Decryption brush face and biometric recognition
Li Ziqing, Research Fellow, Institute of Automation, Chinese Academy of Sciences, Director of the Center for Biometrics and Security Technology Research Face recognition has been particularly hot in the last year or two. There are several reasons. The first is the development brought about by the advancement of technology. The second is that the application has strong demand. The third is the promotion of the big players and the enthusiasm of capital. Today's face recognition technology makes it easy to get a normal application. For example, we want to use face recognition to search for papi sauce, then the results of its return can now basically be done with papi sauce. The result may also include some other girls who are not papi sauces themselves, but it is also innocuous for this kind of face search application. Moreover, people may also be a combination of beauty and talent, right? Some high-end applications, such as brush face payment, there are still some technical and security issues. I believe that Ma Yun himself will authorize his own account to use the brush face to transfer money, then too young too simple (pattern Tusen broken), sometimes naive (sometimes naive)! As early as more than a decade ago, face recognition was a small fire. Bill Gates himself is very optimistic about the application and future of biometrics. In 2001, he disclosed his face recognition technology to the outside media. That is a complete, fully automatic, real-time face recognition system developed by the author. There is a human body behind, with glasses, that is when the author is young. Because our demonstration is very successful. The reporter finally said, he said: "You guys, this week's salary has already fallen." Long before that, Bill Gates released a new version of Windows, and encountered a blue screen of death at the press conference, so before the interview with CNN, we made various plans. Including the careful deployment of this light, and please Bill Gates, his old man can face up, look at the camera, and please do not be too proud, exaggerated expression may cause errors in recognition. Face recognition, such a thing is what we do when we are born. This is our ability to evolve over millions of years. The automatic face recognition algorithm, the first in the world to do this thing is an Englishman. He was commissioned by a government agency to conduct research. The method he used at the time was a semi-automatic method of manually calibrating the key points on the face on the image, then measuring the distance between the eyes and the thickness of the lips as a feature to compare the faces. Correct. In fact, many of my friends, when the author told him that the author is to do face recognition. He will tell the author: "The author knows how to do it, the distance between the eyes, the size of the eyes, the size of the lips." But in fact, the current technology is not like this. After that, there were a lot of major technological breakthroughs. The key point was that in 2001, AdaBoost, a face detection technology, was able to quickly frame faces from photos and pictures. In the last 10 years, the research and application of deep learning has greatly improved the core technology of face recognition and artificial intelligence. The development of image hardware also strongly provides a good image foundation for this face recognition. The process of face recognition is probably like this. First we find this face in the image, then pre-process each face and give it some corrections for lighting, gestures, expressions, and so on. Then on this basis, we use the algorithm to extract a feature from the face part of the image and turn the picture into a two-dimensional code. On this basis, the features are compared and then the identity decision is made. The technical difficulties encountered here are first of all to solve the lighting problem. For example, under all black conditions, we can't even get images. How can face recognition be performed? Under the condition of the left side, an extreme illumination of our current algorithm is not good enough to make an effective and correct comparison. Others include gestures, expressions, and accessories. For example, wearing a sunglasses and blocking the hair. Often girls are more difficult to identify because her hair is blocked. There are also makeup, many people will ask the author, "I went to South Korea to go to the whole capacity, then what should I do when I enter the customs?" In fact, the author believes that if you rely on face recognition, if you are completely different, it should be treated as a different person. Someone asked the author, "I am the twins, they look exactly the same, can you distinguish them?" The author said: "NO!" Because of the appearance, relying on the face, the twins are the same face. There are still some problems, such as the age span, the author from small to large, the author's face changes are very huge, this is also a problem. As mentioned earlier, in order to give a demonstration to Bill Gates, we carefully arranged the lights. After that, the author is thinking about how to solve this lighting problem, this is the first step to solve. A straightforward solution is that we can add a flash to the front like a digital camera. As long as you can take a positive photo, we can identify it with high accuracy. But if you say that every time you do face recognition, you have to use the flash to flash the experience is very