Facial recognition technology is a new era in video analytics, video surveillance and access control systems. Face recognition system using video surveillance systems. Face search algorithm

Taking out a loan, applying for a visa, or simply launching the latest model of smartphone - all this is impossible today without the participation of facial recognition algorithms. They help police officers in investigations, musicians on stage, but little by little they turn into an all-seeing eye that monitors all our actions online and offline.

Algorithms (technologies)

Identifying a person from a photo from a computer’s point of view means two very different tasks: firstly, to find the face in the photo (if it is there), and secondly, to isolate from the image those features that distinguish this person from other people in the database.

1. Find

Attempts to teach a computer to find a face in photographs have been made since the early 1970s. Many approaches were tried, but the most important breakthrough came much later - with the creation in 2001 of Paul Viola and Michael Jones of the cascade boosting method, that is, a chain of weak classifiers. Although there are now more sophisticated algorithms, you can bet that it is the good old Viola-Jones that works in both your cell phone and camera. It's all about remarkable speed and reliability: even back in 2001, the average computer could process 15 images per second using this method. Today, the efficiency of the algorithm satisfies all reasonable requirements. The main thing you need to know about this method is that it is surprisingly simple. You won't even believe how much.

Step 1. We remove the color and turn the image into a brightness matrix.
Step 2. Place one of the square masks on it - they are called Haar features. We go through the entire image with it, changing the position and size.
Step 3. We add up the digital brightness values from those matrix cells that fall under the white part of the mask, and subtract from them those values that fall under the black part. If in at least one of the cases the difference between the white and black areas is above a certain threshold, we take this area of the image for further work. If not, forget about her, there is no face here.
Step 4. Repeat from step 2 with a new mask - but only in the area of the image that passed the first test.

Why does this work? Look at the sign. In almost all photographs, the eye area is always slightly darker than the area immediately below. Look at the sign: the light area in the middle corresponds to the bridge of the nose, located between the dark eyes. At first glance, black and white masks do not look like faces at all, but for all their primitiveness they have a high generalizing power.

Why so fast? One important point is not noted in the described algorithm. To subtract the brightness of one part of the image from another, you would need to add the brightness of each pixel, and there can be many of them. Therefore, in fact, before applying the mask, the matrix is converted into an integral representation: the values in the brightness matrix are added in advance in such a way that the integral brightness of the rectangle can be obtained by adding just four numbers.

How to assemble a cascade? Although each stage of masking produces a very large error (the actual accuracy is not much more than 50%), the strength of the algorithm lies in the cascade organization of the process. This allows you to quickly exclude from the analysis areas where there is definitely no face, and spend effort only on those areas that can produce results. This principle of assembling weak classifiers in a sequence is called boosting (you can read more about it in the October issue of PM or). The general principle is this: even large errors, when multiplied by each other, will become small.

2. Simplify

Finding facial features that would allow one to identify its owner means reducing reality to a formula. We are talking about simplification, and a very radical one. For example, there can be a huge number of different combinations of pixels even in a miniature photo of 64 x 64 pixels - (2 8) 64 x 64 = 2 32768 pieces. Moreover, in order to number each of the 7.6 billion people on Earth, only 33 bits would be enough. Moving from one number to another, you need to throw out all the extraneous noise, but preserve the most important individual features. Statisticians familiar with such problems have developed many data simplification tools. For example, the principal component method, which laid the foundation for identifying faces. However, recently convolutional neural networks have left the old methods far behind. Their structure is quite peculiar, but, in essence, this is also a method of simplification: its task is to reduce a specific image to a set of features.

We apply a mask of a fixed size to the image (correctly called a convolution kernel), and multiply the brightness of each pixel in the image by the brightness values in the mask. We find the average value for all pixels in the “window” and write it into one cell of the next level.

We shift the mask by a fixed step, multiply again and again write the average into the feature map.

Having gone through the entire image with one mask, we repeat with the other - we get a new feature map.

We reduce the size of our maps: we take several neighboring pixels (for example, a 2x2 or 3x3 square) and transfer only one maximum value to the next level. We do the same for cards received with all other masks.

For the sake of mathematical hygiene, we replace all negative values with zeros. We repeat from step 2 as many times as we want to get layers in the neural network.

From the last feature map we assemble not a convolutional, but a fully connected neural network: we turn all the cells of the last level into neurons that, with a certain weight, influence the neurons of the next layer. Last step. In networks trained to classify objects (to distinguish cats from dogs in photos, etc.), here is the output layer, that is, a list of probabilities of detecting a particular answer. In the case of faces, instead of a specific answer, we get a short set of the most important features of the face. For example, in Google FaceNet these are 128 abstract numeric parameters.

3. Identify

The very last stage, the actual identification, is the simplest and even trivial step. It boils down to assessing the similarity of the resulting list of features to those already in the database. In mathematical jargon, this means finding in feature space the distance from a given vector to the nearest area of known faces. In the same way, you can solve another problem - finding people similar to each other.

Why does this work? The convolutional neural network is designed to extract the most characteristic features from an image, and to do this automatically and at different levels of abstraction. If the first levels usually respond to simple patterns like shading, gradient, clear boundaries, etc., then with each new level the complexity of the features increases. The masks that the neural network tries on at high levels often really resemble human faces or fragments of them. Additionally, unlike principal component analysis, neural networks combine features in a non-linear (and unexpected) way.

Where do masks come from? Unlike the masks used in the Viola-Jones algorithm, neural networks do without human help and find masks during the learning process. To do this, you need to have a large training set, which would contain pictures of a variety of faces on a variety of backgrounds. As for the resulting set of features that the neural network produces, it is formed using the method of triplets. Triples are sets of images in which the first two are a photograph of the same person and the third is a photograph of another. The neural network learns to find features that bring the first images as close as possible to each other and at the same time exclude the third.

Whose neural network is better? Facial identification has long left the academy and entered big business. And here, as in any business, manufacturers strive to prove that their algorithms are better, although they do not always provide data from open testing. For example, according to the MegaFace competition, the Russian deepVo V3 algorithm from Vocord currently shows the best accuracy with a result of 92%. Google's FaceNet v8 in the same competition shows only 70%, and DeepFace from Facebook, with a declared accuracy of 97%, did not participate in the competition at all. These numbers need to be interpreted with caution, but it is already clear that the best algorithms have almost achieved human-level facial recognition accuracy.

Live makeup (art)

In the winter of 2016, at the 58th Annual Grammy Awards, Lady Gaga performed a tribute to David Bowie, who had recently died. During the performance, living lava spread across her face, leaving a mark recognizable to all Bowie fans on her forehead and cheek - an orange lightning bolt. The effect of moving makeup was created by video projection: the computer tracked the singer’s movements in real time and projected pictures onto her face, taking into account its shape and position. It’s easy to find a video on the Internet in which it is noticeable that the projection is still imperfect and is slightly delayed during sudden movements.

Nobumichi Asai has been developing the Omote video mapping technology for faces since 2014 and has been actively demonstrating it around the world since 2015, collecting a decent list of awards. The company he founded, WOW Inc. became a partner with Intel and received a good incentive for development, and cooperation with Ishikawa Watanabe from the University of Tokyo allowed us to accelerate the projection. However, the main thing happens on the computer, and similar solutions are used by many developers of applications that allow you to put masks on your face, be it an Empire soldier’s helmet or “David Bowie” makeup.

Alexander Khanin, founder and CEO of VisionLabs

“Such a system does not require a powerful computer; masks can be applied even on mobile devices. The system can work directly on a smartphone, without sending data to the cloud or server.”

“This task is called facial point tracking. There are many similar solutions in the public domain, but professional projects are distinguished by speed and photorealism,” Alexander Khanin, head of VisionLabs, told us. “The most difficult thing in this case is to determine the position of the points, taking into account facial expressions and the individual shape of the face or in extreme conditions: with strong turns of the head, insufficient lighting and high exposure.” To teach the system to find points, the neural network is trained - first manually, meticulously marking photo after photo. “The input is a picture, and the output is a marked set of points,” explains Alexander. “Then the detector is launched, the face is determined, a three-dimensional model of it is built, on which a mask is applied. Markers are applied to each frame of the stream in real time.”

This is roughly how Nobumichi Asai's invention works. Previously, the Japanese engineer scans the heads of his models, obtaining accurate three-dimensional prototypes and preparing a video sequence taking into account the shape of the face. The task is also made easier by small reflective markers that are glued to the performer before going on stage. Five infrared cameras monitor their movements, transmitting tracking data to a computer. Then everything happens as VisionLabs told us: the face is detected, a three-dimensional model is built, and Ishikawa Watanabe’s projector comes into play.

