|
|
 |
|
Law Enforcement
|
 |
|
 |
 |
Your face or mine?
Face recognition technology shows improvement
Jan 22, 2008
|
|
by Laura DiBenedetto Kenyon
|
On Jan. 28, 2001, Super Bowl XXXV security officials in Tampa, FL, used a face recognition system to scan the faces in the crowd and compare them against a database of images containing sought-after suspects. Over a dozen matches were made -- all of them false positive.
If face recognition technology were a football player, it made a spectacular fumble during Super Bowl XXXV. The event kicked off an industry-wide trend where end-users backed away from implementing the technology and sought other identification solutions.
But it's time to get FRT back in the game; in fact, the technology now has the potential to move up from bench warmer to all-star status.
Over the past several years, face recognition technology has improved by leaps and bounds, as illustrated by the 2006 Face Recognition Vendor Test, a large scale independent evaluation test of face recognition systems conducted by the U.S. National Institute of Standards and Technology.
Error Correction
As a number of companies produce facial recognition technology, it has been NIST's charge to determine the efficacy of the technology via its 2000, 2002, and 2006 FRVTs. To do this, the FRVT team takes a set of sequestered facial images and asks the vendor software to match two sets of images together. Based on the results, the FRVT team can then calculate the algorithims' performance. (Note, as the testing is done on a computer at NIST, the participants can't adjust their algorithms to adapt to the test, nor have the algorithms received a prior exposure to the pictures used in the testing. These measures help ensure the FRVT's results are unbiased.)
Dr. P. Jonathon Phillips is in a unique position to see the evolution of face recognition technology, as he has led the 2000, 2002, and 2006 FRVTs. "One of the key findings [of the 2006 test] is that performance is improved," Phillips said. "Between 2002 and 2006, the algorithms' performance has increased by a factor of 20."
 |
| Face recognition technology can be used for detection in highly populated areas like sports stadiums. |
According to Phillips, there are two major ways in which the technology has improved: its false accept and false reject rates. "When you measure performance, there are two types of errors an algorithm can make," he explained. "One is it could think I am someone I am not, the false accept rate." The most recent test found a false accept rate of one in 1,000, meaning "somebody got away with claiming to be somebody else only once in every thousand tries," said Phillips.
The other error occurs is when the algorithm doesn't recognize a face already in the system. "So, when I say, 'I'm a Jonathan Phillips,' and the computers says, 'No, you're not,' that's a false reject rate. "
In 2002, the false reject rate was 20 percent, Phillips said. But in the 2006 test, the best algorithms were performing at the false reject rate of 1 percent.
The other major finding in the 2006 FRVT is the technology's ability to make matches using less-than-ideal images. In this scenario, the false reject rate is about 12 percent.
"So the systems' performance in these uncontrolled situations is better than the corresponding systems in 2002 on the easier problem," said Phillips. What this means is that face recognition technology in the 2006 is better at identifying images -- taken in both controlled and uncontrolled lighting situations -- than its predecessors.
Better Images, Better Results
Dr. Joseph Atick, executive vice president and chief strategic officer, L-1 Identity Solutions, noted additional improvements in the technology. "One of the biggest accomplishments we have seen is the fusion of multiple algorithms as a means to speed up the identification process," he said. "We've seeing a layering of information from the face. There is technology that examines the facial geometry from a macro level and then another application that is scrutinizing the skin's texture. When used in tandem, the biometrics add an additional layer of accuracy."
Of course, it's not just the algorithms that have improved -- so have the digital cameras capturing the images. Today, many images are taken using 4-6 megapixel cameras, making the faces "bigger." As a face takes up more pixels in the picture, it becomes easier for algorithms to analyze.
The combination of these factors has led overall improvement in face recognition technology, making it a convenient way to identify people. It can be used when a user has verification (is this person who they say they are and are they allowed access) or detection (is this person on a watch list).
According to Atick, one way FR technology can be used for verification is via a "one to many" search engine, where a still facial image is compared against those in a database. He said it's already popular and in use for applications like diver's licenses, visas, and passports.
For detection, faces in live video are screened against a database and alerts are triggered when an accurate match is made. "This technology is more commonly used in highly populated areas like airports, casinos, and sports stadiums," he added.
Trend Toward Acceptance
Prior to the 2006 FVRT, it made sense for many end-users to use means other than face recognition technology for identification purposes. But based on NIST findings, that may start to change.
Operating from the perspective that many "identification systems" being used now are human, the NIST decided to directly test human and computer abilities in terms in terms of recognizing faces. The findings? "We found that computers are capable of outperforming humans," said Phillips.
While people are very good at recognizing familiar faces, it's pretty difficult for us to match unfamiliar faces. Therefore, if a solution is needed for identifying faces that have not seen before, the face recognition technology will likely do a better job than humans. "Based on the data we had in 2002, this would not have been the case," added Phillips.
With improved technology and ease of use, it should only be a matter of time before face recognition technology gains wider use. According to Roger Kelesoglu, senior business development executive at Cognitec Systems Corporation, the industry has entered a phase of stabilization.
"This trend includes maturing technologies and off-the-shelf, cost-effective product, constantly growing [the] customer base, and companies dividing into a few top technology providers and systems integrators or solution providers on the other hand," he said. "The market demand is accelerating, the challenge simply is to really understand and reply to the customer's needs."
Atick agreed: "Education is going to be the key driver for facial recognition adoption. People need to learn that this isn't your grandfather's facial recognition technology. Sure, as with any facial recognition technology, there were false starts initially -- but now we have validation that it's accurate and has proven to work effectively and efficiently to solve identification problems."
With a one in 1,000 false accept rate and a one percent false reject rate, can the technology continue to improve, or has it reached the apex of its abilities? Phillips believes the technology can and will continue to improve.
"The best performances have been achieved with frontal images in controlled illumination, or images in hallways that are still very large images," Phillips said. "There are challenges [in identifying] people from different angles, different poses, if the faces are smaller or under more constrained lighting conditions -- these are still challenges for the face recognition community itself."
MORE INFORMATION
Cognitec Systems Corporation cognitech-systems.com
Face Recognition Vendor Test frvt.org
L-1 Identity Solutions l1id.com
|
|
|
Government Video is part of the NewBay Media Community
copyright 2007© NewBay Media LLC.
810 Seventh Avenue, 27th Floor New York, NY 10019 T (212)378-0400 F (212)378-2160
|
|
|
