How long will it take for the phone to recognize your “face mask”?

Editor’s note: This article from the micro-channel public number “The machine can” (ID: almosthuman2017) , of: Xu Dan.

From the outbreak of the domestic outbreak Time has passed, but your phone still cannot recognize you wearing a mask.
                    

The delay of the face recognition technology of the mobile phone mask is related to the privacy and security performance of the mobile phone.
                    

Taking an Apple phone as an example, Apple ’s Face ID is the most advanced 3D facial recognition system, which perceives the user ’s facial features by projecting more than 30,000 invisible points on the user ’s face. It also uses its advanced anti-spoofing neural network and machine learning to prevent the use of photos or masks or any other technology to unlock the iPhone.
                    

However, this mobile phone problem was first widely solved in security scenes-face recognition in masks has gradually become normalized. A mainstream method is the “attention mechanism face recognition algorithm” in the face occlusion environment By identifying key points of the eye as a “attention model”, the global image can be quickly scanned to obtain the eye of the target area that needs to be focused on and form the focus of attention.
                   

After the epidemic, mobile phone users all over the world Pain point-Face ID is too difficult to unlock when wearing a mask. Even the vice president of Huawei’s mobile phone product line also said that the Mate 20 Pro had tried to wear a mask for face recognition, but unfortunately it failed to achieve this function in the end.

During May 1st, Apple quietly launched the latest beta version of IOS 13.5, trying to solve the problem of “mask face” recognition in this version, but unfortunately, this is not the ultimate solution.

Previously, if a user wore a face mask and used Face ID, the Apple lock indicator would shake and the phone would vibrate, indicating a problem. After the face ID times out, the system will prompt the user to swipe up to enter the screen where the password can be entered.

Since Apple does not support fingerprint unlocking, the process of booting with a mask on is very cumbersome and the user experience is poor.

iOS 13.5 beta skipped a step. After the update, the system will automatically recognize the face wearing the mask, without waiting for the recognition failure step, directly jump into the password input step.

At present, Apple has not disclosed when iOS 13.5 will be finalized and released to the public.

With the normalization of epidemic prevention and control, how to recognize a face wearing a mask has become a common problem. Apple has pioneered the “recognition of whether a face is wearing a mask” on a smartphone. There is still a long way to go in “face boot”.

However, in the broader security field, real face recognition has gradually become normalized.

Face recognition mobile phone is difficult to land

Smartphones have been unable to recognize masks. The most important issues are privacy and security.

At present, there are two main schools of 3D face recognition on smartphones. One is Apple-led Structured Light technology, and the other is the TOF (Time Of Flight) commonly used in Android flagships. ).

The former perceives the user’s facial features by projecting invisible points on the user’s face, while the latter perceives by recording the time when the radar light source reaches the face.

Wear masks will cover a lot of facial features. For security and privacy considerations, mobile phones ca n’t recognize human faces when their faces are not completely exposed. After all, mobile phone facial recognition involves payment and other issues.

Apple’s “face mask recognition” can only “recognize a face wearing a mask”, and then automatically jump to the password unlocking stage, instead of directly wearingThe face of the mask is unlocked.

In January, when the domestic requirements for wearing masks began to be widely promoted in China, the vice president of Huawei ’s mobile phone product line, Li Xiaolong, once responded on Weibo that wearing masks could not recognize faces.

Bruce talked about it. Previously, when he released the Huawei Mate 20 Pro, he also tried to wear a mask for face recognition, but unfortunately, he finally failed to achieve this function, because after wearing the mask, almost half of the face was covered Live, and there are too few feature points of the eyes and head exposed, which fail to show people’s facial features and cannot guarantee security, so they finally turned to the facial recognition unlocking technology of the “wear mask / scarf” scene. .

Apple ’s mobile phone is even more so. Apple ’s Face ID is the most advanced 3D facial recognition system. It projects the user ’s facial features by projecting more than 30,000 invisible points on the user ’s face. It also uses its advanced anti-spoofing neural network and machine learning to prevent the use of photos or masks or any other technology to unlock the iPhone.

Face ID only works if the user ’s eyes, nose and mouth are bare and scannable. When the user wears a mask, the system does not work for security reasons, because the system is also used for bank transactions and Apple Pay.

When the user wears a mask, these 30,000 invisible points will be partially covered, which does not allow the Face ID system to measure 3D depth and facial features, and the camera cannot capture and read all biometric information. Therefore, face ID cannot be used with masks.

Overcoming technical difficulties, the security market becomes fertile ground

In addition to mobile phones, a huge application market for face recognition is security. Including video surveillance, access control and so on.

Under the influence of the epidemic, how to identify the face wearing a mask in the security market has become a common problem.

Security Face Recognition is a biometric technology based on human facial feature information for identification, the purpose is to determine the identity of the face image, extract the features of the input face image, the feature is to use one The vector features facial features such as eyes, mouth size, skin color, faceAbstract description.

