- The Washington Times - Thursday, April 17, 2003

Is computerized face-recognition technology ever really going to work? The answer appears to be maybe.
Recent large-scale tests administered by the National Institute for Standards and Technology and other federal agencies show the technology has improved a lot recently, but falls short of perfection.
What with the 2001 USA Patriot Act and the Enhanced Border Security and Visa Entry Reform Act of 2002, there is certainly a market for people-recognition technology, and a lot of firms are looking to make money on it.
Many applications are obvious. Security people at the airport cannot possibly remember the faces of 5,000 criminals and suspected terrorists. A computer can. Further, a computerized watch list is easily updated. If the embassy in London gets a picture of a known terrorist, it can be distributed to all the world's airports instantly.
The idea of face recognition, which has been around for years, is that a computer attached to a camera can compare a photograph stored in memory to an actual face and decide whether they are the same person. (Actually, it would more likely store a set of numeric identifiers calculated from the photograph, rather than the photo itself.)
Such technology is already being used by police in Virginia Beach and Tampa, Fla., to track down criminals. Police in these cities are involved in a digital manhunt for criminals by pairing surveillance cameras with the new identification technology.
Public officials in Virginia Beach are using biometrics technology to identify people by using algorithms that measure faces, fingerprints and irises to help them locate criminals wanted on outstanding felony warrants.
How does it work? Consider the different proportions between different measurements that can be taken of a face. The ratio of the width of your nose to the space between your eyes is probably different from the like ratio calculated from my face.
Your ears may stand out further from your head than do mine. Comparing a large array of such proportions improves the likelihood of an accurate identification.
Real systems use much trickier approaches. And still it's hard. As a geek friend of mine puts it, "Computers are inherently good at adding numbers. They are not inherently good at recognizing people. You have to fool them into it."
Years back, a company called Viisage Technology, based in Littleton, Mass., was much in the news, with journalists suggesting that its technology, originally developed at the Massachusetts Institute of Technology, would be used to catch criminals when they entered buildings equipped with cameras.
The recent federal government tests were done with huge sets of photos from such organizations as the State Department. The conclusion was that it worked reasonably well. It works better indoors, where lighting is controlled, than outdoors. It works better for old people than young people. Men are easier to recognize than women. It's hard to be more specific because the results get into false positives, verification as distinct from identification, and such.
The test found that successful identification using face recognition depends on the number of faces. For the best system, the top-rank identification rate was 85 percent on a database of 800 persons, 83 percent on a database of 1,600, and 73 percent on a database of 37,437. As the watch-list size increases, performance decreases.
Now, you might as a citizen think that an 85 percent chance of catching a terrorist isn't good enough. But if you were a terrorist, would you take an 85 percent chance of 30 years in the slammer? It's really boring there.
False alarms are, of course, not good: People do not like being treated as terrorists when they are not. But this may be a matter for tact rather than abandonment of the technology.

Copyright © 2018 The Washington Times, LLC. Click here for reprint permission.

The Washington Times Comment Policy

The Washington Times is switching its third-party commenting system from Disqus to Spot.IM. You will need to either create an account with Spot.im or if you wish to use your Disqus account look under the Conversation for the link "Have a Disqus Account?". Please read our Comment Policy before commenting.


Click to Read More

Click to Hide