bad, it is not feasible. But we have other methods, such as using a near-infrared active light source. You may see the video above the chat camera and the kind of surveillance video. At night, there are some ways that the naked eye can't see but the sensor and camera can see it. Therefore, the author invented such a method of near-infrared face recognition. Such a product looks like this, is it that everyone looks at the soil? This is the second generation of samples. The first generation of the sample was made of earth, packed in a box of Mengniu milk, and all the circuits, including the diodes, including some sensors, were wrapped inside. But it works, it solves the problem. In 2004, such a face recognition rate at that time was greatly improved. Now this product is grown like this. It is much taller than before. If such a product is sold hundreds of thousands of units every year around the world. Big data and deep learning have greatly advanced face recognition and artificial intelligence, including AlphaGO. There are three elements in the development of such a technology. The first one is big data. To learn such a data for this model, it should be able to cover such data that can be seen in most of our scenarios. The second is the structure of the deep network. It has a very deep layer and it is a non-linear transformation that makes such a function handle some very complex problems like face recognition, speech recognition and machine chess. The third element is that the amount of computation for deep learning is very large. We hope that this model can be trained and learned in a limited and waitable time. This requires GPU acceleration. These three elements are indispensable! In many cases, the posture and expression of the face are uncontrolled. So how to solve this problem? We have developed a three-dimensional deformable model for this purpose. It specifically puts the input image on an internal, three-dimensional model and attaches it to it, and then uses the three-dimensional model to turn it around according to the position of the key point. After turning to the front, we will normalize this expression to it, turn it into a neutral expression, and finally get such an output. In this way, the accuracy of face recognition in a large posture and a large expression condition can be improved. The latest development is that we have combined the three-dimensional deformable model and the deep learning process mentioned above to make it more powerful and solve more difficult problems. A variety of biometrics, including faces, fingerprints, irises, eyes, etc., have some problems. The first is to identify the wrong question. Because there is no algorithm, which artificial intelligence technology can guarantee 100%. The second is that this system will be attacked by a variety of prostheses. A typical case about the problem of wrong recognition rate is that Zhao Wei’s driver sold Zhao Wei’s husband’s house. Perhaps this news is known to everyone because the driver of Zhao Wei has deceived the face recognition system. This driver has grown into a model? He can deceive this face recognition system, and the accuracy and similarity is as high as 98.3%, so I said that I searched online. This is Zhao Wei’s husband, and the author has tried to search. The driver's face, but did not find it. But the very witty author found an expression package for an old driver. If you look closely, is it quite similar, so we should like the face recognition system, it is very smart and accurate! A variety of prosthetic attacks, face recognition, including photo printing, video playback, such as Android 4.0, it launched face unlock. Someone immediately said, "The author took this photo with this mobile phone and unlocked it with this phone." There is also a mask for the face. It is relatively simple to deal with photo printing, we can take a human-computer interaction method. For example, the author can give instructions, you give the author an eye, then the author will see if you blink your eyes. You open the mouth to the author, and you shake the head to the author. When this anti-prosthetic attack technology emerged, another form of attack emerged. He printed the photo and put it out of the eyes and mouth. You let the author blink the eye and blink. You want the author to open the mouth and open the mouth. You let the author shake the head and shake his head, so this technique is a spear. There are shields, we are spiraling and technological progress. Such a face that can be bought and sold on the Internet. In addition to this face, the fingerprint is more common. If you search on Baidu, you can find a variety of imitation fingerprints. It can be used for punching! I don't have to go to work. The author also very much hopes that every Chinese person can survive in a good environment, that is, to live with dignity, so we must eradicate this kind of deception. In order to solve the problem of biometric anti-prosthesis attack, the EU organized 12 teams to carry out systematic and cooperative research. 