The DynaFlash device was introduced by him in 2015: it is a high-speed projector that can track and compensate for the movements of the plane on which the picture is displayed. The screen can be tilted, but the image will not be distorted and will be broadcast at a frequency of up to a thousand 8-bit frames per second: the delay does not exceed an invisible three milliseconds. For Asai, such a projector turned out to be a godsend; live makeup began to work truly in real time. In the video recorded in 2017 for the popular Japanese duo Inori, the lag is no longer visible at all. The dancers' faces turn into either living skulls or crying masks. It looks fresh and attracts attention - but the technology is quickly becoming fashionable. Soon, a butterfly landing on the cheek of a weather forecaster, or performers changing their appearance every time on stage, will probably become the most common thing.

Face hacking (activism)

Mechanics teaches that every action creates a reaction, and the rapid development of surveillance and personal identification systems is no exception. Today, neural networks make it possible to compare a random blurry photograph from the street with pictures uploaded to social network accounts and find out the identity of a passerby in seconds. At the same time, artists, activists and computer vision specialists are creating tools that can give people back the privacy, the personal space that is shrinking at such a dizzying rate.

Identification can be thwarted at different stages of the algorithms. As a rule, the first steps of the recognition process are attacked - detecting figures and faces in the image. Just as military camouflage deceives our vision, hiding an object, violating its geometric proportions and silhouette, so they try to confuse machine vision with colored contrasting spots that distort important parameters for it: the oval of the face, the location of the eyes, mouth, etc. Fortunately, computer vision is not yet as perfect as ours, which leaves great freedom in the choice of colors and shapes of such “camouflage”.

Pink and purple, yellow and blue tones dominate the HyperFace clothing line, the first samples of which were introduced by designer Adam Harvey and startup Hyphen Labs in January 2017. Pixel patterns provide machine vision with an ideal—from its point of view—picture of a human face, which the computer catches as a decoy. A few months later, Moscow programmer Grigory Bakunov and his colleagues even developed a special application that generates makeup options that interfere with the operation of identification systems. And although the authors, after some thought, decided not to make the program publicly available, the same Adam Harvey offers several ready-made options.

A person wearing a mask or with strange makeup on his face may not be noticeable to computer systems, but other people will certainly pay attention to him. However, there are ways to do the opposite. Indeed, from the point of view of the neural network, the image does not contain images in the usual sense for us; for her, a picture is a set of numbers and coefficients. Therefore, completely different objects may look quite similar to her. Knowing these nuances of the AI, you can conduct a more subtle attack and correct the image only slightly - so that the changes will be almost imperceptible to a person, but machine vision will be completely deceived. In November 2017, researchers showed how small changes in the color of a turtle or a baseball caused Google's InceptionV3 system to confidently see a gun or a cup of espresso instead. And Mahmoud Sharif and his colleagues from Carnegie Mellon University designed a spotted pattern for the frame of glasses: it has almost no effect on the perception of the face by others, but computer identification using Face++ confidently confuses it with the face of the person “for whom” the pattern on the frame was designed.

Opened a new era. Facial recognition technology is its main feature. And no one doubts that this unlocking method will be implemented in many other smartphones.

Back in the 1960s, special experiments were carried out in which a computer had to learn to recognize a person’s face. Then this led to nothing, since any emotion led to a failure. Also, the invented system was afraid of changing lighting conditions.

Only at the very end of the 20th century did systems appear that learned to identify people’s faces from photographs, remembering them. At the same time, they no longer fail when a mustache, beard, glasses and other “interference” appear. Most actively, such systems began to be implemented in digital cameras. They also found a place for themselves in the security sector.

Facial recognition systems have long had one significant drawback. They depended heavily on the lighting and angle. However, this problem was not noticeable in security scanners. The face was pressed almost closely to them, then illuminated by lamps. The introduction of stereo photography helped get rid of the above-mentioned disadvantage. Two cameras understand the depth of the scene, and therefore the accuracy of the readings increases several times.

How does facial recognition technology work?

Gradually, a new feature began to appear in smartphones. Here, biometric user identification is implemented so that an unauthorized person cannot unlock the device. Ideally, only the twin can access personal information. There is no need to worry about this. It is unlikely that anyone will seriously hide something from their brother or sister. And no one bothers you to set an additional password to read some particularly secret data.

The operation of the facial recognition system in smartphones can be divided into four stages:

Face scanning. This is done using the front camera or, as in the case of the iPhone X, a special sensor. The scan is 3D, so the photo trick won't work.
Extract unique data. The system focuses on a set of features of the face being scanned. Most often these are the contours of the eye sockets, the shape of the cheekbones and the width of the nose. In advanced systems, scars may also be visible.
Retrieving a template with previously received data from memory.
Search for matches. The final stage in which the system decides whether to unlock the display. The power of modern processors allows you to spend only a split second on “thinking”.

The face recognition function can be implemented even using the front camera - as long as it has two lenses. However, in this case, the operation of this function will be unstable. The fact is that only special sensors will ensure facial scanning even in the dark, while the front camera requires bright lighting. Also, special sensors virtually display a larger number of points on the face, so they are triggered even when a beard, glasses and other obstacles appear. In a word, in some DOOGEE Mix 2 the system will definitely work noticeably worse than in the iPhone X. Another thing is that Apple’s anniversary product costs much more than all other smartphones with facial recognition function.

Is technology the future?

The sensors needed for facial scanning require perfect installation. A shift of hundredths of a millimeter will lead to the fact that the function’s operation will no longer be ideal - therefore, during the production of a smartphone, an increased yield of defects may be observed, and this leads to an increase in its cost. And the sensors themselves are very expensive; it’s not for nothing that only Apple uses them, although it doesn’t have any patents for them.

In a word, for now, Android manufacturers will implement the facial recognition function using the front camera. You can already find it in the Samsung Galaxy S8 and Note 8. But the owners of these devices will tell you that it doesn’t work the best - it’s easier to use a fingerprint scanner. Therefore, nothing can be said about the future of the function yet. We need to wait to see whether Apple will implement the corresponding sensors in more affordable smartphones, and whether they will appear in Android devices.

Conclusion

There is no need to worry about saving your identification data. The template created when scanning a face is located in a separate memory section - this sector cannot be read by a computer or third-party programs. However, this also applies to fingerprints. Which type of identification is more convenient to use is up to you to decide.

Have you ever held a smartphone that can recognize your face? And are you expecting mass implementation of this feature? Share your opinion in the comments, we will be happy about it!

More than three thousand video cameras of the city video surveillance network were connected to the facial recognition system. The video image is automatically analyzed in real time: the system can determine the identity of the person in the video, his gender and age.

The Moscow video surveillance system has been trained to recognize faces. Thanks to an algorithm based on the use of neural networks, video recordings from city cameras are analyzed in real time. The faces on the recordings are scanned so that, if necessary, they can be compared with information in various databases - for example, in photo databases of law enforcement agencies when it comes to finding an offender. In addition, such an analytical system can help law enforcement agencies, when catching a criminal, build a route for his movement around the city. The system itself will select the necessary videos from different surveillance cameras, identifying the suspect in the video. The metropolitan network consists of 160 thousand video cameras and covers 95 percent of the entrances of residential buildings. By the end of the year, citizens will be able to independently install cameras in their homes and connect them to a unified video surveillance system.

“The introduction of video analytics is a powerful driver for increasing the efficiency of both private and city video surveillance systems. City residents now have an additional level of protection,” said Artem Ermolaev, head of the Moscow Department of Information Technologies. — Of course, all these opportunities must be implemented very responsibly. Our priority is to balance privacy and security, and we maintain strict internal controls to ensure that citizens' rights are respected."

Currently, about 16 thousand users are connected to the city surveillance system - these are law enforcement officers, state and municipal organizations. Each has its own access level, which allows you to maintain the confidentiality of information. Law enforcement officers can obtain the necessary data upon request within the framework of current legislation, and employees of government agencies receive access to video cameras only from those territories and routes for which they are responsible. Every call to the tracking system is recorded.

The facial recognition function works online, the identification process takes a few seconds. If the algorithm detects a person whose face is uploaded to the database, it will send an alert to law enforcement agencies.

