Recently, many articles have appeared on Habré devoted to Google's face identification systems. To be honest, many of them smell like journalism and, to put it mildly, incompetence. And I wanted to write a good article on biometrics, it's not my first! There are a couple of good articles on biometrics on Habré - but they are quite short and incomplete. Here I will try to briefly outline the general principles of biometric identification and the modern achievements of mankind in this matter. Including in identification by persons.

The article has a continuation, which, in fact, is its prequel.

As a basis for the article, a joint publication with a colleague in a journal (BDI, 2009), revised for modern realities, will be used. Habré does not have a colleague yet, but he supported the publication of the revised article here. At the time of publication, the article was a brief overview of the modern market for biometric technologies, which we conducted for ourselves before launching our product. Applicability value judgments put forward in the second part of the article are based on the opinions of people who have used and implemented the products, as well as on the opinions of people involved in the production of biometric systems in Russia and Europe.

general information

Let's start with the basics. In 95% of cases, biometrics is inherently mathematical statistics. And matstat is an exact science, algorithms from which are used everywhere: in radars and in Bayesian systems. Errors of the first and second kind can be taken as the two main characteristics of any biometric system). In radar theory, they are usually called “false alarms” or “target misses”, and in biometrics, the most established concepts are FAR (False Acceptance Rate) and FRR (False Rejection Rate). The first number characterizes the probability of a false match of the biometric characteristics of two people. The second is the probability of denying access to a person with a permit. The system is better, the smaller the FRR value at the same FAR values. Sometimes a comparative characteristic of EER is also used, which determines the point at which the FRR and FAR graphs intersect. But it is not always representative. More details can be seen, for example,.
The following may be noted: if FAR and FRR for open biometric databases are not given in the characteristics of the system, then no matter what the manufacturers declare about its characteristics, this system is most likely incapacitated or much weaker than its competitors.
But not only FAR and FRR determine the quality of a biometric system. If this were the only way, then the leading technology would be DNA recognition of people, for which FAR and FRR tend to zero. But it is obvious that this technology is not applicable at the current stage of human development! We have developed several empirical characteristics to assess the quality of the system. "Forgery resistance" is an empirical measure that summarizes how easy it is to spoof a biometric identifier. "Environmental stability" is a characteristic that empirically evaluates the stability of the system under various external conditions, such as changes in lighting or room temperature. "Ease of use" shows how difficult it is to use a biometric scanner, whether identification is possible "on the go". An important characteristic is the "Speed ​​of work", and "The cost of the system". Do not forget that the biometric characteristic of a person can change over time, so if it is unstable, this is a significant minus.
The abundance of biometric methods is amazing. The main methods using static biometric characteristics of a person are identification by papillary pattern on the fingers, iris, facial geometry, retina, hand vein pattern, hand geometry. There is also a family of methods that use dynamic characteristics: identification by voice, handwriting dynamics, heart rate, gait. Below is the distribution of the biometric market a couple of years ago. In every second source, these data fluctuate by 15-20 percent, so this is just an estimate. Also here, under the concept of “hand geometry”, two different methods are hidden, which will be discussed below.

In the article, we will consider only those characteristics that are applicable in access control and management systems (ACS) or in tasks close to them. By virtue of their superiority, these are primarily static characteristics. Of the dynamic characteristics at the moment, only voice recognition has at least some statistical significance (comparable to the worst static algorithms FAR ~ 0.1%, FRR ~ 6%), but only in ideal conditions.
To get a feel for the likelihood of FAR and FRR, one can estimate how often false matches will occur if an identification system is installed at a gated organization with N staff. The false match probability of a fingerprint received by the scanner for a database of N fingerprints is FAR∙N. And every day, about N people also pass through the access control point. Then the error probability per working day is FAR∙(N∙N). Of course, depending on the goals of the identification system, the probability of an error per unit of time can vary greatly, but if one error per working day is accepted, then:
(1)
Then we get that the stable operation of the identification system at FAR=0.1% =0.001 is possible with the number of personnel N≈30.

Biometric scanners

Today, the concepts of "biometric algorithm" and "biometric scanner" are not necessarily interconnected. The company can produce these elements individually, or together. The greatest differentiation of scanner manufacturers and software manufacturers has been achieved in the papillary finger pattern biometrics market. The smallest 3D face scanner on the market. In fact, the level of differentiation largely reflects the development and saturation of the market. The more choice - the more the theme is worked out and brought to perfection. Different scanners have a different set of abilities. Basically, this is a set of tests to check whether a biometric object has been tampered with or not. For finger scanners, this can be a relief check or a temperature check, for eye scanners, this can be a pupil accommodation check, for face scanners, face movement.
Scanners have a very strong influence on the received FAR and FRR statistics. In some cases, these figures can change dozens of times, especially in real conditions. Usually the characteristics of the algorithm are given for some “ideal” base, or just for a well-suited one, where blurry and blurry frames are thrown out. Only a few algorithms honestly indicate both the base and the full FAR / FRR output for it.

And now in more detail about each of the technologies.

Fingerprints

Dactyloscopy (fingerprint recognition) is the most developed biometric method of personal identification to date. The catalyst for the development of the method was its widespread use in forensic science in the 20th century.
Each person has a unique papillary fingerprint pattern, which makes identification possible. Typically, algorithms use characteristic points on fingerprints: the end of the line of the pattern, the branching of the line, single points. Additionally, information about the morphological structure of the fingerprint is involved: the relative position of closed lines of the papillary pattern, "arched" and spiral lines. The features of the papillary pattern are converted into a unique code that preserves the information content of the print image. And it is the "fingerprint codes" that are stored in the database used for searching and comparing. The time for translating a fingerprint image into a code and its identification usually does not exceed 1 s, depending on the size of the base. The time spent on raising a hand is not taken into account.

As a source of data for FAR and FRR, VeriFinger SDK statistics obtained using the U.are.U DP fingerprint scanner were used. Over the past 5-10 years, the characteristics of recognition by the finger have not stepped forward much, so the figures given show a good average of modern algorithms. The VeriFinger algorithm itself has won the International Fingerprint Verification Competition for several years, where fingerprint recognition algorithms competed.

The typical FAR value for the fingerprint recognition method is 0.001%.
From formula (1) we obtain that the stable operation of the identification system at FAR=0.001% is possible with the number of personnel N≈300.

Advantages of the method. High reliability - the statistical indicators of the method are better than those of the methods of identification by face, voice, painting. Low cost devices that scan the fingerprint image. A fairly simple procedure for scanning a fingerprint.
Disadvantages: the papillary fingerprint pattern is very easily damaged by small scratches, cuts. People who have used scanners in businesses with several hundred employees report a high rate of scan failure. Many of the scanners do not adequately treat dry skin and do not let old people through. When communicating at the last MIPS exhibition, the head of the security service of a large chemical enterprise said that their attempt to introduce finger scanners at the enterprise (scanners of various systems were tried) failed - the minimal exposure of employees' fingers to chemicals caused a failure in the security systems of scanners - scanners declared fingers fake. There is also a lack of security against fingerprint forgery, partly due to the widespread use of the method. Of course, not all scanners can be fooled by methods from MythBusters, but still. For some people with “inappropriate” fingers (body temperature, humidity), the probability of access being denied can reach 100%. The number of such people varies from fractions of a percent for expensive scanners to ten percent for inexpensive ones.
Of course, it is worth noting that a large number of shortcomings are caused by the widespread use of the system, but these shortcomings do exist and they appear very often.

