Home > News content

How does Baidu "break the circle" when face recognition hits the wall?

via:雷锋网     time:2020/3/20 15:56:27     readed:507

This spring, face recognition encountered a little trouble. An epidemic has made people wear masks. No one would have thought that the first one to be affected is apple mobile users.Since Apple has been using face ID to unlock mobile phones since iPhone x, face id also uses 3D face recognition technology with high security. Recently, many apple mobile phone users said that "after wearing masks, mobile phones are not easy to use". There are even a lot of teaching videos on the Internet, teaching how to train your iPhone and let your iPhone know you wearing masks.

The training results vary from person to person. Some bloggers also said that they failed to let their "silly son" know who they are wearing masks hundreds of times.

So, when face recognition can no longer recognize your face, will it recognize you?

After returning to work, face recognition hit a wall

In recent years, with the increasing maturity of face recognition technology, related products have been widely used in domestic scenes such as hotel registration, station crowd screening, company attendance, etc. Taking attendance of the company as an example, according to the relevant industry survey data in 2018, the shipment volume of face recognition attendance machine has exceeded one million.

However, in 2020, an epidemic made people wear masks. After the mask became a national "just needed", it began to be out of stock frequently. When you returned to work years later, you found that the face recognition attendance machine you thought was very high-tech at the door of your company did not know you wearing the mask at all.

In fact, wearing a mask is a large area occlusion problem, which is a recognized problem in the field of face recognition. There are three main difficulties:

  • First, the face recognition algorithm mainly determines the identity based on the face features. Wearing masks will cause a large number of facial features such as chin, mouth and nose to be lost;

  • Secondly, the deep learning technology used in current face recognition algorithms relies on massive training data, so it is difficult to collect a large number of photos with masks in a short period of time and label them manually;

  • Third, the face recognition system generally includes face recognition, tracking, living detection, recognition and other modules. Wearing a mask not only affects the work of the face detection module, but also has a great impact on each module of the system.

The popularity of face recognition technology in China is so fast because a large number of Internet companies are promoting it. Bat actually had relevant patents in hand before. During the epidemic, it can also be seen that Tencent Youtu launched "special AI for mask wearing recognition", Shangtang technology published "face recognition algorithm for reading 240 facial feature points around eyes, mouth and nose", and Baidu vision team launched "face recognition algorithm for mask wearing".

When face recognition, it is no longer face recognition

Novel coronavirus pneumonia novel coronavirus pneumonia was reported in January 21st by Zhong Nanshan, who returned from Beijing to Guangzhou. The meeting showed that the new crown pneumonia had been handed down to people in Guangzhou.

Also on this day, baidu vision team began to develop "face recognition algorithm with masks".

  • First of all, aiming at the problem of information loss caused by occlusion, baidu vision team found that,The amount of information contained in each region of the face for identification is not evenly distributed, and the eye region contains more identity information than other locations, so as long as the algorithm pays more attention to the feature learning of the eye region, it can minimize the impact of information loss caused by wearing masks, hats and other occlusions on the recognition rate of the algorithm.

Based on this discovery, the method of spatial location-based attention mechanism feature learning is used to fully obtain the identity information of face wearing mask.

Feature learning of attention mechanism based on spatial location

  • Secondly,In view of the lack of training data of wearing masks, baidu algorithm team solves the problem through manual generation.

Because of the change of face pose and the difference of image distribution in different scenes, the traditional method of superimposing mask images directly on the face can not simulate the real face images in real scenes. The experimental results show that the effect improvement is very limited. Therefore, baidu vision team adopts the 3D image fusion technology based on the key points of face, which not only solves the problem of mask deformation and occlusion caused by the change of face posture, but also produces more natural and real photos. Finally, by collecting all kinds of color, size and style of mask pictures on the market, baidu vision team integrates with the previously accumulated face pictures without masks, which is fast A large number of real training photos with masks were synthesized.

Composite mask picture

  • Finally, in all aspects of face recognition system, baidu algorithm team has made a series of optimization for the problem of wearing masks.