The Department also noted that the introduction of the facial recognition function has already increased the efficiency of investigating offenses and finding criminals. During pilot tests, it was used to detect and detain more than 50 percent of lawbreakers who were sought using analytical algorithms. Before this, some of them could not be found for many years.

Muscovites will be able to connect their surveillance cameras to the general city network. This option will be implemented before the end of the year. Video from such cameras will be transmitted to a unified data storage and processing center (DSDC), and recordings from them can be used as legally significant evidence in court.

This year, more than 3.5 thousand additional cameras were connected to a single data storage and processing center. Access video cameras, cameras installed on the territory and in buildings of schools and kindergartens, at MCC stations, stadiums, public transport stops and bus stations, as well as in parks are connected to a unified system. In addition, by June 2018, CCTV cameras will appear in 25 underground pedestrian crossings in the capital. Recording devices will be installed in underground passages that are not connected to metro stations and are under the jurisdiction of the Gormost State Budgetary Institution.

The global facial recognition market will grow from $4.05 billion in 2017 year before US$7.76 billion by 2022.

How does the facial recognition system work?

In principle, a facial recognition system can be described as the process of matching faces captured by a camera lens with a database of previously stored and identified reference facial images.
Based on the structural implementation of a face recognition system, three common schemes can be distinguished.

Analysis of the video stream on the server

The most common implementation scheme is that an IP camera transmits a video stream to a server; on the server, specialized software analyzes the video stream and compares facial images obtained from the video stream with a database of reference faces.

The disadvantages of this scheme are: high load on the network, high cost of the server; even the most powerful server can connect a limited number of IP cameras, i.e. the larger the system, the more servers.
The advantage is the ability to use an existing video surveillance system.

Video stream analysis on an IP camera

In this case, image analysis will be performed on the camera itself, and processed metadata will be transferred to the server.

Disadvantages - you need special cameras, the choice of which is currently extremely small, the cost of cameras is higher than regular ones. Also, in systems from different manufacturers, the issue of storing and the size of the database of recognized facial references, as well as issues of interaction between the software on the camera and the software on the server will be resolved differently.
Advantages - connecting an almost unlimited number of cameras to one server

Video stream analysis on an access control device

Unlike the first two schemes where IP cameras are used, in this case the camera is built into an access control device, which, in addition to facial recognition, which naturally occurs on the device, performs access control functions, usually through a turnstile or electric lock installed on the door. The database of reference faces is stored on the device, and as a rule, not in the form of photographic images.

Disadvantages - as a rule, all such devices are produced for indoor use.
Advantages - low cost of systems compared to video surveillance systems used for facial recognition.

In any case, the success of facial recognition projects depends on three important factors:
Recognition algorithm
Databases of recognized faces (standards)
Algorithm performance

Facial recognition technology

Typically, the system consists of a video surveillance camera and software that performs image analysis.Facial recognition software is based on image processing and the calculation of complex mathematical algorithms, which require a more powerful server than is typically required for video surveillance systems.

We will be primarily interested in the quality indicators of the software. Secondly, what server power will be needed to analyze the image and process the image database, and thirdly, we will consider the issue of the applicability of IP cameras for facial recognition purposes.Special attention should be paid to the so-called “stand alone” devices that perform image processing directly on the device itself and not on the server; such devices may also have a database of reference faces stored in memory.

2D face recognition
2D (two-dimensional) face recognition technology is based on flat two-dimensional images. Face recognition algorithms use: anthropometric facial parameters, graphs - face models or elastic 2D face models, as well as images with faces represented by a certain set of physical or mathematical characteristics. We will look at the popularity rating of face recognition algorithms below.

2D image recognition is one of the most popular technologies at the moment. Since the main databases of identified individuals accumulated in the world are precisely two-dimensional. And the main equipment, already installed, is also 2D around the world - according to 2016 data - 350 million CCTV cameras. This is why the main demand is for 2D facial recognition systems.

And demand, as we know, stimulates supply, forcing developers to maximize efforts to improve 2D technology.These efforts sometimes bring unexpectedly interesting results, for example in the form of creating a three-dimensional model of a face based on a 2D image.Researchers from the Universities of Nottingham and Kingston have presented a project to 3D reconstruct faces based on a single image.The neural network, through which many three-dimensional 3D models of people and ordinary portraits were passed through, recreates the three-dimensional faces of people based on just one two-dimensional image of the face.

Advantages
A huge advantage of 2D face recognition is the availability of ready-made databases of facial standards and ready-made infrastructure. The maximum demand will be in this segment, and the demand will stimulate developers to improve technologies.

Flaws
Higher FAR and FRR error rates compared to 3D face recognition.

3D facial recognition
3D recognition (Three-dimensional face recognition - English) is usually performed using reconstructed three-dimensional images. 3D facial recognition technology has higher quality characteristics. Although of course it is not ideal.

There are several different 3D scanning technologies. These can be laser scanners that estimate the distance from the scanner to the elements of the object’s surface, special scanners with structured illumination of the object’s surface and mathematical processing of strip bends, or they can be scanners that process synchronous stereo pairs of facial images using the photogrammetric method.

One of the most researched 3D scanners by consumers and experts is the famous Face ID from Apple. The experience of using Face ID is extremely interesting and indicative, because in fact this is the only device with 3D facial recognition technology released to the mass market, if, of course, you can consider a phone for one hundred dollars a device for the mass market.

Apple's 3D technology is the only one in the world that uses vertically emitting lasers (VCSELs), and is rumored to have spent a total of $1.5 to $2 billion on the development of Face ID. The VCSEL supplier to Apple is two companies Finisar Corp (Apple investments - $390 million) and Lumentum Holdings. And judging by the fact that other 3D technologies do not show such effectiveness as Face ID, face unlocking on Android smartphones will not appear soon.

Naturally, Face ID cannot cope with the task of identifying twins, although no one expected this, but even close relatives fail.
Well, rather a comical moment, but at first Face ID did not distinguish Asians, but the problem was fixed so quickly that Apple did not even have time to file a single lawsuit for racism.

Benefits of 3D
Greater accuracy and fewer errors are still unattainable for 2D face recognition systems.

Disadvantages of 3D
Easy enough fake for professionals
Even Face ID, despite all its coolness, was hacked by the Vietnamese company Bkav immediately after it went on sale. The mask was created using a 3D printer. The cost of creating the mask is only $150. Making a mask is quite difficult for the average person, and your mother is unlikely to be able to repeat it, but for professionals it’s like two fingers on the asphalt.

Do not use 3D facial recognition to protect against unauthorized access to laptops, smartphones, or premises with a special level of secrecy; all of them can be easily hacked by professionals.

3D recognition required special cameras for scanning, which several times more expensive conventional CCTV cameras that are used in 2D recognition.
Lack of ready-made databases of identified faces, compared to 2D recognition
Recognizing twins remains a challenging task for facial recognition algorithms. On average, there are 13.1 twins born per 1,000 births worldwide, and this figure varies greatly depending on geographic region.

Face recognition based on facial skin texture
High-resolution images are another factor in the improvement of facial recognition technology; it is thanks to high resolution that very detailed analysis of skin texture has become possible.

In this type of analysis, a specific area of facial skin can be captured as an image and then broken down into smaller units, which are turned into mathematically measurable spaces in which the lines, pores and actual texture of the skin are recorded.

The technology can identify differences between twins, which is not yet possible with facial recognition software.” If face recognition is combined with surface texture analysis, identification accuracy can greatly increase.

Face recognition from thermal imaging
The use of thermal imaging cameras for facial recognition purposes is currently considered a promising area for development, but there are no commercial solutions ready for implementation yet.

The technology is quite promising as it allows us to eliminate the pain points of 2D recognition.

Face recognition in complete darkness and in low light conditions
Makeup, hairstyle, beard, hat, glasses are not a problem for thermal imaging cameras
Allows you to recognize twins

There are two directions in which development is being carried out:
Identification based on pre-created thermograms of identified individuals. The problems here are the same as with 3D recognition, there are no ready-made databases of standards, and the equipment is expensive.
Identification of a person from images obtained from a thermal imaging camera, and a database of ordinary two-dimensional images is used as reference faces. The problem is solved, as you probably already guessed, using deep neural networks.

Face recognition based on skin texture and thermal imaging. It works only in the laboratory, and even then it’s not ideal. But we are watching closely, and if anything happens we will let you know right away.

Software quality

There are several important metrics for assessing software quality.