Market situation

Currently, fingerprint recognition systems occupy more than half of the biometric market. Many Russian and foreign companies are engaged in the production of access control systems based on the fingerprint identification method. Due to the fact that this direction is one of the oldest, it has received the greatest distribution and is by far the most developed. Fingerprint scanners have come a long way indeed. Modern systems are equipped with various sensors (temperature, pressing force, etc.), which increase the degree of protection against counterfeiting. Every day systems become more and more convenient and compact. In fact, the developers have already reached a certain limit in this area, and there is nowhere to develop the method further. In addition, most companies produce ready-made systems that are equipped with everything you need, including software. There is simply no need for integrators in this area to assemble the system on their own, since it is unprofitable and will take more time and effort than buying a ready-made and already inexpensive system, the more choice will be really wide.
Among foreign companies involved in fingerprint recognition systems, one can note SecuGen (USB scanners for PCs, scanners that can be installed in enterprises or built into locks, SDK and software for connecting the system to a computer); Bayometric Inc. (fingerprint scanners, TAA/Access control systems, fingerprint SDKs, embedded fingerprint modules); DigitalPersona Inc. (USB-scanners, SDK). The following companies operate in Russia in this area: BioLink (fingerprint scanners, biometric access control devices, software); Sonda (fingerprint scanners, biometric access control devices, SDK); SmartLock (fingerprint scanners and modules), etc.

Iris

The iris of the eye is a unique human characteristic. The iris pattern is formed in the eighth month of fetal development, finally stabilizes at the age of about two years and practically does not change throughout life, except as a result of severe injuries or severe pathologies. The method is one of the most accurate among biometric methods.
The iris identification system is logically divided into two parts: an image capture device, its primary processing and transfer to a calculator, and a computer that compares the image with images in the database, transmitting a command on admission to the actuator.
The time of primary image processing in modern systems is approximately 300-500ms, the speed of comparing the resulting image with the base has a level of 50000-150000 comparisons per second on a conventional PC. This speed of comparison does not impose restrictions on the application of the method in large organizations when used in access systems. When using specialized calculators and search optimization algorithms, it becomes even possible to identify a person among the inhabitants of an entire country.
I can immediately answer that I am somewhat biased and have a positive attitude towards this method, since it was in this field that we launched our startup. A paragraph at the end will be devoted to a small self-promotion.

Statistical characteristics of the method

The characteristics of FAR and FRR for the iris are the best in the class of modern biometric systems (with the possible exception of the retinal recognition method). The article presents the characteristics of the iris recognition library of our algorithm - EyeR SDK, which correspond to the VeriEye algorithm tested on the same databases. CASIA databases obtained by their scanner were used.

The characteristic value of FAR is 0.00001%.
According to formula (1), N≈3000 is the number of personnel of the organization, at which the identification of an employee occurs quite stably.
Here it is worth noting an important feature that distinguishes the iris recognition system from other systems. In the case of using a camera with a resolution of 1.3 MP, you can capture two eyes in one frame. Since the FAR and FRR probabilities are statistically independent probabilities, when recognizing in two eyes, the FAR value will approximately equal the square of the FAR value for one eye. For example, for a FAR of 0.001% using two eyes, the probability of a false tolerance would be 10-8%, with FRR only twice as high as the corresponding FRR value for one eye with FAR=0.001%.

Advantages and disadvantages of the method

Advantages of the method. Statistical reliability of the algorithm. Capturing an image of the iris can be performed at a distance of several centimeters to several meters, while the physical contact of a person with the device does not occur. The iris is protected from damage - which means it will not change over time. It is also possible to use a high number of methods that protect against forgery.
Disadvantages of the method. The price of a system based on the iris is higher than the price of a system based on finger recognition or face recognition. Low availability of ready-made solutions. Any integrator who comes to the Russian market today and says “give me a ready-made system” will most likely break off. For the most part, expensive turnkey systems are sold, installed by large companies such as Iridian or LG.

Market situation

At the moment, the share of iris identification technologies in the global biometric market is, according to various estimates, from 6 to 9 percent (while fingerprint recognition technologies occupy more than half of the market). It should be noted that from the very beginning of the development of this method, its strengthening in the market was slowed down by the high cost of equipment and components necessary to assemble an identification system. However, with the development of digital technologies, the cost of a single system began to decline.
The leader in software development in this area is Iridian Technologies.
Entry to the market for a large number of manufacturers was limited by the technical complexity of scanners and, as a result, their high cost, as well as the high price of software due to the monopoly position of Iridian in the market. These factors allowed only large companies to develop in the field of iris recognition, most likely already engaged in the production of some components suitable for the identification system (high-resolution optics, miniature cameras with infrared illumination, etc.). Examples of such companies are LG Electronics, Panasonic, OKI. They entered into an agreement with Iridian Technologies, and as a result of joint work, the following identification systems appeared: Iris Access 2200, BM-ET500, OKI IrisPass. In the future, improved system models arose, thanks to the technical capabilities of these companies to independently develop in this area. It should be said that the above companies also developed their own software, but in the end, in the finished system, they prefer the software of Iridian Technologies.
The Russian market is dominated by the products of foreign companies. Even though it's hard to buy. For a long time, Papillon assured everyone that they had iris recognition. But even representatives of RosAtom, their direct purchaser, for whom they made the system, say that this is not true. At some point, some other Russian company appeared, which made iris scanners. I don't remember the name now. They bought the algorithm from someone, perhaps from the same VeriEye. The scanner itself was a system 10-15 years old, by no means non-contact.
In the last year, a couple of new manufacturers entered the world market due to the expiration of the primary patent for recognizing a person by eyes. The most trusted of them, in my opinion, deserves AOptix. At least their preview and documentation does not arouse suspicion. The second company is SRI International. Even at first glance, to a person involved in iris recognition systems, their videos seem very false. Although I would not be surprised if in reality they can do something. Both systems do not show data on FAR and FRR, and also, apparently, are not protected from fakes.

face recognition

There are many face geometry recognition methods. All of them are based on the fact that the facial features and the shape of the skull of each person are individual. This area of ​​biometrics seems attractive to many, because we recognize each other primarily by the face. This area is divided into two areas: 2-D recognition and 3-D recognition. Each of them has advantages and disadvantages, but much also depends on the scope and requirements for a particular algorithm.
I will briefly talk about 2-d and move on to one of the most interesting methods today - 3-d.

2D face recognition

2-D face recognition is one of the most statistically inefficient biometric methods. It appeared quite a long time ago and was used mainly in forensic science, which contributed to its development. Subsequently, computer interpretations of the method appeared, as a result of which it became more reliable, but, of course, it was inferior and every year it is more and more inferior to other biometric methods of personal identification. Currently, due to poor statistical performance, it is used in multimodal or, as it is also called, cross-biometrics, or in social networks.

Statistical characteristics of the method

For FAR and FRR, data for the VeriLook algorithms were used. Again, for modern algorithms, it has very ordinary characteristics. Sometimes algorithms with an FRR of 0.1% with a similar FAR flash by, but the bases on which they were obtained are very doubtful (cut out background, the same facial expression, the same hairstyle, lighting).

