A Neural Net that Knows Faces

TrueFace from Miros (Wellesley, MA) is a neural-network-based facial verification system and one of the first of its kind to be commercially available. It runs under Windows on a 486- or Pentium-based PC. TrueFace is available as a software system (the recognition program; a demonstration program; a C-callable 32-bit library; sample application source code; sample facial images; documentation; and a security dongle) for $4850 in single-unit quantities (the price for 11 or more units is $1750). Miros also offers TrueFace to integrators who incorporate the system into solutions for security at hospitals, businesses, and immigration stations.

Miros sells a proprietary TrueFace neural-network algorithm that has been tr ained on what a face is and how to compare faces. In live use, the algorithm compares a live face image with a compressed reference face image encoded on an identification card or in a computer database. Compressed images require only 500 bytes. The algorithm then decides if the two images are of the same person. The algorithm is adaptive, enabling it to accommodate changes in appearance, like hairstyle.

One unique property of the TrueFace system is an adjustable threshold of acceptance. Some businesses (e.g., banks) place a premium on not irritating their customers. As Dr. Michael Kuperstein, president of Miros, observes, ``Security adds inconvenience.'' Such businesses might select a lenient threshold, which although it might be less accurate, would be less likely to annoy customers. On the other hand, extremely secure installations might require an extremely stringent threshold.