Hendon Publishing - Article Archive Details
Facial Recognition, the Real You
Recognizing the face of a criminal or terrorist at a specific location and time is a challenging task. However, the human face has certain features that lend themselves to computer analysis and matching to known suspects. The face represents a biometric parameter that can be analysis for identification purposes. Other biometric parameters commonly in use include fingerprints, iris and retina patterns, hand geometry, vein patterns, and speech and handwriting patterns.
Facial recognition has become an important biometric tool, especially in situations where a large number of people are being screened in a short period of time, e.g., at airports, border crossings or in buildings with high security areas.
Facial recognition initially dealt with comparison of a 2-dimensional picture of an individual standing before a screening device. As computer technology has progressed over the past two decades, face recognition has advanced to allow a real-time picture of the individual to be compared with a reference picture of that individual in the system database.
The human face is probably an individual’s most unique physical characteristic. For this reason, electronic systems can capture and compare one face to other faces in a databank stored in the computer. This biometric analysis is effective in recognizing people with the exception of identical twins. The face has a number of unique physical features that lend themselves to biometric analysis.
There are 80 specific features, called nodal points, which can be used for analysis. Facial-recognition systems routinely use only 15 to 20 of these nodals. Facial-recognition software measures features such as the size and shape of the chin, the width of the nose, the distance between the eyes, the shape and position of cheekbones, the jaw line, and the shape and thickness of the lips, etc. You can observe these features by looking at yourself in a mirror.
The conventional camera captures a picture of the person on a film that must be chemically developed to obtain the picture. The advent of digital camcorders and cameras has allowed a captured image to immediately be digitized, converted to bytes of data, and passed to a computer databank. With current facial-recognition systems, this capture and transfer can occur within a very small fraction of a second. This allows these systems to be used to capture faces in a moving crowd and process them for identification very rapidly. This is important in systems used for security monitoring at places such as airports.
The steps involved in facial recognition are the same for all systems and include detection, alignment, normalization, presentation, and matching. Detection involves using a video camera to scan a field of view for a face. When one is detected, a digital image is captured and passed to the software system, which starts a multi-level analysis of the data.
The first step is a low-resolution analysis of the detected image to ensure that it is a human head. Then analysis is switched to a high level analysis of the nodals of the image. Each of these steps require a different software algorithm, or set of computer instructions, to perform specific tasks. The alignment step is required because a face moving in a crowd will not usually be facing directly at the camera.
Systems vary, but in general, a face must be within in a certain range or degrees of angle toward the camera for it to register. Alignment allows the system to recognize the heads size, position and pose in order to normalize it, which is the next step in the process.
In the normalization process, the image is scaled to size and rotated so that it can be accurately mapped to images of faces contained in the known data bank. Commercial systems use different parameters such as verifying the positioning of eyes and ears, or head position to ensure that image capture is consistent in all cases.
The normalized facial-image data is now translated into a unique code in a process referred to as presentation, and the acquired data set is compared with or matched to facial data that is stored in FBI, DOJ or police databases. The National Institute of Standards and Technology (NIST) has developed a set of minimal standards for face-recognition systems. It has developed an image standard called CBEFF, common biometric exchange file format, which allows captured facial images stored in different databases to be shared between all vendors of equipment.
Software systems presently allow facial images to be altered to add beards and mustaches, changes hair color and styles, as well as other properties to show how an individual would look under different disguise conditions.
Problems with Facial Recognition
Initially, only black-and-white pictures were available taken from mugshots and other ID systems. More recently, color photo images are being introduced into the databases from passports, driver’s licenses and other sources. Because the problem of face alignment with the camera is a major source of false acceptances, many recognition system manufacturers are now developing software programs for their equipment that will construct a 3-D presentation of the head of the person.
People can disguise the appearance of their faces by growing beards, mustaches, and coloring eyebrows and hair in order to fool a facial-recognition system. However, they cannot change the basic shape of their heads. The nodals of the face can predict what the heads shape should be. This is based on data that has been developed from analysis of a wide variety of head shapes, and the data has been consolidated into a series of codes for each head shape. These codes are what the computer uses for the 3-D recreations.
Facial expressions present a major problem for recognition systems. In most cases, the facial image of an individual initially placed in a databank has a serious facial expression. The image captured by a surveillance system may have smiling, laughing or other expression, and the problem for the recognition system is to match this to the initial data set.
Software algorithms have been developed for many systems that correct for common facial expression variations. Expressions can be removed or added to images as required. These expression codes are developed by scanning different expressions on the same face and then developing a common set of variables to describe the expressions. The data from a large number of samples is then combined to form the final code.
Alignment of facial images for comparisons to known images presents another challenge. New software is being developed to produce 3-dimensional images that can be manipulated for proper alignment. This software takes the captured 2-D information, and through a mathematical process called smooth deformations, generates accurate and precise full 3-D wireframe image.
A wire frame image is a virtual image that can be rapidly processed, changed and then created as a minimal presentation of the 3-D image. Then the full 3-D image can be processed over the frame. This 3-D image can be rotated side to side or tilted up or down allowing a better comparison to the position of a face/head in a database image. It also allows a screener to make identification even from a side view of the person.
Over the past decade, there has been a large volume of scientific research on facial recognition. This has resulted in development of increasingly more sophisticated equipment and software for facial recognition. However, some things still remain a mystery in the recognition process. Among these are why these systems can more accurately recognize men (60%) more of the time than women (40%).
In addition, recognition of races also varies considerably; Caucasians (66%), Asians (25%), while others (11%) only a small percentage of the time. There is also a wide discrepancy in the ability to accurately recognize individuals of various ages. The current systems appear best at accurately identifying people in the 18- to 29-year-old age bracket, being accurate 82% of the time. The identification rate for older groups is much lower: the 30- to 39-year-old group is 11%, the 40- to 49-year-old group is 4%, and those older than 50 were identified accurately only 3% of the time.
Forensic Uses of Facial Recognition
In addition to video surveillance and criminal identification, facial-recognition systems are being used to reconstruct facial images of skulls from skeletons and to age pictures of victims missing in cold case investigations. A very important area is in the investigation of missing or exploited children. From photographs of a child around the time of abduction, this software is able to project what a child would look like many years later. The programs have been successful in finding missing children.
Finally, another area where this software is being used is in preparing composite drawings of suspects from witness information. Sketch artists are expensive, and many departments cannot afford an adequately trained artist. However, software now on the market allows computer sketches to be generated even by novices. These systems contain basic face structures and allow virtually any feature of the human face to be changed electronically. From the sketch, 3-D composites can be developed in a relatively short period of time and electronically distributed to officers in the field.
Doug Hanson, Ph.D., is a biochemist who has operated toxicology and analytical chemistry laboratories for more than 25 years. He also is a freelance writer who has written extensively for law enforcement, EMS and first responder magazines. He published a book, “The Eider Files,” which is a novel about bioterrorism. He can be reached at firstname.lastname@example.org.
Published in Law and Order, Sep 2006
Rating : 9.0
Click to enlarge images.