The characteristic value of FAR is 0.1%.
From formula (1) we obtain N≈30 - the number of personnel of the organization, at which the identification of an employee occurs quite stably.
As can be seen, the statistical indicators of the method are quite modest: this eliminates the advantage of the method that it is possible to conduct covert shooting of faces in crowded places. It's funny to see how a couple of times a year another project is funded to detect criminals through video cameras installed in crowded places. Over the past ten years, the statistical characteristics of the algorithm have not improved, and the number of such projects has increased. Although, it is worth noting that the algorithm is quite suitable for leading a person in a crowd through many cameras.

Advantages and disadvantages of the method

Advantages of the method. With 2-D recognition, unlike most biometric methods, expensive equipment is not required. With the appropriate equipment, the possibility of recognition at considerable distances from the camera.
Flaws. Low statistical significance. There are requirements for lighting (for example, the faces of people entering from the street on a sunny day cannot be registered). For many algorithms, the unacceptability of any external interference, such as glasses, a beard, some elements of a hairstyle. Mandatory frontal image of the face, with very small deviations. Many algorithms do not take into account possible changes in facial expressions, that is, the expression must be neutral.

3-D face recognition

The implementation of this method is a rather difficult task. Despite this, there are currently many methods for 3-D face recognition. The methods cannot be compared with each other as they use different scanners and bases. far from all of them issue FAR and FRR, completely different approaches are used.
The transitional method from 2-d to 3-d is a method that implements the accumulation of information about a person. This method has better characteristics than the 2d method, but just like it uses only one camera. When entering the subject into the database, the subject turns his head and the algorithm connects the image together, creating a 3d template. And when recognizing, several frames of the video stream are used. This method is rather experimental and I have never seen implementations for ACS systems.
The most classic method is the template projection method. It consists in the fact that a grid is projected onto the object (face). Next, the camera takes pictures at a speed of tens of frames per second, and the resulting images are processed by a special program. A beam falling on a curved surface bends - the greater the curvature of the surface, the stronger the bending of the beam. Initially, this used a source of visible light supplied through the "blinds". Then visible light was replaced by infrared, which has a number of advantages. Usually, at the first stage of processing, images are discarded in which the face is not visible at all or there are foreign objects that interfere with identification. Based on the obtained images, a 3-D model of the face is restored, on which unnecessary interference (hairstyle, beard, mustache and glasses) is highlighted and removed. Then the model is analyzed - anthropometric features are highlighted, which are eventually recorded in a unique code entered into the database. Image capture and processing time is 1-2 seconds for the best models.
Also, the method of 3-d recognition based on an image obtained from several cameras is gaining popularity. An example of this is Vocord with its 3d scanner. This method gives positioning accuracy, according to the assurances of the developers, higher than the template projection method. But, until I see FAR and FRR at least in their own database, I won’t believe it !!! But it has been developed for 3 years already, and progress at exhibitions is not yet visible.

Statistical indicators of the method

Full data on FRR and FAR for algorithms of this class are not openly provided on the websites of manufacturers. But for the best Bioscript models (3D EnrolCam, 3D FastPass) working by the template projection method with FAR = 0.0047% FRR is 0.103%.
It is believed that the statistical reliability of the method is comparable to the reliability of the fingerprint identification method.

Advantages and disadvantages of the method

Advantages of the method. No need to contact the scanning device. Low sensitivity to external factors, both on the person himself (the appearance of glasses, a beard, a change in hairstyle), and in his environment (light, head rotation). High level of security, comparable to fingerprint identification.
Disadvantages of the method. Expensive equipment. The complexes available for sale were even more expensive than iris scanners. Changes in facial expressions and noise on the face degrade the statistical reliability of the method. The method is not yet well developed, especially in comparison with fingerprinting, which has been used for a long time, which makes it difficult to widely use it.

Market situation

Facial geometry recognition is one of the "three big biometrics" along with fingerprint and iris recognition. I must say that this method is quite common, and so far it is given preference over recognition by the iris of the eye. The share of face geometry recognition technologies in the total volume of the global biometric market can be estimated at 13-18 percent. In Russia, this technology is also showing more interest than, for example, identification by the iris. As mentioned earlier, there are many 3-D recognition algorithms. For the most part, companies prefer to develop turnkey systems that include scanners, servers, and software. However, there are those who offer the consumer only the SDK. To date, we can note the following companies involved in the development of this technology: Geometrix, Inc. (3D face scanners, software), Genex Technologies (3D face scanners, software) in the USA, Cognitec Systems GmbH (SDK, special computers, 2D cameras) in Germany, Bioscrypt (3D face scanners, software) is a subsidiary of the American company L- 1 Identity Solutions.
In Russia, Artec Group companies (3D face scanners and software) are working in this direction - a company headquartered in California, and development and production are carried out in Moscow. Also, several Russian companies own 2D face recognition technology - Vocord, ITV, etc.
In the field of 2D face recognition, the main subject of development is software, because Conventional cameras are great at capturing images of faces. The solution to the problem of face recognition has reached a dead end to some extent - for several years now, there has been practically no improvement in the statistical indicators of algorithms. In this area, there is a systematic "work on the bugs".
3D face recognition is now a much more attractive area for developers. It employs many teams and regularly hears about new discoveries. Many of the works are in a "just about to be released" state. But so far, only old offers are on the market; in recent years, the choice has not changed.
One of the interesting points that I sometimes think about and which, perhaps, Habr will answer: is the accuracy of kinect enough to create such a system? There are quite a few projects for pulling out a 3d model of a person through it.

Recognition by the veins of the hand

This is a new technology in the field of biometrics, its widespread use began only 5-10 years ago. The infrared camera takes pictures of the outside or inside of the hand. The pattern of veins is formed due to the fact that blood hemoglobin absorbs infrared radiation. As a result, the degree of reflection is reduced and the veins are visible on the camera as black lines. A special program based on the received data creates a digital convolution. No human contact with the scanning device is required.
The technology is comparable in reliability to recognition by the iris of the eye, surpassing it in some ways, and inferior in some ways.
The FRR and FAR values ​​are for the Palm Vein scanner. According to the developer at FAR 0.0008% FRR is 0.01%. No company produces a more accurate schedule for several values.

Advantages and disadvantages of the method

Advantages of the method. No need to contact the scanning device. High reliability - the statistical indicators of the method are comparable with the readings of the iris. Hidden characteristics: unlike all of the above, it is very difficult to obtain this characteristic from a person “on the street”, for example, by photographing him with a camera.
Disadvantages of the method. Exposure of the scanner to sunlight and rays of halogen lamps is unacceptable. Some age-related diseases, such as arthritis, greatly impair FAR and FRR. The method is less studied in comparison with other static biometric methods.

Market situation

Hand vein recognition is a fairly new technology, and therefore its global market share is small, around 3%. However, there is growing interest in this method. The fact is that, being quite accurate, this method does not require such expensive equipment as, for example, recognition methods based on facial geometry or the iris. Now many companies are developing in this area. So, for example, by order of the English company TDSi, software was developed for the palm vein biometric reader PalmVein, presented by Fujitsu. The scanner itself was developed by Fujitsu primarily to combat financial fraud in Japan.
Also in the field of vein identification are the following companies Veid Pte. Ltd. (scanner, software), Hitachi VeinID (scanners)
I don't know any companies in Russia dealing with this technology.

