What is the core problem of face recognition technology?

Author: Xushun Li

This article was issued by James Consulting with the authorization of Xu Shun. If you need to reprint, please indicate the source.

Apple has just released the latest generation of iPhone X products, the most striking of which is the face recognition application based on the hardware of True Depth technology, which has received wide attention in just a few days. Here, I only express my personal opinion, I hope to put aside various technical terms, and use the simplest language to let consumers understand the technical background of face recognition and the changes brought to the mobile phone.

1. The most natural way to identify Apple products have always been known for the best user experience, starting with the first generation of iPhone support touch screen, each generation of products are trying to innovate in the most natural way of interaction. Undoubtedly, face recognition is another revolutionary innovation after the touch screen. The most direct and natural way for humans to know each other is by remembering the various attributes and gestures of the face. For example, I know you, not through fingerprints, irises or exchange passwords. Then a smart machine should know a person in the same way, so that it is closest to the natural attributes of human beings. Therefore, good face recognition technology will bring people the most comfortable experience and it is very easy to be accepted. 2. Reliable sensing system Just as human eyes obtain images of objects in three-dimensional space, the best way to identify a machine is to have eyes like humans. The key technology here is the depth camera (or 3D camera). In order to achieve the ultimate experience, the iPhone X has to open a number of holes in the design, in order to add a really good depth of visual perception to the phone.

Everyone is very familiar with 2D cameras, so how do you evaluate a depth camera? Apple calls it True Depth. I will give you some more specific concepts: First, when you judge a person's face, you see the face of a person in three-dimensional space and the details of the face. Some people wear makeup, wear glasses, partially block the face, Or face to you, or even make faces to you, in most cases you can recognize a person. If the machine is to have such capabilities, it must rely on a precise depth camera and powerful algorithms (details are covered in the next section of this article). Second, if someone holds a photo, a face model, a mask, or a video posing as someone standing in front of you, you must see it at a glance. Then the machine needs to use depth information to reach human wisdom. Third, the mobile phone needs to be woken up dozens or even hundreds of times a day. You need to unlock the mobile phone in indoors, outdoors, and in the dark in various postures. This requires the depth camera to have strong anti-interference ability. Fast response speed, providing special infrared LED for "lighting", so that the face can still be recognized in the dark environment, and it needs to have a long enough life (after all, it is not cheap to carry this high-tech mobile phone) what…).

3. The Importance of Data and Algorithms The process of human cognition is constantly self-learning. Not everyone with normal vision has the same ability to recognize. The ability of children to remember faces is not as good as that of adults. When a foreigner first arrives in China, he may feel that the Chinese are almost the same. After living for a while, it is easy to identify each Chinese, and a specially trained spy. You can remember an unfamiliar face and identify it accurately in a short amount of time. This is the process of learning. The same is true for the machine, which enhances its recognition ability by learning a large number of face samples and good algorithms. Most of the current face recognition solutions are based on 2D photos. Each of our ID card photos, various public photos, photos registered on various websites or systems, etc. can be recognized by computers. Samples, through a large amount of technology accumulation, the current 2D face recognition technology can achieve a false acceptance rate of 0.1 to 0.2%, but under certain conditions may be limited (such as lighting, sunglasses, angles, expressions, etc.). And the most critical point, face recognition based on 2D images, is easily attacked by high-definition photos, so high-level security applications (such as payment) often do not dare to use only 2D information for authentication. In order to overcome the limitations of 2D data and attack vulnerabilities, before the popularity of 3D cameras, the industry has thought of many ways to make up for the shortcomings. The main idea is to first confirm that the user is a real person, not a photo or video, and then use the existing The 2D data and algorithm further identify the face. The most typical two ways:

First, when the user is authenticated, the machine asks for an expression, such as blinking, smiling, or nodding. The current Alipay login uses this method, which avoids photo attacks. If it is not frequently operated, the user can accept it, but if Need to be unlocked frequently in a variety of occasions, then you may be noticed because the expression pack is too rich.

Second, use the depth camera to determine whether it is a living face, and then use 2D algorithm for face recognition. The image says that when you pass a security gate with face recognition, the machine uses your 2D image data to identify it, and you need a staff member to stand next to confirm that you are not cheating with photos (but the staff is not Know you). If the machine itself can use a depth camera to determine if your face is a real face or a photo, then no extra staff is needed. This does avoid attacks and has no expression requirements for the user. In many cases, this is a very good solution. But this machine actually doesn't know you in 3D space. There is still a big gap between true human wisdom and there is no limit to breaking through 2D data in essence.

iPhone X's 3D face recognition technology far exceeds the cognitive limit of 2D recognition, reaching one million acceptance rate. How to do it? The most critical are the data and algorithms. The best 3D face recognition technology, the sample data source should be the real 3D face, that is, using the depth camera to obtain the face model as a sample of machine learning, which is the closest to the human cognitive process. However, because the depth camera technology has only begun to be promoted within a certain range in recent years, in the past few years, except for Apple, there are only a few well-known companies in the industry who are investing in this area. The first well-known consumer category The 3D face recognition application is based on Intel RealSense's Windows Hello, which is used to unlock the Windows operating system of the PC platform. It is only about 4 years old. In this case, the amount of sample data for the public 3D face is very limited. Presumably, Apple has invested a lot of resources to collect data samples from the moment of acquiring the deep camera company PrimeSense (2013).

There are two problems that come with it. Question 1: Who will be after iPhone X? Can we get the same experience on other phones? My understanding is that – data and algorithms are the key, who can get a lot of 3D face data in a short time, and use the best learning method (professional term for machine learning) , haha), who can have the face recognition ability of iPhone X. For companies in this industry, this will be a huge investment. Question 2: Is iPhone X reading my information as a sample every day? unknown. It is clear that as the number of users and usage increases, many new data will be used as a supplement to the sample, constantly improving the machine's perception of the face, and even more understanding of yourself. It’s like the puppy you adopt, getting closer and closer to you. A good face recognition algorithm will learn data in an increasingly enhanced way, giving users better feedback. 4. Security Issues There has been a lot of debate about the security of face recognition. The first is about hacking. iPhone X uses a millionth of the error recognition rate to ensure that your device is not easily unlocked by others, and beyond the fingerprint recognition to reach the payment level, which is also a redefinition of the biometric industry. The standard of measurement. I believe that future technologies based on deep camera technology and excellent algorithms can be quickly followed. Followed by the security specification for face recognition, if face is used as an important security document in all aspects, how your face information will be acquired, how it is stored and how it will be used will become another major problem in this industry. Imagine if you have acne on your face today, and tomorrow, if a cosmetics company sells acne products to you, will you feel voyeuristic? (Of course this is a joke) We expect leading companies in the industry to play a social responsibility and promote the safety and health of the entire ecology. In short, in the field of face recognition, only companies or solution providers with strong hardware, software, data and other technical strengths and sufficient security can provide consumers with a comfortable experience and confidence. The release of iPhone X seems to be a propositional essay on face recognition, not only for mobile-related industries, but also for a wider range of areas, even each of us. Whether you like it or refuse, the prelude to the era of brushing has been opened, are you ready?