The software developed for the post office was modified somewhat to zero in on differences in the handwriting. Such things as the spacing of letters, and how the letters are slanted, and how "loops" in numbers and letters are opened and closed, were easily recognized by the software. There is considerable variation in how hard we press down when writing, and the computer could detect that by the darkness of the letters.
The team ended up with 11 features that characterize the structure of the writing, such as the size of margins and spacing of characters and lines, and 512 features of individual letters and numbers. Two people might make an identical "e," but it is less likely the same two people would also make an identical "b," and so on.
"So we set up an experiment for our computer," Srihari says. "Could the computer program tell whether two documents were written by the same person, or two different people?"
In each case the researchers knew the answer, "and we wanted to see if the computer could figure it out also."
The results were a bit astounding. If given a substantial document, say a complete page of handwriting, the computer got it right 96 percent of the time. Even if just given a couple of words, the computer still nailed it more than 80 percent of the time.
"That's pretty high, and it's surprising because we were using pretty simple features," Srihari says.
A human, properly trained, should do even better, he adds.
There are far more sophisticated features, like style of writing and phrasing, that are much more difficult to analyze with software than the simpler features used by the researchers. Yet they are routinely used by handwriting experts.
One area that Srihari's research did not address is how hard, or easy, it might be to fool the computer, or the expert. It might not be all that difficult to trick the computer, but he suspects it would be a lot tougher to hoodwink the human expert.
A human, for example, might detect changes in writing style in a ransom note designed to disguise the identity of the writer, but still pick up individualistic features that would point to the true author.
"One could use all kinds of approaches to overcome the disguise issue," he says.
Srihari found the research quite convincing, and it is being published in the July issue of the Journal of Forensic Sciences. It won't convince everyone, of course, because some still regard handwriting analysis as akin to astrology.
And some will have ample reason to challenge the results. A criminal case could hinge upon the determination of whether a defendant really penned that ransom note, so lawyers will have much to fight about.
But so far, the findings have held up well.
"There have been a couple of federal court cases where the decision went in favor of admitting handwriting evidence" after hearing his report, Srihari says.
Maybe this is the guy we need to sic on the spammers. Surely his computer program could find out who they are, give us their home addresses, and put a few of them in jail.
Lee Dye’s column appears weekly on ABCNEWS.com. A former science writer for the Los Angeles Times, he now lives in Juneau, Alaska.