Retina

Until recently, it was believed that the most reliable method of biometric identification and authentication of a person is a method based on scanning the retina. It contains the best features of identification by the iris and by the veins of the hand. The scanner reads the pattern of capillaries on the surface of the retina. The retina has a fixed structure that does not change over time, except as a result of a disease, such as cataracts.
Retinal scanning uses low-intensity infrared light directed through the pupil to the blood vessels at the back of the eye. Retinal scanners have become widely used in access control systems for highly secret objects, as they have one of the lowest percentages of denied access to registered users and there are practically no erroneous access permissions.
Unfortunately, a number of difficulties arise when using this biometric method. The scanner here is a very complex optical system, and a person must not move for a considerable time while the system is induced, which causes discomfort.
According to EyeDentify for the ICAM2001 scanner with FAR=0.001%, the FRR value is 0.4%.

Advantages and disadvantages of the method

Advantages. High level of statistical reliability. Due to the low prevalence of systems, there is little chance of developing a way to "cheat" them.
Flaws. Difficult to use system with high processing time. The high cost of the system. The lack of a wide market offer and, as a result, the insufficient intensity of the development of the method.

Hand geometry

This method, quite common 10 years ago, and originating from forensic science, has been declining in recent years. It is based on obtaining the geometric characteristics of the hands: the length of the fingers, the width of the palm, etc. This method, like the retina of the eye, is dying, and since it has much lower characteristics, we will not even enter a more complete description of it.
It is sometimes believed that geometric recognition methods are used in vein recognition systems. But in the sale, we have never seen such a clearly stated. And besides, often when recognizing by veins, only the palm of the hand is taken, while when recognizing by geometry, a picture is taken of the fingers.

A little self-promotion

At one time, we developed a good eye recognition algorithm. But at that time, such a high-tech thing was not needed in this country, and I did not want to go to the bourgeoisie (where we were invited after the very first article). But suddenly, after a year and a half, there were still investors who wanted to build a “biometric portal” for themselves - a system that would eat 2 eyes and use the color component of the iris (for which the investor had a world patent). In fact, this is what we are doing now. But this is not an article about self-promotion, this is a brief lyrical digression. If anyone is interested, there is some information, and sometime in the future, when we enter the market (or do not), I will write a few words here about the ups and downs of the biometric project in Russia.

conclusions

Even in the class of static biometric systems, there is a large selection of systems. Which one to choose? It all depends on the security requirements. The most statistically reliable and tamper-resistant access systems are iris and arm vein access systems. For the first of them, there is a wider market for proposals. But this is not the limit. Biometric identification systems can be combined to achieve astronomical accuracy. The cheapest and easiest to use, but with good statistics, are finger-tolerance systems. 2D face tolerance is convenient and cheap, but has a limited scope due to poor statistics.
Consider the characteristics that each of the systems will have: resistance to forgery, resistance to the environment, ease of use, cost, speed, stability of the biometric feature over time. Let's place marks from 1 to 10 in each column. The closer the score is to 10, the better the system is in this regard. The principles for choosing grades were described at the very beginning of the article.


We also consider the ratio of FAR and FRR for these systems. This ratio determines the efficiency of the system and the breadth of its use.


It is worth remembering that for the iris, you can increase the accuracy of the system almost quadratically, without loss of time, if you complicate the system by making it for two eyes. For the fingerprint method - by combining several fingers, and recognition by veins, by combining two hands, but such an improvement is possible only with an increase in the time spent working with a person.
Summarizing the results for the methods, we can say that for medium and large objects, as well as for objects with a maximum security requirement, the iris should be used as a biometric access and, possibly, recognition by hand veins. For facilities with up to several hundred employees, fingerprint access will be optimal. 2D facial recognition systems are very specific. They may be required in cases where recognition requires the absence of physical contact, but it is not possible to place the control system on the iris. For example, if it is necessary to identify a person without his participation, with a hidden camera, or an outdoor detection camera, but this is possible only with a small number of subjects in the database and a small flow of people filmed by the camera.

Young technicians take note

Some manufacturers, such as Neurotechnology, have demo versions of the biometric methods they release on their website, so you can plug them in and play around. For those who decide to delve into the problem more seriously, I can advise the only book that I have seen in Russian - "A Guide to Biometrics" by R.M. Ball, J.H. Connell, S. Pancanti. There are many algorithms and their mathematical models. Not everything is complete and not everything corresponds to the present, but the base is not bad and comprehensive.

P.S.

In this opus, I did not go into the problem of authentication, but only touched on identification. In principle, from the characteristics of FAR / FRR and the possibility of forgery, all conclusions on the issue of authentication suggest themselves.

In recent years, biometrics is increasingly penetrating into our lives. The leading countries of the world have already put into circulation or are planning to introduce in the near future electronic passports containing information about the biometric characteristics of their owner; many office centers have implemented biometric sensors in corporate access control systems; laptops have long been equipped with biometric user authentication; security services are armed with modern means of identifying any wanted criminal in a crowd of people

Andrey Khrulev
Head of biometric department
and integrated security systems
Technoserv Group of Companies, Ph.D.

There are more and more examples of the use of biometric systems. The success of biometrics is easy to explain. Traditional means of personal identification based on the principles of "I am what I have" (identification cards, tokens, certifying documents) and "I am what I know" (passwords, pin codes) are not perfect. The card is easy to lose, the password can be forgotten, besides, any attacker can use them, and no system will be able to distinguish you from a figurehead.

In addition, traditional means of identification are absolutely useless when it comes to tasks of hidden identification of a person, and there are more and more such tasks:

• recognize the criminal in the crowd;
• check whether the passport is really presented by its owner;
• find out if a person is wanted;
• find out if the person was previously involved in financial fraud with loans;
• identify potentially dangerous fans at the entrance to the stadium, etc.

All these tasks can be solved only with the use of biometric identification tools based on the principle "I am what I am." This principle allows the information system to identify the person directly, and not the objects that he presents, or the information that he reports.

The uniqueness of facial biometrics

Among the variety of biometric characteristics of a person used for personal identification, it is worth noting the image of the face. Facial biometrics is unique in that it does not require the creation of specialized sensors to obtain an image - a face image can be obtained from a conventional camera of a video surveillance system. Moreover, a photograph of a face is present on almost any identity document, which means that the introduction of this technology in practice is not associated with a variety of regulatory problems and the difficulties of social perception of the technology.

It is also worth noting that a face image can be obtained implicitly for the person himself, which means that facial biometrics is optimally suited for building monitoring systems and covert identification.

Any face recognition system is a typical image recognition system, the task of which is to form a certain set of features, the so-called biometric template, according to the mathematical model embedded in the system. It is this model that constitutes the key know-how of any biometric system, and the effectiveness of face recognition directly depends on such factors as the resistance of the biometric template to various kinds of interference, distortions in the original photo or video image.

The effectiveness of face recognition directly depends on such factors as the resistance of the biometric template to various kinds of interference, distortions in the original photo or video image

Despite the huge variety of face recognition systems presented both on the Russian market and in the world, many of them use the same biometric engines - the actual software implementations of methods for constructing and comparing mathematical face models. In Russia, such biometric engines as Cognitec (developed by Cognitec Systems GmbH, Germany), Kaskad-Potok (developed by Technoserv, Russia), FRS SDK (developed by Asia Software, Kazakhstan), FaceIt (developed by L1 Identity Solutions, USA).