11 of them are members of the EU member states. We are invited as the only non-European team. an item. We proposed in this project to use a multispectral approach. It is multi-spectral, including ultraviolet, near-infrared, and thermal infrared imaging, which is invisible to the naked eye, but it can be distinguished by imaging in various spectral situations. The difference between this real person and the prosthesis is that it requires a special kind of hardware. The author below compares the identification of the machine with the identification of the manual. This is a picture provided by China Merchants Bank. Then we have to show our ID card at the counter and the teller will check it out. China Merchants Bank statistics is that the error rate of manual verification is about 5%, and the error rate of automatic machine identification is between one thousandth and one ten thousandth. So this machine has far exceeded the artificial one. Identification, but this is conditional. This case is such a water passenger in the Luohu Customs in Shenzhen. Because there are only a few thousand faces that people can recognize, especially for people who are not familiar with it, it is difficult for the author to identify him. Like the author's own words, the author is particularly blind, I think the author can only recognize at least a thousand faces, far below this average. The system successfully identified more than two hundred water users during the first three days of activation. Comparing machine recognition and manual recognition under normal conditions, the success rate and accuracy of robot face recognition is much higher than that of manual recognition, but when the machine does face recognition, it only looks at such a part of the face. It does not use some external information, clues, such as what hairstyle, tall and thin. It is not seen, and people can see this. In addition, the machine, just the author said, the author can search for faces in big data. For example, in a case we did, searching for 10 million copies took less than a second, and the author just said that he can only know thousands of people in his life, but the machine is more vulnerable to such forgery attacks. Use a photo or a video, or take a cell phone, but people can easily identify such a prosthetic attack. In addition to the human face, there are various biological features. We are more familiar with fingerprints, irises, like those in the large piece of palm print, palm vein, gait, signature, finger vein. Then we combine different kinds of biological features, so that it can improve its accuracy and security. For example, when the author pats the face, the author also shoots the pupil, the iris. This will not affect the convenience of your use, but also improve your recognition rate; if you use fingerprints, the author will not only affect the convenience of using fingerprints, but also improve its accuracy. Sex and safety. Finally, talk about technology development. Deep learning is already one of the core methods of face recognition, object recognition, speech recognition and artificial intelligence. Everyone has a feeling that deep learning has not really developed much in recent years. The reason why it is applied successfully is because we are doing data collection on many projects. Then I train, then adjust the parameters and adjust the application. I don't think that the improvement of this technology can be attributed to an engineering result. The author believes that the deep learning theory itself has great potential to be dug up. Many engineering application problems can be attributed to an optimization problem. After defining the objective function, we have to find an optimal solution to solve the problem. This involves a global optimization problem. For example, the author knows that the highest peak in the world is Mount Everest, but if the author climbs up to any mountain, perhaps I will climb to the top of the mountain, it is not Mount Everest. So how to jump out, to avoid such a not too good local extremum, to find the best of the global. In fact, this is already a problem in the mathematical and theoretical circles, or is it not well solved. So further to take this artificial intelligence to a higher level, we need a cross-border effort, we need to have a deeper understanding of our brain. So now it is known that there are different areas in the human brain that perform different functions. This part is to identify the face of the face. This part is to identify the cat. This part identifies whether there is Zhang San Li Si in the face. Also in a different cell, this is the so-called sparse expression problem, or the problem of calling grandmother cells in brain science. Deep learning depends on big data. Without big data, your deep learning is still useless, but when we learn a new concept, we don't use big data. For example, when we grew up in China, the author has never seen a tropical fruit like durian, and some other tropical fruits. I only need to see it once, and the author will recognize it in the future. I don't need to take a lot of durian to see it. Right? So this is a big difference between human cognition and the machine intelligence we are doing now. Therefore, we need cross-border cooperation, we need to do brain science, we need to do biology, physics, chemistry, and we need to do artificial intelligence. Let's discuss cooperation together. All in all, the advancement of technology and the application needs have brought us into the era of brushing the face, and further progress requires us to further explore and break through in science and technology. In addition, there are some security and privacy issues in the application of face recognition. In addition to the further improvement of technology, some standards, laws and regulations must be formulated to avoid possible risks. Silicone Plate,Silicone Suction Plate,Silicone Divided Plate,Suction Silicone Plate MeMe Baby Product (GZ) LLC , https://www.kupbo.com