Facing a challenge
The federal government has been researching facial recognition technologies since the 1960s and, although it has significantly improved over the decades, it is still far from meeting operational needs.
'Facial recognition is the weakest of the biometric modalities,' said Matthew Malanowski, an intelligence specialist at the Army Intelligence Center at Fort Huachuca, Ariz. 'People have to be full frontal into the camera, and you can't do anything covert.'
Facial recognition is one of the three methods, along with fingerprint and iris recognition, used by the Defense Department's Biometrics Automated Toolset. BAT, which resides on 75 networked servers, is used by 1,500 clients in Iraq and at the National Ground Intelligence Center and other commands worldwide. To improve BAT's accuracy, the Army in April issued a request for information (GCN.com/1159) to assess the state of the technology in being able to recognize people under less than ideal conditions. Those conditions include when the face is turned, under poor lighting or the when the face is partially obscured by a branch.
'What we are looking for is software that can manipulate a two-dimensional capture and make it three dimensional,' said Malanowski. That way, other images of a person can be compared with the 3-D image, no matter what angle the person is facing, according to the RFI.
'Depending on the resolution of the input facial images and the amount of manual intervention during the modeling process, the results can be highly accurate,' said Demetri Terzopoulos, Chancellor's Professor of Computer Science at UCLA's Computer Science Department.
In the RFI, the Army was looking for a 95 percent identification rate with no more than a 1 percent false acceptance rate. The RFI also called for a module that would identify facial profiles with a 90 percent identification rate and no more than 10 percent of false positives.
The Army evaluated responses from five vendors and found them lacking.
'The entire purpose of the RFI was to determine whether the capability exists now,' said Malanowski. 'We determined it does not; that technology needs to be developed.'
So, the Army plans to release a request for proposals to create the software needed to move closer to what is needed.
In the meantime, other agencies are also trying to develop better face-recognition software. The interagency Technical Support Working Group has spent more than $40 million on such research over the past decade. The National Institute of Standards and Technology has conducted a series of Face Recognition Vendor Tests (GCN.com/1160). To the test
From 2004 to 2006, the FBI and other agency groups sponsored the Face Recognition Grand Challenge, in which researchers and developers were given a set of increasingly difficult set of problems to solve, using a data set of 50,000 images. This year the group is hosting the Multibiometric Grand Challenge, which combines face recognition with iris recognition.
'Progress has been made in increasing the degree of modeling automation and the imaging technology has improved, increasing the resolution, which yields more accurate facial geometry and texture,' Terzopoulos said.
Since last summer, Portugal has been using the eBox RAPID system at its airports in Lisbon and Faro. In less than 20 seconds it verifies individuals match the data in the electronic passport and opens a gate to let them through. This is fairly simple in cases where there is a one-on-one match between one person and one passport. However, the chance of error increases logarithmically with the number of identities in a database. So, when dealing with something the size of the terrorist watch list ' nearly 1 million identities ' the automatic recognition systems need to be near perfect.
'You have people who are going through security at an airport, you should be able to match them with a watch list,' said Malanowski. 'Ideally that would be great, but we don't have that capability yet.'