Dotting the I's, Crossing the T's

By

If handwriting recognition systems make your beanie twirl, go to the Department of Commerce this August, close your eyes, hold out your hand and say "please."

When you look at your palm, you might find the software for an operational handwriting recognition package, courtesy of the National Institute of Standards and Technology. The system represents the culmination of six years' work by scientists at NIST's Image Recognition Laboratory.

If handwriting recognition systems make your beanie twirl, go to the Department of Commerce this August, close your eyes, hold out your hand and say "please."

"It has the latest-greatest technology we've been putting together in the lab," said Michael Garris, a NIST computer scientist.

The software, which uses a combination of traditional image processing, neural nets and advanced statistics, is designed to be an "open box."

"They'll be able to open this up, see how it works and learn from it," Garris said. The software is written in a modular fashion to allow selective use of code. The program is in C and runs on Unix workstations.

The package isn't designed for commercial use, and lacks a friendly interface. Rather, it is meant to be a tool for computer scientists to speed the evolution of computer-based handwriting recognition.

According to Garris, the system performs "competitively" against marketed products and has an accuracy of 90 percent on digits, 70 percent on lower-case letters and 80 percent on upper-case letters.

Unlike machine print recognition systems, which are common and highly accurate, the irregularities of handwriting present special challenges for computers. Form-consuming federal agencies like the Social Security Administration and the Internal Revenue Service are in desperate need of better and faster recognition systems.

Loral Federal Systems, recent winner of the Document Processing System mega-contract to the IRS, will be a beta-test site for NIST's software this summer.

William Klein, advisory software engineer for Loral is eager run it and has particular interest in the character-recognition and image-separation codes. The NIST software can "learn" to decipher an individual's writing style -- like when an E and S connect -- and separate them.

"This is NOT ho-hum stuff," Klein said. "They're doing some very good research."

NIST's Garris expects the beta testing will be done by the end of the summer, when the software will be free to all comers.


NEXT STORY: Spies, PIs and Snitches on Net