CEDAR associate director.

"There are only a small set of street names that will correspond to both the street number and the ZIP code," he said.

The main architects of the technology are Srihari, Govindaraju and CEDAR research scientists Paul Palumbo; Evelyn Kleinberg; Ajay Shekhawat; John Favata and Gyeonghwan Kim, as well as several former CEDAR researchers.

During the past five years, hundreds of undergraduate and graduate students at UB and postdoctoral researchers at CEDAR have contributed to the success of the HWAI system.