As a rule, face recognition in any biometric engine is performed in several stages: face detection, quality assessment, template building, matching and decision making.

Stage 1: face detection

At this stage, the system automatically selects (detects) people's faces in a stream of video frames or in a photograph, and the range of angles and scales of faces can vary significantly, which is extremely important for building security systems. It is not necessary that all selected faces be recognized (as a rule, this is impossible), but it is extremely useful to detect the maximum number of faces in the stream and, if necessary, place them in the archive (Fig. 1).

Face detection is one of the key stages of recognition, since the detection of a face by the detector automatically means that further identification is impossible. The quality of the detector operation is usually characterized by the probability of face detection P0. For modern biometric systems operating in the conditions of a flow of people, the value of the probability of face detection is from 95 to 99% and depends on the video recording conditions (lighting, camera resolution, etc.).

One of the most promising trends in the development of the biometrics market is the emergence of intelligent digital video cameras that implement the face detection function based on built-in logic (Fig. 2). Intelligent video cameras allow you to receive not only a high-quality video stream, but also associated metadata containing information about the found faces.

This approach can significantly reduce the load on the hardware capacity of the recognition system, which, in turn, reduces the final cost of biometric systems, making them more accessible to the end user. In addition, the requirements for data transmission channels are reduced, since with this approach we do not need gigabit communication lines to transmit high-quality video, but the presence of standard networks is sufficient to transmit compressed video and a small stream of detected face images.

Stage 2: quality assessment

This is a very important stage of recognition, at which the biometric engine selects from the entire array of detected faces only those images that meet the specified quality criteria.

Often developers of biometric systems are cunning, claiming that their system provides a high level of recognition if the face images in the video stream meet the quality requirements defined in GOST R ISO/IEC 19794-5. However, this GOST imposes very strict (almost ideal) conditions on the quality of face photographs (frontal view of the face with a deviation of no more than 5 degrees; uniform illumination; neutral facial expressions, etc.), which cannot be performed in real conditions of systems video surveillance. Such requirements of GOST are fully justified by the fact that, in fact, this standard is intended to unify the format for storing electronic photographs in passport and visa documents of a new generation - the so-called biometric passports. In practice, biometric identification systems have to deal with much less favorable operating conditions:

• deviation of the face from the frontal position at angles exceeding 20 degrees;
• strong illumination;
• covering part of the face;
• the presence of shadows on the face;
• small image size, etc.

It is the stability of the biometric engine in such difficult conditions that determines its quality. In modern biometric engines, at the stage of quality assessment, as a rule, the following are evaluated:

• face angle (should not exceed 20–30 degrees);
• face size (estimated by the distance between the pupils of the eyes and should be more than 50–80 px);
• partial face closure (face closure should not be more than 10-25% of the total face area).

There is a common misconception that if the eyes are closed in the image of the face (by blinking or glasses), then the system will allegedly not be able to recognize the person. Indeed, early face recognition algorithms used the centers of the pupils of the eyes as a base for further image processing, in particular for standard face scaling. However, at the moment, many modern biometric engines (for example, Cognitec or Kaskad-Potok) use more complex face coding schemes and are not tied to the position of the centers of the pupils.

Stage 3: building a template

This is one of the most complex and unique stages of face recognition and constitutes a key know-how of biometric engine technology. The essence of this stage is a non-trivial mathematical transformation of a face image into a set of features combined into a biometric template.

Each face has its own unique biometric template. The principles of constructing biometric templates are extremely diverse: a template can be based on the textural properties of the face, on geometric features, on characteristic points, on a combination of various heterogeneous features.

The most important characteristic of a biometric template is its size. The larger the template size, the more informative features it includes, but the lower the speed and efficiency of searching for this template. A typical size value for a face template in biometric systems is between 1 and 20 kB.

Stage 4: comparison and decision

This is a combined stage of the recognition system, which compares the biometric face template built on the basis of the detected face with an array of templates stored in the database. In the simplest case, matching is performed by simply enumerating all templates and evaluating their similarity measure. Based on the estimates obtained and their comparison with the given thresholds, a decision is made on the presence or absence of an identical person in the database.

In modern systems, matching is implemented according to complex optimal matching schemes that provide matching speeds from 10,000 to 200,000 comparisons per second or more. Moreover, it should be understood that the matching process can be parallelized, which allows identification systems to work almost in real time even for large arrays of images, for example, 100,000 people.

The quality of work of face recognition systems is usually characterized by identification probabilities. Obviously, two types of errors may occur during biometric identification.

1. The first error is related to the possibility of missing and not recognizing the person actually in the database - this is often called a type one error. And often they do not indicate the value of the error of the first kind, but one minus the probability of an error of the first kind. This value is called the probability of correct recognition PPR.
2. The second error reflects cases when the system recognizes a person who is not actually in the database or confuses him with another person - it is commonly called a type 2 error. For modern face recognition systems, the typical value of the probability of correct recognition, as a rule, is in the range from 80 to 97%, with an error of the second kind not exceeding 1%.

Conditions for successful identification

It should be understood that face recognition is not an absolute technology. You can often hear criticism of biometric systems that it is not possible to achieve the same high performance on real objects as in "laboratory" conditions. This statement is only partly true. Indeed, it is possible to effectively recognize a face only under certain conditions, which is why it is extremely important when introducing facial biometrics to understand the conditions under which the system will be operated. However, for most modern recognition systems, these conditions are quite achievable on real objects. Thus, to improve the efficiency of face recognition in identification zones, a directed flow of people (doorways, metal detector frames, turnstiles, etc.) should be organized to provide the possibility of short-term (no more than 1–2 s) fixation of the face of each visitor. At the same time, video recording cameras should be installed in such a way that the angle of deviation of the recorded faces from the frontal position does not exceed 20–30 degrees. (for example, installing cameras at a distance of 8–10 m from the passage zone with a suspension height of 2–3 m).

Compliance with these conditions when introducing recognition systems allows you to effectively solve the problem of identifying a person and searching for people of particular interest, with probabilities as close as possible to the values ​​of indicators of successful identification declared by the developers.

Today, several types of such systems are presented on the market at once and they perform tasks of different levels of complexity: from remote recognition in the crowd to accounting for working hours in the office. Facial recognition solutions are available to customers on different platforms - these are server architecture, mobile and embedded solutions, and cloud services.

Modern systems work on deep learning neural network algorithms, so recognition accuracy is maximum even for low-quality images, they are resistant to head turns and have other advantages.

Example 1: Public Safety

Ensuring security is a kind of starting point from which the introduction of biometric identification systems began. Remote facial recognition systems are used to ensure the security of crowded facilities.

The most difficult task is to identify a person in a crowd.

The so-called non-cooperative recognition, when a person does not interact with the system, does not look into the camera lens, turns away or tries to hide his face. For example, at transport hubs, metro, major international events.

Cases

One of the most significant projects of 2017 for our company was the largest international exhibition EXPO-2017, which took place in Kazakhstan this summer. Specialized cameras were used in the remote biometric face recognition system.

