Software spots forged signaturesBy Kimberly Patch, Technology Research News
Researchers from the University of Buffalo are working on a handwriting analysis system aimed at determining who penned a ransom note or forged a check.
The move was motivated by several high court rulings that required expert testimony to be substantiated by scientific evidence, according to Sargur Srihari, a professor of computer science and engineering at the University of Buffalo and director of the school's Center for Document Analysis and Recognition. "Since individual handwriting had not been subjected to such study, we undertook this work," he said.
The researchers' handwriting recognition program extracts handwriting features like character shape, line and word separation, and stroke slant and thickness. To compare handwriting from different documents, the program computes the differences in these features between the two samples measured, then determines if the differences fall within the limits of an individual writer's variability, according to Srihari.
In tests to determine whether or not two documents were written by the same person, the results were 95 percent accurate, said Srihari. In a second type of handwriting task, where the program determines which member of a known group of writers wrote a certain document, the performance varies from 98 percent for two writers to 89 percent for 1,000 writers, he said.
The program is based on a pair of well-known pattern recognition algorithms: Artificial Neural Network and Nearest Neighbor. It differs from conventional handwriting recognition in a key way, however. "Our work is about determining the variability between writers and within writers, [while] in handwriting recognition, the goal is to identify the message by averaging out the differences between writers," said Srihari.
The researchers' next steps are to capture finer features of handwriting in order to increase the accuracy. "For example, we currently are not measuring very many word-level features such as ascenders or descenders, or the presence of garlands or arcades," Srihari said. Ascenders and descenders are the parts of lowercase letters that extend above or below most other lowercase letters. Garlands resemble circles and arcades, arches.
Capturing features like these should make it possible to boost accuracy past 99 percent, he said.
The research is scientifically thorough, and is certainly useful, but is aiming for a difficult goal, said Nasser Sherkat, an associate professor of real-time machine vision at Nottingham Trent University in England.
"Variability within a writer' s [handwriting] is very high, especially when time elements and conditions of writing are taken into account. The ultimate question is whether we can tell if the writer is the same when the constraints [used in the study] have been removed," Sherkat said.
The program can eventually be used as the basis for human experts testimony in court, and eventually as a purely objective handwriting evaluation tool that doesn't require human intervention, said Srihari. This would be useful, said Sherkat. "Checking for forgeries manually is expensive and we could do with automating at least part of the process," he said.
The software could be available in less than a year, said Srihari.
Srihari's research colleague is Sung-Hyuk Cha of the University of Buffalo. The research was funded by the National Institute of Justice.
Timeline: < 1 year
TRN Categories: Applied Computing
Story Type: News
Related Elements: Technical paper, "Handwriting Identification: Research to Study Validity of Individuality of Handwriting and Develop Computer-assisted Procedures for Comparing Handwriting," downloadable from www.cedar.buffalo.edu/NIJ/publications
March 21, 2001
Net inherently virus prone
Rubber stamp leaves electronic mark
Software spots forged signatures
Magnetic transistor means changeable chips
Filters distill quantum bits
Research News Roundup
Research Watch blog
View from the High Ground Q&A
How It Works
News | Blog | Books
Buy an ad link
Ad links: Clear History
Buy an ad link
© Copyright Technology Research News, LLC 2000-2006. All rights reserved.