Research Breakthrough:
A new approach for reliable face recognition

One of the most well-known biometric identifiers is also the most basic--the human face. Everyday we use faces to identify the people we know. But what happens as we age? Or what if we grow our hair out, or cut it short, or what if we grow a full beard when once we were clean shaven? These kinds of things occur everyday and are a challenge for biometric systems--especially when security is involved.

Not only do people change over time, they change during the day. “Two pictures of the same person can be more different than pictures of two different people,” says computer science professor Harry Wechsler. “I would look very different tomorrow morning after a transatlantic flight, unshaven and exhausted, than I would today when I board.”

Wechsler is an internationally known expert on the topic of facial recognition. The director of Mason’s Distributed and Intelligent Computation Center, Wechsler was involved in the development of the FERET database, which has become the standard facial database for benchmark studies, experimentation, and performance evaluation. The FERET database made it possible for researchers to conduct analysis and develop algorithms using a common database.

One of the major challenges for facial recognition software is how to account for these superficial changes. Physical changes to someone’s face aren’t the only complicating factors. The distance at which an image is captured, along with lighting, pose, and distortion are other factors that must be considered when comparing facial images, says Wechsler.

Current security systems using the face biometrics are particularly ineffective when temporal changes, involuntary or not, occur. In addition, the working hypothesis for face recognition research so far has not been particularly concerned with the possibility that subjects would seek to foil their true biometric signatures. The very purpose of biometrics, however, is to provide security from impostors and from those seeking to breach security. It is obvious that such subjects are well motivated to interfere with the proper acquisition of their biometric signatures, and they will attempt to hide and/or alter the information needed for their identification. As occlusion and disguise usually affect only parts of the face, Wechsler proposes a novel recognition–by–parts approach for reliable face recognition.

The recognition–by-parts approach involves using individual face components and their sequential recognition. Faces can be partially occluded and/or disguised while parts of them remain unchanged, but the method proposed by Wechsler can still detect and authenticate them. How it works is the face representation, or picture, used spans a multi-resolution grid that captures partial information from the face at different scales to accommodate different surveillance scenarios including human identification from distance. The face parts represent the eyes, nose, mouth, eye and nose, nose and mouth, and the like. Using computer analysis, the parts that appear to be most relevant in facial appearance are captured and compared electronically, allowing the program to distinguish one face from another one through the use of comparison and elimination.

Wechsler has found that the recognition–by–parts approach is reliable enough to cope with geometry, illumination, and temporal changes, as well as missing or disguised parts of the face. Reliable face recognition of occluded and disguised faces will constitute a major step forward and will make face recognition the medium of choice for biometric identification. Surveillance in crowded environments using Closed Circuit TV (CCTV), when only parts of faces are visible from time to time will become feasible and will greatly benefit homeland security personnel and others.

Since face recognition technology requires little or no cooperation from the subject, it is becoming one of the top choices for biometrics and is starting to move into the commercial market. It is already being used by the government.

Over the years, Wechsler’s research has been funded by the Navy Research Laboratory, Technical Support Working Group/Department of Defense, and the Army Research Laboratory. In his most recent book, Reliable Face Recognition Methods: System Design, Implementation and Evaluation, Wechsler discusses his work and that of others in great detail, examining the evolution of face recognition research and exploring promising new avenues for research and development.

Yet experts say facial recognition will never be 100% accurate. False-positives (misidentifying someone) and false-negatives (not identifying the person) do still occur, but with Wechsler’s groundbreaking research, accurate results are becoming more likely.

Article submitted by Jennifer Freeman

If you would like to collaborate or apply the technology and resources to your own research program,
DIC ’s web address is:
http://cs.gmu.edu/~wechsler/DIC_Center/

Back to the Main Page