The most important of them are FRR and FAR
False Reject Rate - FRR (False Reject Rate) - the probability that the system does not identify or verify the authenticity of a registered user.

How is FRR calculated:
Let Nt be the number of image standards in the database. FR - number of false non-recognitions (False Reject - Ivanov, not recognized as Ivanov),

False Acceptance Rate - FAR (False Acceptance Rate) - the likelihood that the facial recognition system will falsely identify an unregistered user or confirm his authenticity.

How is FAR calculated:
Let Nt be the number of image standards in the database. FA - number of false recognitions (False Acceptation - Ivanov is recognized as Petrov),

The first and most important thing you need to know about these two indicators is that they are not absolute, but relative, i.e. they may change depending on the settings of the face recognition algorithm.

The second is that these indicators are interrelated - the lower the FAR, the greater the FRR.

Approximate values of FRR and FAR for face recognition systems and their relationship are presented in the table:

Comparison of FAR and FRR of various biometric identification methods:

Developers of facial recognition algorithms

A recognition algorithm is, as a rule, not a ready-made software product, but a software algorithm that has yet to be packaged into a software product and hardware.

There are quite a few manufacturers of recognition algorithms in the world, fortunately there are independent organizations that test the effectiveness of algorithms. The most famous: NIST - US National Institute of Standards of Technology and MegaFace - University of Washington, Labeled Faces in the Wild, there are others. The results of the competitions are constantly updated. Any company can update its result at any time by taking the test again. Not long ago, NtechLab declared itself the winner, but today they are only in 4th place.

We will publish NIST testing with results as of 5/13/2018. Since NIST, from my point of view, is more interesting since the testing of algorithms takes place on a closed database of individuals, which eliminates the developer’s preparation for testing.

Algorithm - megvii-000 from Megvii, China
Chinese company Megvii with its main product Face++. ByKommersant estimatesThe company's turnover was about $100 million.
Algorithms: 2nd place - visionlabs-003, 7th place - visionlabs-002, VisionLabs, Russia
Algorithms: 3rd place - morpho-002, 17th place - morpho-000. OT-Morpho, France
The first heavyweight in the ranking with a turnover of almost 3 billion euros for 2017. Joint venture Oberthur Technologies (OT) and Safran Identity & Security (Morpho)
Algorithms: 4th place - ntechlab-003, 13th place - ntechlab-002 from NtechLab, Russia
Moscow company, which became famous as a developersolutions for finding porn actors .
Received investments from RT - Business Development (a subsidiary of Rostec) and the VB Partners fund. The investment amount is not disclosed. As a result, the Rostec subsidiary received 12.5% of the company, the New Dimension Fund Variable Capital Investment fund, managed by VB Partners, received 25% of the company. NtechLab plans to enter the national security systems market and accelerate development in the commercial sector.
Algorithm - cogent-000 from Gemalto Cogent, USA
Develops a full range of biometric solutions with a focus on law enforcement, border control and civilian identification. Annual global sales are approximately $205 million.
Algorithm - vocord-002 from Vocord, Russia
The Vocord company was founded in 1999 by MIPT graduates Dmitry Zavarikin and Alexey Kadeishvili. According to SPARK-Interfax, in 2014 the company’s revenue amounted to 302 million rubles; more recent data could not be obtained at the time of writing.
Algorithms: - fdu-000, 9th place - fdu-001. Fudan University, China
Algorithm - neurotechnology-003. Neurotechnology, Lithuania
From the company website you can download demo version of the software for PC and Android smartphone, demo version of SDK. The company has an informative YouTube channel. Prices are published on the website. The company also offers its own cloud service www.skybiometry.com
Algorithm - itmo-003. Saint Petersburg State University of Information Technologies, Mechanics and Optics, Russia
Algorithm - 3divi-001. Tridivi (3DiVi Inc.), Russia
Algorithm - yitu-000. Yitu Technologies, China
Yitu's main product is the Dragonfly Eye facial recognition system, which is used by government security systems in various cities in China. In the first three months of using the system in Shanghai, 567 lawbreakers were detained using Dragonfly Eye. The system stores 1.8 billion photos, and the database includes photos not only of Chinese citizens, but also of all tourists crossing the country’s border. The system is also deployed at public events: during the beer festival in Qingdao, cameras helped detain 22 wanted people. Local authoritiesreport on successes : in one city, the Yitu system helped reduce pickpocketing by 30%, in another, it solved 500 crimes in two years. In some incredible way, the system even helped identify the murder victim by his skull five years after the crime.
Algorithm - gorilla-000, Gorilla Technology, Taiwan
Algorithm - cyberextruder-002, CyberExtruder, USA
Algorithm - tongyitrans-002,TongYi Transportation Technology, China
Algorithm - yisheng-001,Zhuhai Yisheng Electronics Technology, China

Total: 5 representatives from Russia, which is good news, 5 from China, which is not even surprising.

In fact, there are many more manufacturers of recognition algorithms; you can find many missing here in the MegaFace rating. But even if you make a single list, it will still not be complete. Almost all giants of the IT industry are developing their own facial recognition algorithms - Facebook, Google (considers its recognition system the most accurate), Baidu, Microsoft, Yandex (tests driver authorization by face and voice), VKontakte, Toshiba and many others.

There are even .

From all this diversity, several simple conclusions can be drawn:

Competition in this market will intensify, and its consequence has already been a multiple reduction in prices. For example, back in 2017 Macroscop reduced its recognition module prices by 18 times, which they joyfully reported on their website, as if conveying a “big hello” to all their clients who were lucky enough to buy a recognition module before 2017.

It is obvious that prices will continue to decline.The quality indicators of recognition algorithms are constantly growing, and in many cases they differ slightly from each other, the price differs significantly, as you can see below, the performance differs even more significantly, naturally, such a parameter as performance must be tested on a database of the maximum size.

It is also easy to notice that there are practically no manufacturers of equipment for video surveillance systems in the ratings, and without video cameras and storage devices, this whole story with algorithms is just playing on a computer. But the fact that they are not there does not mean that they do not see this market and do not understand its significance. Here is facial recognition from Panasonic, from NEC, Amazon and many others. In general, this market will soon become very hot. Except software solutions(this is when recognition directly occurs on the server), there is also Stand Alone solutions are when recognition occurs on the reading device.

Facial recognition software for CCTV systems

Testing the effectiveness of face recognition algorithms is of course interesting, like any competition, but it is more like an exhibition of national economic achievements. It seems impressive, but how exactly to start using it and how much it will cost is unclear.

The result of the work of algorithms for face recognition will be a match or non-match with the base of standards. And then, depending on the specifics of your system, a pre-programmed action should occur. For example, when a VIP client logs in, the senior manager receives a notification with all the client data from your database.

Or vice versa, when a person from the black list enters, security receives a notification. Or when a person from the blacklist tries to pass through the checkpoint, the access control system blocks the passage - this is already the integration of a facial recognition system with an access control system.

The operation of a face recognition system in real conditions is a whole complex of software and hardware interaction.

To organize such interactions, there are a bunch of integrated platforms that allow you to set up interactions with access control systems, video surveillance systems, security systems, fire safety systems, CRM systems, enterprise management systems, and many others.

So if you’re not into checkers, but lucky, the next couple of sections are simply a “Must Have” for you.

Integration platforms is a big name; it applies to the developers listed below to varying degrees, so when choosing solutions for facial recognition, you need to familiarize yourself with all the capabilities of the software (platform). Taking into account both the current needs of the enterprise and development opportunities, both the qualitative characteristics of the face recognition algorithm and integration capabilities.

Software developers for facial recognition systems and prices for their modules

ISS, Russia, Software “SecurOS® Face”

Face capture module license - price 41,275 rubles

To the channel. Installed on a face recognition server or face capture server

License of the face recognition module (up to 1000 people in the database) - price 665,760 rubles.
To the face recognition server.

Servers for software for facial recognition purposes

Face recognition, like any other video analytics, is a processor-intensive task, so to deploy even a small face recognition system you will need fairly powerful and not at all cheap servers. The characteristics of the server are selected individually and depend on many factors - from the number of recognition channels, to the expected size of the database of reference faces, and the duration of storage of the video archive.

Servers for facial recognition software - price from 101,567 rubles

The choice of servers is not limited to those presented in this catalog; in most cases, we assemble a server depending on your stated requirements.

Best IP Cameras for Facial Recognition

We discussed the software and servers above, but for the system to work, you need IP cameras. It is the quality characteristics of the cameras that will greatly determine how well the facial recognition system will work.