The selection of faces in the frame occurs in the camera itself and only the image of the face is transmitted to the server, this unloads the channel and significantly reduces the cost of the network infrastructure. The cameras monitored four entrance groups, in different parts of the complex. The architecture of the system was designed in such a way that the input groups worked separately or all together, while the correct operation of the system was provided by only 4 servers and 48 cameras.

With the help of online video analytics, suspects and missing people are searched for at large geographically distributed facilities, accidents and incidents are investigated, and passenger traffic is analyzed.

At some airports, by the end of 2017, biometrics will also be used to check in passengers for a flight. According to the Tadviser portal, 12 European countries (Spain, France, the Netherlands, Germany, Finland, Sweden, Estonia, Hungary, Greece, Italy, Romania) also plan to introduce smart gate systems at airports.

And the next step should be the introduction of face recognition systems for border and migration control. With state support, the introduction of facial identification can become as commonplace as metal detector frames in the next three to five years.

Example 2. Know your customer by sight

Business is also betting on biometric facial identification. First of all, it is retail.

The systems recognize the gender and age of customers, the frequency and time of visiting retail outlets, accumulate statistics for each individual store in the chain.

After that, detailed reports are automatically displayed for the department both for the whole network and for a breakdown by outlets. Based on these reports, it is convenient to draw up a "client profile" and plan effective marketing campaigns.

Unfortunately, we cannot disclose customers. Among them are the largest retailers and DIY (Do It Youself) networks, which include expensive tools and components.

How it works

Many fear leaks of confidential information, but we specifically emphasize that no personal data of recognized people is stored in archives. Moreover, not even the image is stored, but its biometric template, according to which the image cannot be restored.

With repeated visits, the biometric face template is “pulled up”, so the system knows exactly who and how many times was in the store. For the safety of personal data, you can be calm.

For small shops, car dealerships, pharmacies, the mechanism for collecting marketing analytics is implemented in a cloud recognition service. For small and medium-sized businesses, this option is more preferable, since it does not require the cost of server hardware, hiring additional staff, updating software, and so on. Firstly, this is a convenient tool for evaluating the efficiency of outlets, and secondly, it is an excellent assistant to detect thieves. That is, one system performs several functions at once.

Example 3. Access control and management systems

In addition to the above functions, it is convenient to use the face recognition system as an alternative to Proximity cards in access control and management systems (ACS).

They have a number of advantages: provide high reliability of recognition, they cannot be deceived, copied or stolen identifier, they are easy to integrate with existing security equipment. You can even use existing surveillance cameras. Biometric face identification systems work remotely and very quickly with the recording of events in the archive.

On the basis of a biometric ACS, it is convenient to keep track of employees' working hours, especially in large office centers.

case

We implemented such a system at a large Indian enterprise that specializes in the field of logistics last year. The number of permanent employees is more than 600 people. At the same time, the company works around the clock and practices a “floating” work schedule. With the help of our remote biometric identification system, the customer received a complete and reliable record of employees' working time, a preventive facility security tool and an access control system.

Example 4. Fan pass to the stadium

At the time of buying a ticket at the box office, the face of each buyer is automatically photographed and uploaded to the system. This is how the base of match visitors is formed. If the purchase was via the Internet or a mobile application, then authorization is possible remotely using a “selfie”. In the future, when a person comes to the stadium, the system will recognize him without any passports.

The identification of visitors to sports competitions has become mandatory in accordance with Federal Law No. 284-FZ “On Amending Article 20 of the Federal Law “On Physical Culture and Sports in the Russian Federation” and Article 32.14 of the Code of Administrative Offenses of the Russian Federation.

It is the one who bought the ticket who will enter the stadium, it is impossible to transfer the ticket to another person or go through with a fake ticket. Remote face recognition at stadiums works on the same principle as at large geographically distributed transport facilities: if a person is included in the list of persons who are denied access to the stadium, the system will not let him through.

case

In March 2016, as part of a joint project between Vocord and the Khanty-Mansiysk branch of PJSC Rostelecom, a remote facial recognition system was used to ensure the security of the Biathlon World Cup held in Khanty-Mansiysk. Since 2015, the same system has been successfully operating in the Arena Omsk multifunctional sports complex. It is one of the six largest sports facilities in Russia, is the largest sports and entertainment facility in Siberia and the base of the Avangard hockey club.

Example 5: Internet banking and ATMs

Another niche in which facial recognition has settled is the banking sector. Here, the introduction of new technologies is intensive, since the financial sector is more interested in the reliability and safety of personalized information than others.

Today, biometrics is gradually beginning, if not to displace the usual and well-established "paper" documents, then to go on a par with them. At the same time, the degree of protection when making payments is significantly increased: to confirm the transaction, it is enough to look into the camera of your smartphone. At the same time, the biometric data itself is not transmitted anywhere, therefore, it is impossible to intercept it.

The introduction of biometric identification technologies is directly related to the massive use of electronic services and devices, the development of online commerce and the spread of plastic cards instead of cash.

With the advent of high-performance graphics processing units (GPUs) and ultra-compact hardware platforms based on them - such as NVIDIA Jetson - facial recognition began to be introduced into ATMs. Now only the cardholder can withdraw cash or conduct account transactions, for example, through Tinkoff Bank ATMs. And the PIN may soon be retired.

Modern integrated security systems are able to solve problems of any complexity at various industrial, social and domestic facilities. Video surveillance systems are very important tools of security complexes, and the requirements for the functionality of the segment are steadily growing.

Integrated security systems

A single platform includes modules for security and fire equipment, access control and management, video surveillance or security television (SOT). Until recently, the functions of the latter were limited to video monitoring and registration of the situation at the facility and the adjacent territory, archiving and storage of data. Classical video systems have a number of significant disadvantages:

• Human factor. Inefficient work of the operator when broadcasting a large amount of information.
• The impossibility of surgical intervention, untimely analysis.
• Significant time spent to search for and identify an event.

The development of digital technologies has led to the creation of "smart" automated systems.

Strength in the intellect

The basic principle of intellectual is video analytics - a technology based on methods and algorithms for pattern recognition and automated data collection as a result of video stream analysis. Such equipment, without human intervention, is able to detect and track in real time given targets (a car, a group of people), potentially dangerous situations (smoke, fire, unauthorized intervention in the operation of video cameras), programmed events and timely issue an alarm signal. By filtering video data that is not of interest, the load on communication channels and the archive base is significantly reduced.

The most popular video analytics tool is a face recognition system. Depending on the functions performed and the tasks set, certain requirements are imposed on the equipment.

Firmware and hardware

For efficient operation of the system, several types of IP cameras with different performance characteristics are used. The detection of an object in the controlled area is recorded by panoramic cameras with a resolution of 1 megapixel or more and a focal length of 1 mm, and scanning devices point at it. These are more advanced cameras (from 2 megapixels, from 2 mm), producing recognition using simple methods (3-4 parameters). To identify an object, cameras with good image quality are used, sufficient for applying complex algorithms (from 5 megapixels, 8-12 mm).

The most popular software products for face recognition "Face Intellect" (developer - House Control company), Face director (Synesis company) and VOCORD FaceControl (VOCORD) demonstrate:

• High probability of object identification (up to 99%).
• Support for a wide range of camera rotation angles.
• The ability to highlight faces even in a dense pedestrian mass.
• Variability in the preparation of analytical reports.

