That face! Those eyes! How recognizable?
Technology for computerized facial recognition is improving, according to a recent NIST report.
- By Wilson P. Dizard III
- Apr 03, 2007
Technology for computerized facial recognition is ten times more accurate now than it was four years ago, and the best of the systems outperform humans, the National Institute of Standards said.
The federal government has pressed the private sector to improve facial and iris recognition technology dramatically so as to pave the way for improved biometric systems, and NIST has overseen the process in tests called the Face Recognition Vendor Test (FRVT) 2006 and the Iris Challenge Evaluation (ICE) 2006.
The facial-recognition test has compared vendor systems on in their ability to recognize high-resolution still images and three-dimensional facial images, under both controlled and uncontrolled illumination. The ICE 2006 test reported iris recognition performance from left and right irises. The study compared the facial recognition test results with an earlier evaluation called the FRVT 2002. ICE 2006 reported iris recognition performance from left and right irises.
According to a NIST report issued in late March
, the facial recognition systems it tested in the FRVT 2006 trials showed an 'order of magnitude,' or tenfold, improvement over comparable tests conducted four years ago.
The NIST study found that under the tests' conditions, the algorithms' recognition performance was about the same for very-high resolution still face images, 3-Dthree-dimensional face images and single-iris images.
The dramatic performance improvement was one of the goals of the government's Face Recognition Grand Challenge. 'In an experiment comparing human and algorithm [system] performance, the best-performing face recognition algorithms were more accurate than humans,' NIST reported.
Eight authors, including specialists from NIST, Science Applications International Corp, Schafer Corp., Notre Dame University and the Universityuniversities of Notre Dame and Texas worked together on the final research paper, titled "FRVT 2006 and ICE 2006 Large Scale Results."
In general, state-of-the-art facial recognition systems have reduced their error rates from about 0.73 percent in a 1993 evaluation that was partially automated to 0.01 in the fully automated FRVT 2006, the report said.
The test results rated performance by algorithms from 14 organizations, including vendors and universities.
The NIST evaluation team reported major differences in the speed at which different facial- recognition algorithms processed iris images. One of the mathematical routines, called Cambridge-2, processed the experimental iris data in six hours. Two others, called the Sagem-Irdian or SI-2 and Iritech or Irtech-2 algorithms, took about 300 hours each. The iris data tests sets were composed of about 60,000 right- and left- eye images from 240 people.
The NIST tests were nothing if not statistically sophisticated, and gleaned large quantities of data about the algorithms' performance in different variations of the evaluation, such as lighting conditions and the resolution of the test data.
The study authors emphasized in a footnote that they were not implying endorsement of any of the products associated with the algorithms tested. The complexity of the data and the variations in the algorithms' performance, as well as their own professional caution, likely led the scientists to shun a simple algorithm ranking. The statistical data did not appear on its face to show that any of the algorithms is unacceptable.
L-1 Identity Solutions Inc. issued a statement pointing to the fact that its facial and iris recognition algorithms achieved tier one performance results in both the FRVT 2006 and the ICE 2006.
'With performance results near or at the top of every single test conducted, the NIST test results offer important third- party validation of the top-tier performance and functionality of the algorithms that underpin L-1's facial and iris biometric solutions,' the statement read.
L-1 was formed earlier last year when Viisage Technology Inc. bought Indentix Inc. for $770 million in stock. The NIST evaluations of facial and iris technology covered algorithms submitted by both Viisage and Identix.