The features of the input face image are compared with the features of the base library face image one by one to find the base library image with the highest similarity to the input image feature. If the similarity is greater than the preset similarity threshold, the base library The image and the input image are the same person, otherwise the identity of the input image cannot be determined.

The current technical difficulties are as follows:

  • First of all, the facial features often displayed with masks are small. In a scene like a train station, the amount of data is large every day, and the accuracy requirements of the verification personnel are relatively high, which is often difficult to achieve.
     

  • Secondly, there are many types of masks and other obstructions, and the degree of occlusion is different, which again makes it difficult to obtain information on non-occlusion areas;
     

  • In addition, it is difficult to collect a large number of face images wearing masks in a short time, and the algorithm training is difficult;
     

  • Finally, the face recognition system contains multiple modules such as detection, tracking, and recognition, which will affect them all.

    However, in the security-grade market, the difficulty of wearing face recognition technology can be overcome. The normalization of epidemic prevention and control is forcing manufacturers to update their technology to meet the needs of the market and users.

    In this respect, China has made many breakthroughs. As early as in late February, the State Council responded to the “Corporate Prevention and Control Measures for the Resumption of Production and Reproduction of Enterprises and Institutions” issued by the Joint Prevention and Control Mechanism for New Coronavirus Infection Pneumonia. It was proposed that all units should temporarily disable the fingerprint attendance machine and use other methods. Register the entry and exit personnel.

    At present, domestic Internet and technology companies such as Tencent Youtu, Baidu, Alibaba, Shangtang, Hanwang, etc. have developed face recognition technology for masks and have put them into application.

    The technology developed by the Baidu vision team can make the face detection accuracy of wearing masks exceed 99%. In Baidu Park, employees can wear face masks at work to “face” the face.

    Wang University National Multimedia Software Engineering Technology Research Center Professor Wang Zhongyuan developed a mask that masks face recognition accuracy of 90%; at the same time, the world ’s first publicly available face mask sample recognition set is constructed.

    “Attention Mechanism Face Recognition” breaks through the difficulty of masks

    Major facial recognition technology problems, domestic pan-security manufacturers actively conquered during the epidemic. The heart of the machine is in the “AI war epidemic special reportOne of them is Xiaoshi Technology, which has been reported in. The main method of the technology they use is “Attention Mechanism Face Recognition Algorithm” which is a more mainstream idea.

    According to the technical staff of Xiaoshi Technology introduced to the heart of the machine, in the face occlusion environment, by identifying key points of the eye as a “attention model”, quickly scan the global image to obtain the focus of attention Target the eye area and form the focus of attention. Then invest more attention resources in this area of ​​the eye to obtain more detailed information of the target that needs attention, while suppressing other useless information.

    In simple terms, it extracts the key points in the face and recognizes different faces through the key points. When almost half of the face is blocked, the key points of facial features are mainly concentrated in the eyes and eyebrows.

    Break through the algorithm model, use the fusion of local features such as eyes and eyebrows and the overall facial features, and combine the attention mechanism to enhance the eye features. By training the model of the key points of the eye, the model can be improved in the mask Face recognition rate under occlusion.

    Using the same principle, at the algorithm level, for the method based on the combination of global features and local features of the face, the existing face recognition algorithm model can be optimized. Accurate identification.

    However, the attention technology recognition algorithm requires a large amount of face data wearing masks to train the algorithm. This also requires data size. The larger the training data size, the higher the recognition accuracy rate. Usually requires hundreds of thousands to millions of sample size, huge investment.

    In order to obtain the face data of wearing masks as much as possible, the traditional method is to “stick” the mask on the existing face image. By collecting mask pictures of various colors, sizes and styles on the market, and merging with the previously accumulated face pictures of unweared masks, various scenes and massive real mask training pictures are quickly synthesized.

    Synthetic mask pictures

    In addition to the commonly used attention mechanisms, there are some face recognition technologies that improve the accuracy of recognition by collecting human body information such as clothing, posture, and hairstyle. alsoSome reconstruct the face images of wearing objects such as glasses, masks, hats, etc. into face images without wearing accessories through image reconstruction networks, and then realize face recognition through comparison.

    There are also teams that use very technically difficult methods. For example, Chuanda Zhisheng collects three-dimensional portraits of users to increase the amount of facial information collected in a limited area of ​​the face, and constructs detailed information on the user ’s facial stereoscopic geometric structure to achieve the wearing of masks. face recognition.

    In short, China is at the forefront of the world in the application of face recognition with masks. The most important factor is to have a mature Internet environment and massive data.

    Not only that, the industry now generally believes that American artificial intelligence is superior in basic science and core technology, while China focuses on application scenarios, which helps the commercialization of technology.

    Due to the rapid development and popularization of the domestic Internet, huge amounts of data are generated every day, involving all walks of life, providing convenience for mining the value of data, and also becoming a “training ground” for machine learning. Continue to deepen and apply.

    Thus, aside from factors such as privacy and security, China is now the world’s most suitable country for the commercialization of various artificial intelligence application scenarios.