Fundamentals of pattern recognition

Any biometric recognition systems are based on identifying the compliance of the read physiological characteristics of a person with a certain predetermined template.

Scanning takes place in real time. The IP camera broadcasts the video stream to the terminal, and the face recognition system determines whether the image matches the photographs stored in the database. There are two main methods. The first is based on static principles: based on the results of processing biometric parameters, an electronic sample is created in the form of a unique number corresponding to a specific person. The second method models the "human" approach and is characterized by self-learning and robustness. Identification of a person by a video image takes into account age-related changes and other factors (presence of a headdress, beard or mustache, glasses). This technology allows you to work even with old photographs and, if necessary, with x-rays.

Face search algorithm

The most common face detection technique is using Haar cascades (sets of masks).

The mask is a rectangular window with various combinations of white and black segments.

The mechanism of the program is as follows: the video frame is covered with a set of masks, and based on the results of convolution (counting the pixels that fall into the white and black sectors), the difference is calculated and compared with a certain threshold value.

To improve the performance of the classifier, positive (frames with people's faces) and negative (without them) training samples are created. In the first case, the result of the convolution is above the threshold value, in the second - below. The face detector with an acceptable error determines the sum of the convolutions of all cascades and, if the threshold is exceeded, signals the presence of faces in the frame.

Recognition technologies

After detection and localization at the preliminary stage, the brightness and geometric alignment of the image takes place. Further actions - calculation of signs and identification - can be carried out by various methods.

When scanning a full-face face in a room with excellent illumination, algorithms that work with two-dimensional images show good results. Analyzing unique points and distances between them, the face recognition system determines the fact of identification by the coefficients of difference between the "live" image and the registered template.

Three-dimensional technologies are resistant to changes in the light flux, the permissible deviation from the frontal view is up to 45 degrees. Here, not only points and lines are analyzed, but also the properties of surfaces (curvature, profile), the metric of distances between them. For the operation of such algorithms, the maximum quality of video recording with a frequency of up to 200 frames / s is required. The system is based on stereo video cameras with a matrix of 5 megapixels, high optical resolution and a synchronization error minimized. Additionally, they are connected by a special clock cable for transmitting clock pulses.

The state of the modern systems market

The first, due to their high cost, were developed only for state military facilities and only in the mid-90s became available to commercial organizations. The rapid development of technology has made it possible to increase the accuracy of systems and expand the scope of their application. In the market of our country, the leading positions belong to American and Western European manufacturers of security systems. The sales leader is the equipment of ZN Vision Technologies and Visionics corporations. The most promising among domestic developers are the research and products of Vocord, NTechLab, Soling, VisionLabs LLC and the STC group, which, among other things, are also engaged in adapting foreign complexes to Russian conditions.

Computer face control

The most extensive area of ​​application of contactless identification is the fight against terrorism and crime. The image of the criminal's face is stored in the database. In crowded places (airports, train stations, shopping malls, sports facilities), people are being filmed in real time to identify wanted people.

The next area is access control systems: a sample of a photo image on an electronic pass is compared with a model obtained as a result of processing data from video cameras. The procedure takes place instantly, without requiring any additional actions from those undergoing (unlike retinal scanning or fingerprinting).

Another rapidly growing industry is marketing. An interactive billboard, having scanned a person's face, determines his gender and age, visualizes only those advertisements that will be potentially interesting to the client.

Trends and development prospects

Facial recognition systems are in great demand in the banking sector.

Following the results of last year, after installing 50,000 intelligent video cameras in their offices, the management of Post Bank managed to save millions of rubles by preventing fraud in the lending and payments segments. Experts say that by 2021 the necessary infrastructure network will be created and any operations at ATMs will become possible only after biometric identification of the client's face.

In the next decade, high technology will allow opening a chain of full self-service stores: the buyer walks in front of the windows, selects the product he likes and leaves. The face and image recognition system will determine the identity of the buyer, the purchase and write off the necessary amount from his account.

Work is underway to create systems for recognizing the psycho-emotional state. The analysis of human emotions will be in demand in multimedia fields: animation, cinematography, the industry of creating computer games.

Everyone knows scenes from science fiction films: the hero comes to the door and the door opens, recognizing him. This is one of the clear demonstrations of the convenience and reliability of using biometric technologies for access control. However, in practice it is not so simple. Today, some firms are ready to offer consumers access control using biometric technologies.

Traditional methods of personal identification, which are based on various identification cards, keys or unique data, such as, for example, a password, are not reliable to the extent that is required today. A natural step in improving the reliability of identifiers was the attempt to use biometric technologies for security systems.

The range of problems that can be solved using new technologies is extremely wide:

1. prevent intruders from entering protected areas and premises by forging, stealing documents, cards, passwords;
2. restrict access to information and ensure personal responsibility for its safety;
3. ensure access to responsible facilities only for certified specialists;
4. avoid overhead costs associated with the operation of access control systems (cards, keys);
5. eliminate the inconvenience associated with the loss, damage or elementary forgetting of keys, cards, passwords;
6. organize access and attendance records for employees.

The development of technologies for pattern recognition by various biometric characteristics began to be dealt with quite a long time ago, the beginning was laid in the 60s. Our compatriots have made significant progress in developing the theoretical foundations of these technologies. However, practical results were obtained mainly in the West and only "yesterday". The power of modern computers and improved algorithms have made it possible to create products that, in terms of their characteristics and ratio, have become accessible and interesting to a wide range of users.

The idea of ​​using individual characteristics of a person to identify him is not new. To date, a number of technologies are known that can be used in security systems for personal identification by:

1. fingerprints (both individual and the hand as a whole);
2. facial features (based on optical and infrared images);
3. iris of the eye;
4. voice
5. other characteristics.

All biometric technologies have common approaches to solving the problem of identification, although all methods differ in ease of use and accuracy of results.

Any biometric technology is applied in stages:

1. object scanning;
2. extraction of individual information;
3. template formation;
4. comparing the current template with the database.

The biometric recognition system matches the specific physiological or behavioral characteristics of the user to some predetermined pattern. Typically, a biometric system consists of two modules: a registration module and an identification module.

Registration module“trains” the system to identify a specific person. During the registration phase, a video camera or other sensors scan the person in order to create a digital representation of their appearance. The face scan takes about 20 to 30 seconds, resulting in multiple images. Ideally, these images will have slightly different angles and facial expressions, allowing for more accurate data. A special software module processes this representation and determines the personality traits, then creates a template. There are some parts of the face that hardly change over time, such as the upper contours of the eye sockets, the areas surrounding the cheekbones, and the edges of the mouth. Most of the algorithms developed for biometric technologies take into account possible changes in a person's hairstyle, since they do not use the facial area above the hairline for analysis. Each user's image template is stored in the database of the biometric system.

Identification module receives an image of a person from a video camera and converts it into the same digital format in which the template is stored. The resulting data is compared with a template stored in the database to determine if the images match each other. The degree of similarity required for verification is a certain threshold that can be adjusted for various types of personnel, PC power, time of day, and a number of other factors.

Identification may be in the form of verification, authentication or recognition. Verification confirms the identity of the received data and the template stored in the database. Authentication - confirms the correspondence of the image received from the video camera to one of the templates stored in the database. During recognition, if the obtained characteristics and one of the stored templates are the same, then the system identifies a person with the corresponding template.

