Iris recognition trumps facial recognition, says report
WASHINGTON—Iris recognition technology's performance in one-to-many applications is getting faster, is more accurate than facial recognition, and five organizations are leading the pack in terms of quality, according to what the National Institute of Standards and Technology (NIST) says is the first public and independent test of commercially available iris recognition software.
Released in April, the NIST report evaluated 92 iris recognition algorithms from 11 organizations—nine private companies and two university labs—on their one-to-many performance. The task was to identify individuals from a database, provided by the U.S. government, of eye images taken from more than 2.2 million people.
The Iris Exchange (IREX) III report revealed that no software was perfect, but that some iris recognition systems performed better than others on various technical indicators. Success rates ranged between 90 percent and 99 percent, and some produced as many as 10 times more errors than others. The tests also found that the speed of an algorithm—some were fast enough to run through a data set equivalent to the size of the entire U.S. population in less than 10 seconds using a typical computer—came at the expense of accuracy.
Patrick Grother, a scientist in NIST's Information Access Division, declined to say one algorithm was "best" since that qualitative judgment is "application dependent."
"I would caution against trying to pick a best algorithm," Grother told Security Director News. "The report advances a number of technical indicators, but depending on the user's application one could warrant more weight than others in deciding which algorithm to use, and so the best algorithm might mean the most accurate, or it might mean the fastest algorithm, or the algorithm with the smallest template size, or the one that's most stable across different cameras, or a number of other things."
However, with a little additional prodding, Grother identified five providers that the tests determined offer fast, accurate and mature algorithms. They are L1 Identity Solutions; Safran Morpho (Safran acquired L-1 last year); Cogent; Lithuania-based Neurotechnology; and a non-commercial entity, the University of Cambridge. (Cambridge's algorithm is noteworthy because it was supplied by John Daugman, a professor of computer vision and pattern recognition, and widely considered one of the primary inventors of iris recognition technology.)
"You ask for the best and on any given technical indicator any of those might be best," Grother said. Other companies that submitted algorithms were Crossmatch Technologies; IrisID; Iritech; Kynen; Smartsensors; and Southern Methodist University. Some organizations chose not to submit their algorithms to the tests, so not every technology provider is represented. Though, Grother said the test is still relevant as it included "a large majority of commercial providers."
One of the reasons Grother cited for the wide variations in effectiveness is how well the software could find the iris in low-quality images. "The challenges are finding the iris in the presence of reflections and maybe some motion blur and an occlusion from an eyelid," Grother said. "If you can find the iris then you can proceed to find the info within the iris texture, and so that's what makes some algorithms better than others, the capability to do that."
The report also concludes that iris recognition technology is "a more uniquely identifying technology" than facial recognition, Grother said.
"What the report shows is that face [recognition] is incapable of doing things that iris [recognition] can in terms of a accuracy," Grother said. "The iris simultaneously offers you low miss rates and also low false-alarm rates, whereas with the face you can't get low values for both of those things."
Concerning the accuracy of iris recognition technology, NIST issued a related report on how accuracy could be improved on the front end by the camera operators who are collecting the iris images. "It's pretty much true across all of biometrics that considerable benefits can be reaped by having some care, attention and training applied to the acquisition process," Grother said. "Because, as usual, garbage in, garbage out. You can't send broken data to the matching algorithm and expect to get anything useful."
Grother believes the NIST evaluation will be useful to end users and systems integrators who need to choose between providers, or between using one eye or two, or using iris recognition over some other biometric, like the face. However, the NIST evaluation may have a bigger impact on the technology developers themselves, helping them design the next generation of iris recognition systems.
"There's a lot of quantitative data that is understandable by the engineers who built the algorithms and they can say, 'Ah ha! We didn't know our algorithm did that,' [because] they routinely don't have the number of images that we used in this test," Grother said. "So they will glean useful information for algorithm development going forward."
NIST plans to continue testing iris recognition systems with IREX IV, which will have slightly different objectives and perhaps contain some new algorithms, Grother said.