For example, the detection algorithm uses Baidu's latest pyramid Lite detection algorithm, and adds more than 100000 masks to face data training, with the same accuracy rate, the recall rate increases by 30%; the mask classification algorithm extracts facial features with rich facial semantic information based on the face key point network, and judges mask occlusion based on this feature, through paddleslim Row model compression, model size compression more than three times.

Obviously, the face recognition algorithm with mask is no longer a face recognition algorithm, but a more centralized and unobstructed eye with "identity information", which is bound to reduce the accuracy of the face recognition algorithm. In order to ensure the accuracy of the face recognition algorithm with mask, the algorithm team also needs to make great efforts to further optimize the algorithm.

On February 21, baidu vision team completed the research and development and optimization of "face recognition algorithm with mask", which was then applied to Baidu park.

人脸识别碰壁,百度如何“破圈”?

Also applied to Baidu's own park is its recently launched "enterprise AI admission solution" for multi person, mask wearing and remote face detection.

人脸识别碰壁,百度如何“破圈”?

Baidu vision technology, will it be someone else's child

As one of the AI technologies with high commercial value, computer vision is not only closely watched by internet giants, but also closely pressed by many AI Unicorn enterprises.

At present, we can refer to two aspects of strength to measure the ability of computer vision, one is the ability of technology research and development, the other is the ability of technology commercialization.

In terms of technology R & D capability,Baidu opened its paddlepaddle as early as 2016. In the analysis report of deep learning framework and platform market share released by IDC in 2019, the platform ranked the third in the domestic market, with Google and Facebook holding nearly 80% of the domestic market share.

Baidu's computer vision technology capability and related products are all built on this platform. In the process of developing the face mask recognition algorithm, the training database based on Baidu PaddlePaddle (PaddlePaddle) has been trained by using the large scale Classification Library PLSC of Baidu PaddlePaddle (ID). A high-performance face recognition model is proposed, and the rapid deployment of cloud and mobile terminals is realized by using paddlelite. And on February 13, baidu announced the first open-source face detection and classification model of face masks based on paddle in the industry. This model can effectively detect all faces in densely populated areas and judge whether they wear masks. At present, it has been open-source through paddle paddlehub.

In addition, CVPR, a top academic conference on computer vision In 2020, the paper collected was published recently, with the enrollment rate at a new low, down 3% compared with last year, only 22%. However, baidu still has 22 papers selected, among which hambox, an online high-quality anchor mining strategy for the papers collected in face recognition and detection, facescape, a database of large-scale and high-precision face 3D models, can be controlled by high-precision pre measurement from a single image, and facescape, a database of large-scale and high-precision face 3D models, are proposed e. Based on the neural network architecture, the model and method of feature extraction and feature pyramid bfbox are searched simultaneously.

On the ability of technology commercializationThe "face recognition algorithm with mask" developed by Baidu has also been combined with Kesheng intelligence, Shiniu information, Huajie electronics and intelligent technology in application fields such as entrance control transformation of Village rental house in the city, intelligent campus, public transportation face recognition, taxi driving behavior analysis, etc. In addition, more than 100 online stores have started to use such application scenarios as senseless face brushing shopping, smart community face brushing access control, and enterprise employees face brushing attendance.

It can be seen that Baidu vision technology and its R & D team have become children of other families, who are promoting the continuous commercialization of computer vision in various industries.

"Eat by face" will come

Although the current user's interaction habit still stays in the key, touch screen, voice interaction mode, "unlocking" habit still stays in the password, fingerprint and other recognition methods, but due to the convenience and security of face recognition, the future will eventually be a "eat by face" era.

Before that, how to make the adaptability, maturity and security of visual algorithm to the extreme in the complex environment still depends on the long-term polishing of each R & D team and open source power.

China IT News APP

Download China IT News APP

Please rate this news

The average score will be displayed after you score.

Post comment

Do not see clearly? Click for a new code.

User comments