When using biometric systems, especially face recognition systems, even with the introduction of correct biometric characteristics, the decision to authenticate is not always correct. This is due to a number of features and, first of all, to the fact that many biometric characteristics can change. There is a certain degree of probability of a system error. Moreover, when using different technologies, the error can vary significantly. For access control systems when using biometric technologies, it is necessary to determine what is more important not to miss “alien” or to miss all “friends”.

An important factor for users of biometric technologies in security systems is ease of use. The person whose characteristics are being scanned should not experience any inconvenience. In this regard, the most interesting method is, of course, face recognition technology. True, in this case, other problems arise, primarily related to the accuracy of the system.

Despite the obvious benefits, there are a number of negative biases against biometrics that often raise questions about whether biometrics will be used to spy on people and violate their privacy. Due to sensational claims and unfounded hype, the perception of biometric technologies differs sharply from the real state of affairs.

And yet, the use of biometric identification methods has gained particular relevance in recent years. This problem became especially acute after the events of September 11 in the United States. The world community has realized the degree of the growing threat of terrorism throughout the world and the complexity of organizing reliable protection by traditional methods. It was these tragic events that served as the starting point for increased attention to modern integrated security systems. It is a well-known opinion that if the control at airports were stricter, then misfortunes could be avoided. And even today, the search for those responsible for a number of other incidents could be significantly facilitated by using modern video surveillance systems in integration with face recognition systems.

There are currently four main face recognition methods:

1. "eigenfaces";
2. analysis of "distinguishing features";
3. analysis based on "neural networks";
4. method of "automatic processing of the face image".

All these methods differ in the complexity of implementation and purpose of application.

"Eigenface" can be translated as "own face". This technology uses two-dimensional grayscale images that represent the distinguishing characteristics of a face image. The "eigenface" method is often used as the basis for other face recognition methods.

By combining characteristics 100 - 120 "eigenface" it is possible to restore a large number of faces. At the time of registration, the "eigenface" of each particular person is represented as a series of coefficients. For an authentication mode in which an image is used for identity verification, the live template is compared with an already registered template in order to determine the difference factor. The degree of difference between patterns determines the fact of identification. The "eigenface" technology is optimal when used in well-lit rooms, when it is possible to scan the face in front.

The "distinguishing" analysis technique is the most widely used identification technology. This technology is similar to the "Eigenface" technique, but is more adapted to changing the appearance or facial expressions of a person (smiling or frowning face). Distinguishing features use dozens of distinctive features of different areas of the face, taking into account their relative location. The individual combination of these parameters determines the characteristics of each particular person. The face of a person is unique, but quite dynamic, because. a person can smile, grow a beard and mustache, put on glasses - all this increases the complexity of the identification procedure. Thus, for example, when smiling, there is some displacement of parts of the face located near the mouth, which in turn will cause a similar movement of adjacent parts. Taking into account such shifts, it is possible to uniquely identify a person with various mimic changes in the face. Since this analysis considers local areas of the face, tolerances can be up to 25° in the horizontal plane, and up to approximately 15° in the vertical plane, and requires sufficiently powerful and expensive equipment, which accordingly reduces the degree of distribution of this method.

In a method based on a neural network, the characteristic features of both faces - registered and verified - are compared for a match. "Neural networks" use an algorithm that matches the unique parameters of the face of the person being checked and the parameters of the template located in the database, while applying the maximum possible number of parameters. As the comparison proceeds, inconsistencies between the person being checked and the template from the database are determined, then a mechanism is launched that, using the appropriate weight coefficients, determines the degree of compliance of the person being checked with the template from the database. This method increases the quality of face identification in difficult conditions.

The method of "automatic face image processing" is the simplest technology, using the distances and the ratio of distances between easily defined points of the face, such as the eyes, the end of the nose, the corners of the mouth. Although this method is not as powerful as "eigenfaces" or "neural network", it can be used quite effectively in low light conditions.

Facial recognition systems on the market

To date, a number of commercial products designed for face recognition have been developed. The algorithms used in these products are different and it is still difficult to assess which technology has the advantage. The leaders at the moment are the following systems: Visionic, Viisage and Miros.

• Visionic's FaceIt application is based on a local feature analysis algorithm developed at Rockefeller University. One commercial company in the UK has integrated FaceIt into a television anti-crime system called Mandrake. This system searches for criminals using video data from 144 cameras connected in a closed network. When an identity is established, the system informs the security officer. In Russia, the representative of Visionic is DanCom.
• Another leader in this field, Viisage, uses an algorithm developed at the Massachusetts Institute of Technology. Businesses and governments in many US states and several other countries use the Viisage system along with identification credentials such as driver's licenses.
• ZN Vision Technologies AG (Germany) offers a number of products on the market that use face recognition technology. These systems are presented on the Russian market by Soling.
• Miros' TrueFace facial recognition system uses neural network technology, and the system itself is used in the Mr.Payroll Corporation's cash dispensing complex and is installed in casinos and other entertainment establishments in many US states.

In the United States, independent experts conducted a comparative test of various face recognition technologies. The test results are presented below.

Rice. 1. Comparative analysis of the effectiveness of face recognition in different systems

In practice, when using face recognition systems as part of standard electronic security systems, it is assumed that the person to be identified is looking directly into the camera. Thus, the system works with a relatively simple two-dimensional image, which significantly simplifies the algorithms and reduces the intensity of calculations. But even in this case, the task of recognition is still not trivial, since the algorithms must take into account the possibility of changing the level of lighting, changing facial expressions, the presence or absence of makeup or glasses.

The reliability of the face recognition system depends very much on several factors:

• Image quality. The probability of error-free operation of the system is markedly reduced if the person we are trying to identify does not look directly into the camera or is shot in poor lighting.
• The relevance of the photograph entered in the database.
• Database size.

Facial recognition technologies work well with standard video cameras that transmit data and are controlled by a personal computer and require a resolution of 320x240 pixels per inch at a video rate of at least 3 to 5 frames per second. For comparison, an acceptable quality for a video conference requires a video stream speed of 15 frames per second. Higher video bit rates at higher resolutions lead to better identification quality. When recognizing faces from a long distance, there is a strong relationship between the quality of the video camera and the identification result.

The volume of databases when using standard personal computers does not exceed 10,000 images.

Conclusion

The methods of face recognition offered today are interesting and close to widespread implementation, however, it is not yet possible, like in the cinema, to trust the opening of the door only to face recognition technology. It is good as an assistant for a security guard or other access control system.

It is this method that is used in many situations when it is required to make sure that the presented document really belongs to the person who presented it. This happens, for example, at an international airport, when the border guard checks the photo on the passport with the face of the passport holder and decides whether it is his passport or not. A computer access system operates according to a similar algorithm. The only difference is that the photo is compared with the template already stored in the database.

Technologies have already appeared that are based on face recognition in infrared light. The new technology is based on the fact that the thermal image created by heat radiation from the blood vessels of the face or, in other words, the thermogram of a person's face, is unique for everyone and, therefore, can be used as a biometric characteristic for access control systems. This thermogram is a more stable identifier than the geometry of the face, since it almost does not depend on changes in the person's appearance.