When choosing an IP camera for facial recognition, we recommend paying attention to the following characteristics.

WDR (Wide Dynamic Range)

Despite the fact that recently cameras with WDR have been appearing for 5,000 rubles, the image quality of such cameras is much inferior to cameras from a higher price segment. In our experience, cameras with the best WDR cannot cost less than 80,000 rubles.

Frame rate of at least 60 frames per second

The higher the frame rate per second, the more likely it is that you will get a picture with the best orientation of a person’s face relative to the camera, which will directly affect the quality of face recognition.

Varifocal lens

The more pixels there are on a person’s face, the larger the image will be.

Tests have shown that for successful face recognition it is required that the face be represented by at least 160 pixels per oval of the face, and ideally at least 50 pixels at the distance between the eyes. No matter how carefully you choose the location of the camera, to achieve these values it will have to be adjusted locally depending on many factors. This is why you need a varifocal lens.

CCTV cameras with recommended facial recognition specifications installed- price from 10,000 rubles

Then, as they say, the choice is yours. If you are building a recognition system from scratch, then you should think about choosing the truly best tested IP camera models.

A fairly common and inexpensive functionality, as a rule it is always present in the main facial recognition software, but can also be purchased separately. If you have never been interested in video surveillance systems in your life. Watch the video and it will explain the essence as briefly as possible.

ITV, Russia, Intellect software

Search for faces in the archive (for 1 video channel) - price 6,200 rubles

Trassir, Russia, Software “Trassir Face Search”

Module for searching for a specific person in the Trassir Face Search archive - price 36,990 rubles

Most developers have this functionality in one form or another, so we probably won’t lengthen this already short article.

Manufacturers of equipment with integrated facial recognition algorithms

If the section above with software developers is real, i.e. This is where the main solutions showing maximum efficiency at the moment are concentrated. This section is about the future that is already coming.

In the first case, the video stream from the camera is transmitted over the network to a server with installed software, and it is there that facial recognition occurs. The stream from one IP camera is approximately 5 Mbit/s, and this stream must be transmitted over the network to the server and processed there. In the case of one camera, everything looks acceptable, but if there are hundreds of cameras, this is a problem that needs to be solved separately. It can be solved mainly by dozens of servers for data processing; any video analytics is a processor-intensive task. So servers will be a significant expense.

It is much more efficient to recognize on board the device, and transfer already processed results over the network, which will reduce the load on networks and servers by orders of magnitude.

In addition to the fact that such devices already exist, they already show amazing efficiency and speed. I would divide all the equipment into two large groups: “CCTV cameras with built-in face recognition” and “Equipment for access control systems with built-in face recognition.”

CCTV cameras with built-in facial recognition

Smart cameras with built-in facial recognition algorithms are some of the most advanced in the industry. They allow you to process the video stream directly on the camera itself, and send processed metadata to the server.

2MP camera iDS-2CD8426G0/F-I with two lenses - price 135,550 rubles

HikVision, China, the largest Chinese manufacturer of video surveillance systems.
Sensor - 1/2.8’’ Progressive Scan CMOS
Sensitivity - Color: 0.005 Lux @ (F1.2, AGC ON), 0.0089 Lux @ (F1.6, AGC ON), 0 Lux with IR
Electronic shutter speed - 1s ~ 1/100000s
Resolution 2MPHardware WDR 120dB, frame rate 25fps@2MP, slot for microSD up to 128GB, IR illumination up to 10m

The facial recognition camera, with two lenses, is a compact device with DeepinView deep learning algorithms with a facial recognition system on board.

The camera supports several video compression codecs (H.265, H.264, MPEG-4 and MJPEG) and can process up to five video streams. The size of the video camera is 180.4 x 147 x 117.9 mm, the weight of the device is 1500 grams. Structurally, it is a two-lens camera with binocular stereo technology, which reads a large number of facial characteristics for more accurate recognition.

It is equipped with a lens with a fixed focal length of 4 mm and a viewing angle of 86°. The camera automatically captures, selects and displays the optimal image of a person's face.

Performs face recognition, instant comparison of captured faces with on-board libraries, and supports setting up alarm activation based on an identified face.

HikVision claims operating temperatures ranging from -10°C to 40°C and humidity levels of up to 95 percent.

The camera automatically switches between day and night modes. Infrared illumination operates at a distance of up to 10 meters.

DVR iDS-96128NXI-I16 with facial recognition system - price 3,299,990 rubles

Video recording up to 12MP, Video output up to 4K
128 channels, Synchronous playback 4 channels@4K
16 SATA HDDs up to 10TB each
1/2 audio input/output, 16/8 alarm input/output
Network interface 4 RJ-45 10M/100M/1000M Ethernet

The DVR memory is designed for 16 libraries of pictures of people (up to 100,000 photos in total)

The DVR supports Smart functions for searching for similar people, analyzing behavior, and detecting faces and cars.
It is possible to work with thermal imagers, detect fire, detect sea vessels, measure temperature, maintain statistics of thermal map cameras and count visitors.
The iDS-96128NXI-I16 is capable of detecting people on 32 channels, and modeling faces at a speed of 64 photos per second.

The recorder has interfaces 1 RS-232, 1 RS-485, RS-485 for the keyboard, and two USB 2.0 and USB 3.0 connectors, as well as 16 alarm inputs and 8 outputs.

The IDS-96128NXI-I16 supports the use of RAID0, RAID1, RAID5, RAID6 and RAID10 raid arrays.

CCTV camera DH-IPC-HF8242F-FR with facial recognition system on board - price 100,000 rubles
Dahua Technology, China
1/1.9", 2 MP progressive scan CMOSSmart codec 265+/H.264+, 3-stream encoding
Starlight, true WDR 120dB, 3DNR, Day/Night (ICR), AWB, AGC, BLC
Multiple Network Monitoring: Web Viewer, CMS (DSS/PSS) & DMSS
Auto back focusing (ABF)

Face capture is a software application that automatically captures faces from a digital image or video sequence from a video source. Dahua cameras use advanced Deep Learning algorithms, which allows the camera to quickly and accurately recognize and match faces.

The DH-IPC-HF8242FP-FR face recognition camera uses Deep Learning technology to effectively recognize and match faces. The analytical functions of the device allow you to determine age, gender, mood, presence or absence of a mask/glasses/beard or mustache.

The video camera has a people counting function and generates a heat map.
The camera memory holds up to 10,000 faces, which can be divided into 5 categories, allowing for real-time face capture and comparison.

Thanks to Dahua's Starlight technology, the camera is ideal for working in difficult environments with limited lighting.
Its low light sensitivity ensures color picture performance with minimal ambient light. Even in extreme low light conditions, in almost complete darkness, Starlight technology is able to display a color image.

CCTV camera IPC2255-Gi4N with facial recognition system on board - price 100,000 rubles

Kedacom, China
1/1.9" CMOS Starlight matrix, 0.001 lux with color image 1080@30 fps in H.265 / H.264 / MJPEG
Recognition of up to 18 targets simultaneously (image, face)
Hardware WDR, Adapted IR illumination up to 100m
2 alarm inputs / 1 output, Protection class IP66, Temperature range -40°C +60°C

Axis P1367 video camera with built-in Ayonix algorithm - price 68,448 rubles

Ayonix, Japan
1/2.9” matrix with progressive scan
Variable focal length 2.8–8.5 mm
Superior video quality with 5 MP resolution
Lightfinder and Forensic WDR technologies, Zipstream technology
Advanced Image Analysis Capabilities

Japanese facial recognition software developer Ayonix has developed software to work on board the Axis P1367 camera.

Thanks to the ACAP platform, third-party developers can develop applications to install directly on Axis cameras.

Terminals for time and attendance systems with built-in facial recognition

FacePass Pro time and attendance terminal - price 23,000 rubles

Anviz, China
Memory capacity for 400 users
Identification time< 0,1 сек
Distance for user identification: from 30 cm to 80 cm
Recognition percentage: >99%
Sensitive 2.8” TFT touch display
Built-in Web Server for easy terminal setup

Two scanning cameras ensure the most accurate identification, and the high-speed Samsung ARM processor minimizes the time it takes to recognize employees' faces

Even such factors as skin color, facial expression, gender, hairstyle, as well as the presence or absence of facial hair do not affect the accuracy and speed of identification.

Anviz FacePass Pro is an employee time tracking system with facial recognition, contactless RFID cards or password.

