Ever wondered what truly identifies us as being ourselves? What distinguishes us from more than six billion people with whom we share this world? The answers can be surprising. We tend to think that our name, address, date of birth, ID card number is parameters sufficient for our positive identification. Unfortunately, matters are not that simple. Information age has sparked a staggering growth in illicit exchange of personal information. And since these traditional means of identification are not an intrinsic part of who we are, crimes such as identity theft are fast becoming ubiquitous.
Unique identification.
Biometrics involves identification based upon a person’s unique physical attributes. Methods of biometric identification include fingerprinting, DNA matching, retinal scanning, voice recognition, and face recognition. Biometrics is expected to grow by an order of magnitude, from a $200 million business in 1998 to $2 billion industry by 2004. Current major application of biometrics is authentication and identification for security purposes.
Specific biometric methods have varying applicability in different situations. Fingerprinting and DNA matching are ideal when a suspect has been arrested and detained. Retinal scanning and voice identification are excellent techniques for granting restricted access to high security areas such as government or industrial research facilities. While Face Recognition is considered an optimum technique for airport security where DNA matching or retinal scanning of all passengers is simply unrealistic.
History.
Face Recognition Technology, or (FERET), as termed now by US Department of Defense, is considered a valuable addition to the crime fighting toolbox. Although facial recognition techniques are not new to the research communities, the first commercial products began appearing not until mid 1990s. Interestingly, the biggest users, other than the law enforcement agencies, were the casinos! A privacy debate ensued after the Super Bowl in 2001, where cameras took pictures of all the fans entering the stadium and compared their pictures with the existing criminal database, without informing them.
Theory
To a computer, a face, like any other image, is a matrix of several hundred by several hundred pixels. Dealing with many faces, in the form of pictures, can be very time consuming and difficult. Therefore we have to use a more intelligent technique to make the task manageable.
There are two different approaches being used currently.
- Local feature analysis.
- Principal component analysis.
Feature analysis: This technique uses facial ‘landmarks’ like the tip of the nose or the distance between eyes to distinguish among different people. This is also called template matching. A topographical map is generated by determining the position of eyes, nose, and mouth to each other. Once the map is created the pixel data is discarded and only the landmarks on the map are used for matching.
Due to statistical regularity of human faces, if we can locate just a few landmarks on the face we can predict the position of others. Hence LFA provides a redundant description. Also it is much less computationally intensive to compare a highly compressed description of landmark configuration rather than the whole facial image.
Component analysis: The basic idea here is that face images can be economically represented by projecting them onto a small number of basis images called eigenfaces.
Eigenfaces are stereotyped building blocks that can be weighted and combined to reconstruct real face. Unlike local feature analysis, the eigenface approach always looks at the face as a whole. This is a more intuitive approach since it is based on how the human mind perceives and distinguishes among different images.
These eigenfaces are the basis images that are formed by finding the most significant eigenvectors of the pixel wise covariance matrix for a set of training images. They form a much smaller dimensional space upon which a test image can be projected and compared.
Neural net.
This is a relatively new technique being used for face recognition. The Back Propagation algorithm is used here for identification. There are some other innovative techniques like PCA on Wavelet Sub band, Hidden Markov Model, etc. that are still on research stages.
Different techniques.
The Local Feature Analysis has high discrimination power but it depends heavily on feature extraction algorithm. A small error in landmark extraction can cause high rate of false identification. On the other hand the Principal Component Analysis has a high computational load but is much more stable.
Commercial products.
- FaceIt System by Visionics Corp. This system uses local feature analysis for identification.
- FaceFinder by Viisage Technology Inc. The FaceFinder software is based on the eigenface approach.
Just like other emerging technologies, for instance stem cell research or digital copyrighting, face recognition has its share of opponents. Despite the impressive figures of research paper turnout and seminars concerning this subject, it is still considered a controversial field. The main objections being the inadequacy of current criminal databases and the threat it poses to privacy.
These are serious concerns but at least for airports, at about five cents a ticket, this certainly is a cost effective security tool. As for privacy invasion, according to Thomas J.Colatosti, president of Viisage Inc, “the technology only automates what is easily observable by any law enforcement officer . . . its only the privacy of criminals that is in danger.”
Moreover, the technology is a better alternative to racial and ethnic profiling. Not to mention other applications such as automatic search in visual databases, contact less man machine interaction, very low bit rate compression and image tracking.