Worried about new coronavirus pneumonia virus infection, dare not take off your mask? You can now "recognize your face" while wearing a mask.
Recently, many factories, enterprises, and communities across the country have seen such a scenario. Users wear face masks to recognize faces and punch cards. In just a few seconds, they have completed identification and temperature monitoring, which greatly reduces the risk of new coronavirus infection in personnel-intensive places.
In order to ensure the resumption of work and production across the country, representative companies such as Baidu, Alibaba, Tencent, Shangtang Technology, and Yuncong Technology have developed many face recognition products with masks. These products have recently been put into use.
Face recognition technology includes image acquisition, face localization, identity confirmation, and other disciplines.
As early as the 1960s, scientists began research on face recognition technology.
Researchers have found that human facial information, such as the distance between the corners of the eyes and the distance between the nose and the nose, is fixed, and this identity can be used to determine each person's identity.
At present, the face recognition technology is relatively mature, and the recognition accuracy and speed are higher than the naked eye.
Liu Shuo, an engineer of the artificial intelligence department of the China Academy of Information and Communication Technology, said that the previous face recognition was mainly scanning for the entire face.
After the outbreak, the R&D personnel strengthened the identification of key areas such as eyes and eyebrows in consideration of residents' masks.
Can face recognition with masks emerge during the epidemic and still maintain high accuracy?
Deng Wei Hong, professor of pattern recognition laboratory at Beijing University of Posts and Telecommunications, admitted that face recognition under masking conditions such as masks and sunglasses is actually an "old" technology.
Previously, researchers have long studied this technology when solving military criminal investigation and video surveillance problems and developed many mature applications.
Therefore, the stability and accuracy of the technology are fundamental. The previous technical foundation has not completely dispelled people's concerns about the "congenital deficiency" of face recognition when wearing a mask.
Compared with the past, masks cover the face, which reduces the amount of facial information collected by the face recognition system.
Deng Wei Hong said that the key information for face recognition is concentrated on the eyebrows and eyes. As long as the model is trained properly, the accuracy of face recognition with a mask will not decrease significantly.
The epidemic has allowed this niche application to enter the public life. With the popularity of the technology, its application scenarios will be extended to personal consumption, transportation, and education.
Although there are various face recognition products with masks, the algorithms of these products are mostly based on convolutional neural network technology, and various research and development institutions have made some adjustments based on this.
Wisesoft product director Lu Xuebin told China Science Daily that in the subsequent adjustments, many R&D institutions are training based on two-dimensional images.
Two-dimensional images of users are readily available, and some security issues have been exposed in recent years.
In order to improve the recognition accuracy, a method of combining local features with global face features is also favored.
"This method has requirements on the size of the training data, usually requiring hundreds of thousands to millions of samples, and huge investment, often only by developers with strong funds." Deng Wei Hong said that it is undeniable that in image quality Under the premise of guarantee, the larger the training data size, the higher the recognition accuracy rate.
In addition, in order to obtain individual information as much as possible, some face recognition technologies also collect human body information such as dress, body shape, and hairstyle to improve the recognition accuracy.
There are other technologies that take a different approach and use image reconstruction networks to reconstruct face images of objects such as glasses, masks, hats, etc. into face images without accessories, and then realize face recognition through comparison.
Deng Wei Hong said that some implementations may "look beautiful", but the implementation is very difficult, and the recognition stability is difficult to maintain, making the technology difficult to apply.
Wisesoft is one of them.
Lu Xuebin said that he hopes to increase the number of facial information collected in a limited area of the face by collecting three-dimensional portraits of the user, and build fine geometric information about the user's face, so as to achieve face recognition with a mask.
At present, the technology has been used in West China Hospital, Beijing South Railway Station, schools, etc. In response to the epidemic, the research team also developed a mask-based 3D face recognition and automatic rapid body temperature screening system.
"The system can not only identify users who wear masks, but also users who do not wear masks. The recognition accuracy rate is over 96%, which basically meets the needs of the scene." Lu Xuebin said.
The application of face recognition technology is not difficult. Liu Shuo said that at present, most face recognition apps and hardware devices directly purchase mask/face recognition software packages/toolkits provided by the aforementioned R&D companies and can be used after debugging.
"Regular software packages/toolkits are basically sufficient for real-world applications and save development time," Liu Shuo said.
In order to give the application party more autonomy, some R&D companies have also opened software packages/toolkits, and the application party can obtain a mask recognition wearing model in a short time.
Recently, Baidu has open-sourced a face detection and classification model for masks through the Paddle Hub. This model can effectively detect all faces in dense crowd areas and determine whether they are wearing masks.
An AI mask detection application launched by CNPC Ruifei, an information technology company affiliated with CNPC, is based on this open source model.
The application can recognize, and voice alarm the person who does not wear a mask in the work area, and the recognition accuracy rate is more than 96.5%.
An R&D engineer on Baidu's project told China Science Daily that, unlike regular users, CNPC Ruifei uses an internal LAN office, so R&D personnel customize it as needed, incorporating modules such as the propeller main frame, pre-trained model management, and migration learning tool Paddle Hub After being packaged in the form of image, it was deployed to CNPC Ruifei LAN, which solved the problems of video data processing and model test optimization.
It is worth mentioning that the face recognition system with masks is not limited to the epidemic period.
Deng Wei Hong introduced that after the epidemic, the relevant application system can be debugged into a normal face recognition mode, which minimizes the input cost of the application side.
Liu Shuo reminded that just like face recognition, behind mask recognition, there should still be concerns about how to protect privacy and how to balance security with high efficiency.
Compared with face recognition, we currently lack standards related to face recognition for masks, and cannot comprehensively evaluate whether existing technologies/products can meet the needs of different application scenarios.