The combination of the new BioNANO algorithm with high-performance hardware guarantees user identification in less than 0.1 (!) seconds.

Unique infrared illumination ensures stable operation of the device both in rooms with varying lighting conditions and in complete darkness.
The speed and quality of identification is not affected even by factors such as skin color, facial expression, gender, hairstyle, as well as the presence or absence of a beard or mustache on the face.

The dynamic numeric keypad and sensitive 2.8” TFT touch display guarantee comfortable operation.

Multibiometric time and attendance terminal ZKTeco Pface202-ID - price 26,500 rubles

ZKTeco, China

Memory for 600 palm vein patterns, 1200 faces (up to 3000 with 1:1 verification), 2000 fingers and 10000 cards
Log capacity 100,000 per event
Touch LCD display 4.3’’
Built-in card reader Em-Marin

Network biometric terminal for time attendance and access control systems with identification by faces, palm veins, fingerprints, RFID card reader and ethernet connection.

Multibiometric time and attendance terminal ZKTeco uFace302-ID - price 27,405 rubles

Memory capacity is 1,200 face templates, 2,000 fingerprints and 10,000 cards
Event log for 100,000 entries
High recognition speed
Interfaces TCP/IP, RS232/485, USB Host, Wiegand output
Outputs for connecting an electric lock, door status sensor, exit button, alarm output
Using hardware encryption to protect firmware

The biometric terminal for time tracking and access control UFace302-ID provides identification by face, fingerprint, card and code. Face302-ID is able to distinguish the face of a real person from a photo image. An advanced and friendly user interface is provided by a 4-inch touch screen (Touch Screen).

Biometric time and attendance terminal ZKTeco uFace800 - price 27,405 rubles

Built-in dual high-resolution face scanning camera with infrared illumination
Memory for up to 3,000 faces, 5,000 fingers, 10,000 cards and 100,000 events
ZMM220_TFT platform, ZK Face 7.0, ZK Finger 10.0 algorithms
6 event statuses upon registration
Fingerprint scanner, Built-in card reader
Interfaces TCP/IP, USB Host, WiFi (optional), Wiegand output

Network biometric terminal for time attendance and access control systems with facial identification, fingerprint scanner, RFID card reader and ethernet connection.
uFace800 supports lock control, door sensor control, bell, burglary sensor, exit button connection.

Equipment for access control systems with built-in facial recognition

Face recognition terminal FaceStation 2 - price 80,856 rubles

Suprema, Korea. The world's largest manufacturer of biometrics, one of the top 50 largest global manufacturers of security systems.
Microprocessor system: 1.4 GHz Quard Core, Memory: 8 GB Flash + 1 GB RAM
Autonomous memory for 30,000 users, 5,000,000 events, 50,000 photos
Fast identification – comparison of 1:3,000 patterns per second
Touch display 4" LCD Touchscreen, Backlight up to 25000 lux
Stable operation in any light thanks to 25,000 lux backlight

FaceStation 2 is a high-performance facial recognition platform. Face recognition can be used in both identification and verification modes. In addition to facial identification, identification by smartphone and contactless cards is supported.

To solve specific ACS problems at a real object, the user can select different identification (1:N) or verification (1:1) modes from the wide list offered by FaceStation 2.
Thus, it becomes possible to choose the optimal balance between the level of security and speed of operation in each specific case.

Depending on the selected mode, various combinations of biometric sensors, a built-in Smart card reader and a touch keyboard for entering a PIN code will be used.
Contactless facial identification and the ability to use a smartphone instead of an access card make FaceStation 2 extremely user-friendly.

High performance facial recognition terminal, with built-in multi-frequency card reader (125kHz EM & 13.56Mhz MIFARE, DESFire/EV1, FeliCa, NFC, ISO14443A/B, ISO15693).

Thanks to the expanded recognition area, the device, installed according to the instructions, “sees” the faces of people with a height of 145 cm to 210 cm.
The terminal is equipped with advanced security features such as infrared anti-fake face technology and multi-band RF reading technology that supports the latest RFID standards.

There is also a terminal modification FaceStation 2 (model FS2-AWB) - price 93,850 rubles, featuring a built-in MultiCLASS SE reader (125kHz EM, HID Prox & 13.56Mhz MIFARE, DESFire/EV1, FeliCa, iCLASS SE/SR, NFC, ISO14443A/B, ISO15693).

And of course, solutions from our Chinese partners, who would not be Chinese partners if they did not offer modern technologies 3-4 times cheaper. Cheapness certainly does not come without consequences, one of the main differences is the size of the database, which for the Chinese is at least 5 times smaller, and speed and accuracy also suffer a little. But in principle, these are completely working, local solutions for small businesses.

Multifactor biometric identification device ZKTeco VF680 - price 13,410 rubles

ZKTeco®, China, the largest Chinese biometric manufacturer, independently develops recognition algorithms.
ZEM810 platform, ZK Face 7.0 algorithm
Built-in dual high-resolution face scanning camera with infrared illumination
Memory for 800 faces, up to 10,000 and 100,000 events
Verification speed is no more than 1 second
TCP/IP connection

The VF680 can operate standalone or connect to networked access control systems. The terminal is equipped with the ZEM810 platform, 3.0-inch touch screen and ZK Face 7.0 algorithm, supporting 800 faces.
The reader can be programmed from the built-in keyboard or using software.

Multifactor biometric identification device ZKTeco MultiBio700id - price 28,530 rubles

Algorithm version: ZK Face v7.0 and ZK Finger v10.0

100,000 events in the log
Recognition within 1 second

The ZKTeco Multibio700 multi-biometric terminal provides access by face, fingerprint, contactless card and code.
The device captures the relative position, size, and shape of the eyes, nose, cheekbones, and jaws and forms a biometric template from this data for subsequent comparison.
User recognition is accurate and fast within 1 second. Infrared illumination helps to successfully carry out identification in low light conditions.

For direct control of the lock, the terminal can be used autonomously, or connected as a reader to network access systems using the Wiegand interface.

Biometric reader ZKTeco SpeedFace V5

Processor Quad-Core A17 1.8Ghz, Memory 2G RAM / 16G ROM
Dual camera: IR camera + Visible Light camera
Operation under lighting 0~40,000Lux
5 inch touch screen
Memory for 6,000 ~ 10,000 (1:N) faces
Verification speed less than 1 second
The reader supports recognition by face, fingerprint and RFID cards EM-Marine or Mifare.

Visible light face recognition is many times superior to IR face recognition and the recognition distance has been increased to 2 meters, which greatly simplifies the situation during rush hour. There is no need to stand in front of the device's camera for a long time. The user can quickly walk in the desired direction next to the device so that the face falls into the field of view of the camera.

Thanks to the use of CNN and the creation of a 3D face model, recognition from different viewing angles became possible. Using an intelligent CNN algorithm, the anti-spoofing function effectively prevents masks, photos and videos from being passed through.

Biometric access terminal HikVision DS-K1T606M - price 49,990 rubles

HikVision, China
Memory for 3,000 face templates, 5,000 Mifare format cards and 100,000 events in the log
2 alarm inputs and 1 output
TCP/IP communication interfaces; Wi-Fi; EHome protocol; RS-485; Wiegand 26/34
Operating conditions -20 °C to +50 °C, humidity 10% - 90%
Suitable for outdoor installation

Reader with a built-in controller and face recognition algorithm ST-FR040EM - price 26,824 rubles
Smartec, Russia, an umbrella Russian brand, places orders with a bunch of Chinese factories and sells in Russia under the single Smartec brand.
Algorithm version: ZK Face v7.0 and ZK Finger v10.0
High resolution infrared camera, 3" touch display
Memory capacity for 400 face templates, 2000 fingerprints and 1000 contactless cards
Recognition within 1 second
Built-in RFID reader Em-Marine 125kHz

If it seemed to you that it is similar to ZKTeco MultiBio700ID, then you are right. For Smartec, this reader is produced by ZKTeco, and it is a complete copy of MultiBio700ID.

Intercom with built-in facial recognition system DS06M - price 14,300 rubles

Bevard, Russia
1.3 MP SONY Exmor sensor, sensitivity 0.01 Lux
Operation in the temperature range from -40 to +50°С, protection class IP54
Installed 4 GB microSDHC card, recording to memory card
Two-way audio
Camdrive cloud service support

Do not rush to rejoice at the low price, the database contains a maximum of 30 individuals. But if you have a small business and want to create a WOW effect on your visitors, this might be the way to go.

