Tencent Youtu overcomes the problem of mask recognition, the accuracy of mask wearing recognition exceeds 99%

With the battle against the new crown epidemic officially launched, masks have played a crucial role in controlling the epidemic, but the wearing of masks for all people has also challenged scenes such as high-speed rail gates that require face recognition:  people wearing masks have a large face area Obscured by the mask, the existing algorithm cannot accurately detect the position of the face and locate the key points of the facial features, greatly reducing the effect of the existing face recognition algorithm . In addition, removing the mask in a public place and relying on manual inspection not only consumes a lot of manpower, and the efficiency of inspection is low, but also increases the risk of infection of front-line workers. In order to solve this problem, Tencent Youtu quickly set up a tough team during the Spring Festival to develop and optimize algorithms for different mask wearing scenarios , and finally overcome the problem.

Youtu focuses on face detection, face registration (keypoint positioning), face attributes, face recognition and other technologies. Currently, you can detect face wearing masks in real time, accurately identify five different mask wearing situations, and Early warning is provided to those who do not wear masks or wear masks by mistake. On this basis, Youtu DDL face recognition technology further enhances the ability to distinguish the visible area of ​​the face, and achieves more robust face recognition.

Face Detection

Based on the open source DSFD face detection algorithm of Youtu , for the facial features occlusion in the mask wearing scene, Tencent Youtu performs local feature enhancement on the model design to increase the weight of the visible area. At the same time, in view of the problems such as the variety of masks and the variety of wearing positions, corresponding strategies are designed in terms of data enhancement to improve the robustness of the model. At present, the accuracy rate of face detection algorithms in mask scenes exceeds 99% , and the recall rate exceeds 98% .

Face registration (keypoint positioning)

In order to solve the problem of large-scale occlusion of the face area caused by masks, based on the multi-branch lightweight neural network self-developed by Youtu , Youtu quickly synthesizes large amounts of face mask data through image editing technology for algorithm optimization and improvement Precise facial features positioning effectively assists the subsequent algorithm module to improve the effect.

Mask attribute recognition

At present, the excellent picture algorithm can accurately recognize the following five situations: not wearing a mask , wearing a mask incorrectly and covering the mouth , wearing a mask incorrectly and covering the chin , wearing a mask incorrectly without covering the face , and wearing the mask correctly . This attribute recognition is based on the FAN attribute recognition of Youtu open source , and more attention mechanisms are added to the position of the face that the mask may be distributed, which can accurately identify whether the face is worn correctly. At present, the accuracy of identifying whether or not to wear a mask exceeds 99%. Community managers can freely combine these categories according to the needs in different scenarios. At the same time, various enterprises and institutions can also use this technology to detect the situation of employees in time to ensure safe resumption.

Face recognition with masks

Utopia provides a flexible and safe algorithm solution. Two types of analysis are used to judge the mask occlusion and extract the occlusion area of ​​the face masked by the mask using the Youtu face quality model . Among them, the judgment of mask masking has reached an accuracy rate of over 99.5% . For application scenarios with extremely high security requirements, such as payment scenarios, people who wear masks or masks that are severely masked can be screened out based on the mask blocking judgment results, and further guided to perform other forms of identity verification. The algorithm is based on the self-developed DDL technology framework , combined with the occlusion area judgment ability of the Youtu face quality model , so that the data model can adaptively pay attention to the face discrimination information in the non-mask area when responding to the face wearing a mask, so as to extract More robust facial features.

Youtu DDL Face Recognition Technology

Conventional face recognition algorithms, even when applied to face recognition with masks under coordinated conditions, performance will be greatly reduced. The Youtu face recognition algorithm, based on the above optimization methods, can increase the recall rate of face recognition with masks to be close to the normal rate of face recognition , which basically meets the face recognition application in the scene of wearing masks. Compared with the face recognition technology, the human body recognition technology (ReID) based on the image “search for people” is more robust to the occlusion, orientation and sharpness of the human body image. There is no hard requirement.

During the epidemic, most people who go out will wear masks, and facial recognition technology will reduce the success rate of people wearing masks. For front-line workers in the community, the failure of face recognition technology to confirm the identity of the person wearing the mask will greatly increase their workload of investigation and registration, and the removal of the mask for identification will increase the potential transmission risk.

Based on Tencent Youtu’s current industry-leading ReID technology , Tencent Youtu and Tencent Heiner use the combination of human body features and face recognition to confirm the accessibility of people wearing masks that are not traceable under traditional face recognition, thereby enhancing the community The staff ranks the efficiency of registering outsiders.

At present, related technologies have been successively implemented in application scenarios in many different regions , and the value of AI continues to be exerted in this national war against epidemic.