By the way, any IP intercom can be equipped with a facial recognition system, preferably with a normal IP camera.
And even more, the IP intercom is installed exactly at face level or slightly lower, which is ideal for high-quality facial recognition.

The same Bevard intercom, but connected to the Makroscope facial recognition system, in this case the Makroscope software is installed on the server. In this configuration, the size of the database is limited only by, excuse me, your financial capabilities.

Glasses with facial recognition system are already uses Zhengzhou City Police

LLVision Technology, China

Connected to a police database, the glasses provide a person's name and address in 2-3 minutes. Over the course of a week and a half, seven wanted people and 26 with fake ID cards were detained using glasses at the Zhengzhou railway station.

In addition to hackers, there is also the ubiquitous Comrade Major, who shows no less interest in biometric data. WikiLeaks published a message (url prudently blocked by RosKomNadzor) about the possible theft of the CIA Aadhaar database, using equipment used for scanning fingerprints and iris from the Cross Match company (Remember this American company, it is still actively promoting its equipment, including including in international markets)

Naturally, Aadhaar itself provides access to commercial companies, for example, Microsoft uses Aadhaar to verify the identity of users of a special version of Skype for India.

Giving commercial companies access to government biometric databases makes a lot of sense. As a rule, the biometric data themselves are not transmitted, only the identification result is transmitted. In addition to the fact that these services are paid, it is with these revenues that state bimetric databases are created and maintained.

Russia also has its own Aadhaar, it is obvious that the Russian database is unlikely to be able to avoid all the problems that Aadhaar went through.

It is planned to provide access to the Russian database to banks.

There are also examples of a different approach; Belgium is the first country to ban the use of facial recognition systems by commercial organizations.

Databases atappropriate for commercial companies
The state, even though it has a monopoly on compulsory data collection, is usually not the most effective data collector.

That's why we have a huge number of commercial databases of biometric information. The largest - VKontakte database (more than 97,000,000 people use monthly VKontakte), for example, NtechLab uses it for its website FindFace, Facebook and other social networks and dating sites also have a large database.

Commercial databases are created so that other commercial companies can use them for modest money.

For example, Bitrix24 in its products Face tracker, Face cards for 1C, Bitrix24.Time and visitor tracker uses the VKontakte database.

Best Database
All of the above databases have one significant drawback: they have nothing to do with your business. And they contain only certain sets of data, often extremely useful sets, but without taking into account the specifics of your business, their use is very limited.

One of the most important indicators of the quality of your database will be the quality of facial images, references.
The most important quality indicators of a reference image database are:
Number of pixels
Contrast and detailing of the face
The background on which the main part of the face is located
No disturbing parts on the face area, etc.

It is also important to observe more or less identical conditions for obtaining images of faces (lighting, size of the face itself against the background of the entire image).

It is important to take these indicators into account at the design stage of the system as a whole, paying special attention to the system for obtaining recognized facial standards.
Start building your database today!

Use of biometric databases
Many have reasonable concerns about both the government's use of biometric data and even greater concerns about the use of this data by commercial entities.

Indeed, these concerns are not without merit, but they should not stop the implementation of the technology. In our near future it is our biological identity will allow distinguish humans from artificial intelligence.

Screen with data about number of bots and real people visiting websites, research conducted by Imperva Incapsula

Even at the moment, the number of bots and real people visiting sites is approximately the same. With the development of the Internet of Things and artificial intelligence, the number of bots will grow in mathematical progression, as will their capabilities; there are already solutions that can call on your behalf, for example, to a hairdresser or pizzeria.

Areas of application of facial recognition systems

The specific application of facial recognition technology differs in the criticality of errors depending on the scope of application.

Access control systems

One of the best applications of facial recognition systems at the moment is in access control systems. Firstly, the employee himself is interested in providing him with access and will not deliberately sabotage the operation of the facial recognition system. Secondly, you control all external factors affecting the quality of recognition - lighting, background, employee movement pattern. Using all this you can create ideal conditions.

Facial recognition systems can be used in access control systems in two modes:

Identification mode - the decision on admission is made based only on data from the facial recognition system. That is, for example, a database of your employees consists of 100 people, and the task of the recognition system is to compare the face of the current person with a database of 100 people. That is, the comparison occurs 100:1. If a person is identified as an employee, they will be granted access.

Face recognition terminals from HikVision

This mode is most effectively used in tasks of detecting strangers in a controlled area. As a rule, it makes sense to use enterprises in specially protected areas where access is allowed to a limited number of persons. All cameras installed in a given area are connected to the recognition system; if any person is detected that is not contained in the database, the security service is informed.

Verification mode - identification in this case is carried out using another technology, for example RFID (if you are a conservative), or mobile identifiers can be used, or fingerprints or vein pattern of a hand or finger , if you decide which way the wind blows in modern ACS trends, and don’t want to throw money away.

A person brings the card to the reader, the system identifies him, that is, it establishes that it is Ivanov, and Ivanov is allowed access at that time. The facial recognition system in this case already knows that this is Ivanov, and using only Ivanov’s photo from the database, compares the RFID card bearer with Ivanov’s photo in the database. That is, the comparison occurs 1:1.

In verification mode it generally works perfectly, since the verification task is very simple even for average quality face recognition systems.

This mode is advisable to use at any entrance points - business centers, manufacturing enterprises, institutes, schools.

The task of the facial recognition system is to verify the card holder. Typically this task is performed by a security guard or watchman. And this is not the best idea, unless you are a folklorist, and do not pursue the goal of compiling an “encyclopedia of modern culture”

The security guard displays a photo of the person on his monitor when he presents the contactless card to the reader; the security guard’s task is to compare the photo and the card bearer (according to science, this is called verification). The security guard does this job poorly - like any other monotonous, routine, repetitive work.

A facial recognition system will not only do this job much more efficiently, but also prevent abuse by security.

Face recognition in transport

In transport, facial recognition can be used for several purposes:

Search for missing people
Search for wanted criminals
Extracting people's demographic information for better service
Measuring people's satisfaction from their faces
Counting the number of passengers using public transport

Online passenger data will help manage urban public transport networks faster, more flexibly and efficiently.

Also, counting the number of passengers when comparing this data with the number of payments will make it possible to establish violations of fare payment rules.

Facial identification for fare payment purposes

Facial recognition for the purpose of paying for travel can currently be carried out in verification mode, and will avoid unauthorized use of multiple-use travel tickets, for example, the use of one travel card by several persons.

The use of a recognized face as the only passenger identifier for automated fare payment, at this level of technology development, seems possible only in small corporate transport networks, and is in no way suitable for mass public urban transport.

Time tracking

Until recently, recording working hours without the use of blocking devices was an unattainable dream. Today this is reality.
Working time recording is, of course, one of the functions of the access control system, but working time recording can be carried out separately, only with the help of facial recognition systems.

One of the main advantages of using facial recognition systems to track time worked is the absence of requirements for facial cleanliness. Within reason, of course - see the “Sabotage” section.

Also, the advantages of working time tracking using a facial recognition system will be:

The absence of blocking devices, which of course increases comfort
Ability to use time tracking secretly, without informing employees

Working time tracking is just one of the metrics, and in general, in isolation from other data on the operation of the enterprise, it does not say much. But having all the knowledge she has, she is perfectly integrated into the analysis of the company’s performance.

It is worth monitoring traffic especially closely during a crisis; PricewaterhouseCoopers itself directly points this out to us. If you do not take attendance, you will receive 2 additional days of absence from work, which you will not know about, but which you will pay for. Which will increase your financial losses from employee absence from the workplace by 1.3 times.

Facial recognition for worker accounting purposes can be implemented in two types.

Server + software + good IP cameras and all this for a lot of money. The royal option is when working hours can be recorded without informing employees.

Specialized terminals are an option when an employee needs to approach the terminal, thereby going through the identification procedure. This only works if you have announced that anyone who does not check in with the facial recognition device will not be paid for the day. This simple administrative measure magically reduces FAR and FRR errors to absolute zero.

Face recognition in a crowd

Speaking about the facial recognition system, as a rule, our imagination draws precisely scenarios for identifying criminals on the streets of the city. This is the most desired, most in demand, and the most difficult task at the moment.

Finding missing people in China using facial recognition

Difficulties
Uneven lighting (day, night, shining sun, all these are different conditions that will greatly affect the percentage of face recognition)
A large number of people in the frame

pros
The effect of surprise
Facial recognition, although a promising technology about which a lot is written, is written in specialized “geek” publications. Therefore, the number of people who are aware is microscopic, on the scale of the total population. Most criminals simply will not perform actions that impede identification.

Network coverage
There are a lot of CCTV cameras in most big cities. It is this aspect that will make its own adjustments to the operation of the facial recognition system. For example, in the UK, a person gets caught on camera about 300 times a day. And this is not a record, and not the limit given the current low cost of IP cameras.

Age determination

We move from security tasks to marketing tasks. When they talk about joining the “Security Systems” industry to the big IT industry, this is exactly what they mean - with the help of equipment that was previously considered capable of solving only security problems. Today I solve a gigantic range of different problems that have nothing to do with “Security Systems” as such.

The age composition of visitors is invaluable information for any marketer, and if you believe our Minister of Health, who stated that average life expectancy could increase to 120 years, although Ms. Skvortsova did not specify in which country this would happen, which obviously reveals that she is an intelligent person. (I personally believe that in Russia), one way or another, the relevance of the task of determining age will definitely increase.

Online services for age determination
You can test how accurate the age determination will be on several online services. Upload your photos and test.

For age determination purposes, you will need:

Facial recognition software www.axis.com/products/axis-demographic-identifier/

Determination of sex

If you are not going to analyze the gender of Eurovision participants, this is a fairly simple task for modern facial recognition systems.

You don’t need to be a great marketer to understand that the different gender composition of your customers requires different marketing, advertising, PR and any other strategies related to customer interaction.

The Cinema Park and Formula Kino cinema chains have already launched a collection of the age and gender of their visitors.

You can test how accurate the gender determination will be using the online services you are already familiar with. Upload your photos and test.

www.skybiometry.com/demo/face-detect/
www.how-old.net

For age determination purposes you can use:
Facial recognition software, in many cases this is one of its functions.

Ready-made solutions from Axis, HikVision - Smart DVR + IP cameras

Sweden
AXIS Demographic Identifier
https://www.axis.com/products/axis-demographic-identifier/

Counting unique visitors

The classic problem of counting the number of visitors has always been solved with infrared or laser sensors, which simply show the number of crossings of a virtual line. For example, a cart will give a separate intersection, using them as an indicator of the average temperature in the hospital.

There are modern video analytics systems, usually combined with additional sensors. They already know how to count specific people, but a security guard, or a local city madman, who has walked back and forth 50 times, can reduce the data to almost complete uselessness.

For the first time, thanks to modern facial recognition systems, marketers can obtain truly actionable data - number of unique visitors. And coupled with gender and age indicators, this is a Yandex metric for your store.

Trassir Face Analytics facial analysis module - price 36,990 rubles
Intelligent face analysis module. Functional:
1. counting unique persons
2. demographic analysis of individuals (gender, age)
3. race identification
4. recognition of facial attributes (glasses, hat, mustache, hair color). Cost for processing 1 video channel.

Soon we will publish an article on our blog with a review of modern video analytics systems for stores, it will be a bomb, subscribe so you don’t miss it - we are available on all platforms -

Until recently, security systems with facial recognition options seemed like something fantastic, and you could only see them in the movies. But a lot has changed over the past few years. New developments have appeared that have changed the idea of security systems.

The quality and comfort of society depends on the correct approach to organizing personal safety and property protection. It's no surprise that security requirements are constantly growing. One of the innovations was the appearance of a facial recognition function. What are its features? Where is it used? On what principle does it work? We will consider these and other questions in detail in the article.

Areas of application

The benefits of facial recognition are difficult to overestimate. Security systems with this function are used in various fields - when organizing a pass system in large organizations, to search for intruders, to protect private facilities, and so on.

Generally speaking, with the help of such a security system it is possible to solve the following problems:

Organize a reliable and effective access system at the checkpoint of the company or other closed facilities. For greater efficiency, video surveillance is combined with turnstiles. As a result, you can quickly recognize your employees and strangers.
Create a theft protection system at points of sale and private facilities. It is no secret that various shops, shopping centers, supermarkets and other establishments are faced with problem customers who are prone to theft. In most cases, theft is carried out by the same people. If there is an appropriate database, the facial recognition function allows you to identify a person in time and inform the security guard. As a result, it is possible to take additional measures to protect property.
Organize a security system that provides protection against unauthorized entry into closed buildings and private households. Even with careful observation, a security guard is not always able to distinguish an intruder from another object. This is especially true if the camera is installed in an area with low lighting levels. Installation of special systems with facial recognition function helps to quickly identify a person even in the dark. What is beyond the control of a security worker can be easily solved by a computer module.
Ensuring face control in night establishments. The presence of the systems in question in clubs guarantees 100% protection from “problem” visitors.

How it works?

Of greatest interest is the operating principle of the system, which is capable of not only transmitting an image to a monitor, but also recognizing people’s faces. The task of a special module is to read information, as well as its subsequent comparison with the data available in the database. Such complexes are capable of identifying a person’s face at a distance of up to 10 m from the camera.

One of the features of the system is high “sensitivity”, which allows you to recognize a person even when your appearance changes. The module cannot be knocked down by using glasses, changing the hairstyle, beard or other additional elements of camouflage on the face. This is due to the fact that it is not facial features that are analyzed, as many believe, but the structure of the skull and its biometric parameters. Such characteristics are individual, just like fingerprints, which eliminates the possibility of error.

Information is scanned and processed in real time. It is enough for the visitor to turn his face towards the scanner, and the system determines the person and gives commands to other organs. If the facial recognition module is connected to turnstiles or other blocking devices, they are activated automatically. In addition, a photograph of a suspicious person is stored in memory for further processing and analysis by security.

Systems with an identification function are most widespread in large companies where there is great competition. It is no secret that the success of an enterprise depends on the level of security. This is especially true for organizations that work in the defense sector, are engaged in the development of new projects or biological research.

The task of the system is to compare employees and check individuals against the existing database. If a person is not on the list, a signal is sent to the security guards, after which the latter take measures to prevent unauthorized persons from entering the facility. In this case, the location of detection is accurately recorded on the electronic map, and security department employees identify the offender within a few minutes.

Installation Features

When installing a system with facial recognition options, it is worth considering that video cameras can operate in one of 2 modes - 2D or 3D. In the first case, the analysis is performed on the basis of a flat image, and two-dimensional cameras are highly sensitive to light. It follows from this that when installing 2D cameras, special attention should be paid to the lighting of the protected object and the coverage of the protected areas.

As for cameras with 3D, they work with a three-dimensional object based on the image transmitted by the device. In this case, you can not pay attention to the level of illumination, because the system copes well with the functions assigned to it even in the dark. The only danger is that the texture of the face will be slightly distorted.

What types of such systems exist?

When choosing systems that have a face recognition function, it is important to focus on several factors - goals, objectives and installation location. In addition, it is worth taking into account the types of such devices:

Detection systems. The video camera has a resolution of 1 megapixel and a focal length of 1 mm. The operation of the device is aimed at recording the fact of penetration of unauthorized entities into protected objects. The peculiarity of the scanner is its ability to distinguish a person from an animal, but it will not be possible to identify the person.
Recognition system. This complex is more complex, and it includes a 2-megapixel camera with a focal length of six millimeters. The task is to recognize faces and identify them according to the “friend or foe” principle. If you watch a video, the picture will not be clear. The system detects unauthorized persons, but in case of theft it will be difficult to find the thief using the saved image
Identification devices. When organizing such a system, cameras with a resolution of 2 MP or more and a focal length of more than eight millimeters are used. Such complexes are capable of performing the functions discussed above. The advantage is that the resulting image is sufficient to identify a thief from a photograph. The available footage can be used during the investigation process and even submitted to court.

The above description discusses the minimum requirements for security systems in terms of focal length and image resolution. This means that when purchasing equipment, you should focus on products with the best characteristics that provide better quality shooting. For example, 2 MP cameras with a focal length of 8 mm are more suitable for recognition systems. As for identification complexes, the recommendations here are even more serious. It is advisable to use video cameras with a resolution of 5 MP and a 12 mm focal length.

Let's summarize briefly:

A video camera with a resolution of 1MP allows you to distinguish a person from an animal. In this case, it will not be possible to identify the subject.
To capture faces and compare them with the existing database, the fixation device must have a resolution of 2 MP or more.
To identify a person, it is advisable to use a 5-megapixel camera.