Home > News content

Tencent Youtu overcomes the problem of mask recognition, and the accuracy rate of mask wearing recognition is more than 99%

via:博客园     time:2020/2/23 10:08:41     readed:417

With the battle against the new crown epidemic officially launched, masks play a key role in controlling the epidemic, but the wearing of masks by the whole people also challenges the scenes that need face recognition, such as high-speed railway gate Because the face area of people wearing masks is covered by masks in a large range, the existing algorithms can not accurately detect the position of the face and locate the key points of the facial features, which greatly reduces the effect of the existing face recognition algorithms. In addition, removing masks in public places and relying on manual screening not only cost a lot of manpower and low efficiency of screening, but also increased the risk of infection of front-line staff. In order to solve this problem, during the Spring Festival, Tencent Youtu quickly set up a team to research, develop and optimize algorithms for different scenes with masks, and finally overcome the problem.

The recognition accuracy of wearing mask is higher than 99%, and the face recognition technology under mask is conquered by utu

Utu focuses on face detection, face registration (key point positioning), face attributes, face recognition and other technologies. At present, it can detect face wearing mask in real time and accurately recognize five different situations of wearing mask, and timely detect and warn those who do not wear mask or wear mask wrongly. On this basis, the optimal graph DDL face recognition technology further enhances the ability to distinguish the visible area of the face, and realizes more robust face recognition.

In the aspect of face detection, based on the open-source DSFD face detection algorithm of optimal graph, Tencent optimal graph enhances the local features in the model design and improves the weight of the visible area in the face mask scene. At the same time, in order to improve the robustness of the model, the corresponding strategies are designed in the aspect of data enhancement. At present, the accuracy of face detection algorithm in mask scene is more than 99%, and the recall rate is more than 98%.

orgsrc=//img2018.cnblogs.com/news/66372/202002/66372-20200222214826850-373075971.jpg

Face detection effect under different occlusion conditions (pictures from the network)

In face registration (key point location), in order to solve the problem of large-scale occlusion of face area brought by masks, based on multi branch lightweight neural network developed by Youtu, Youtu quickly synthesizes massive face mask data through image editing technology for algorithm optimization and promotion, realizes accurate facial features location of people wearing masks, and effectively assists the effect improvement of subsequent algorithm modules.

orgsrc=//img2018.cnblogs.com/news/66372/202002/66372-20200222214826868-907050475.jpg

Left: before optimization, due to large-scale occlusion of the face, the key point location deviation of the face is large

Right: after the optimization of the algorithm, the algorithm can effectively estimate the coordinates of the correct key points of the face with mask

(image from: MafA dataset)

In terms of mask attribute recognition, at present, the optimal graph algorithm can accurately recognize the following five situations: not wearing a mask, wearing a mask mistakenly and covering the mouth, wearing a mask mistakenly and covering the chin, wearing a mask mistakenly and not covering the face, and wearing a mask correctly. The attribute recognition is based on the open-source fan attribute recognition of optimal graph, and more attention mechanisms are added to the face position where the mask may be distributed, so as to accurately recognize whether the face is correctly worn. At present, the recognition accuracy of whether or not the mask is worn is more than 99%. Community managers can freely combine these categories according to the needs of different scenarios. At the same time, enterprises and institutions can also use the technology to detect the situation of employees in time to ensure safe return to work.

orgsrc=//img2018.cnblogs.com/news/66372/202002/66372-20200222214826849-360574107.jpg

Five types of masks that can be identified

In face recognition with face mask, excellent image provides a set of flexible and convenient algorithm solutions. In this paper, we use the optimal face quality model to judge the mask occluded face and extract the occluded area. At present, the accuracy of mask occlusion judgment is over 99.5%. For application scenarios with high security requirements, such as payment scenarios, people who wear masks or face masks are seriously occluded can be screened based on the mask occlusion judgment results, and further guide them to carry out other ways of authentication. This algorithm is based on the DDL technology framework developed by ourselves, combined with the occlusion area judgment ability of the excellent face quality model, so that the data model can adaptively focus on the face discrimination information of the non mask area when dealing with the face wearing mask, so as to extract more robust face features.

orgsrc=//img2018.cnblogs.com/news/66372/202002/66372-20200222214826876-1165919303.jpg

Optimal DDL face recognition technology

The performance of the conventional face recognition algorithm, even when it is applied to face recognition with mask, will be greatly reduced. The optimal face recognition algorithm, based on the above optimization method, can improve the recall rate of face recognition with masks to close to the recall rate of normal face recognition, which basically meets the application of face recognition in the scene with masks.

Assist the community in personnel management and screening in combination with human body recognition under masks

Image based compared to face recognition

During the epidemic, the vast majority of people who go out will wear masks, and face recognition technology will reduce the success rate of people who wear masks. For the community front-line workers, face recognition technology fails to confirm the identity of the people wearing masks, which will greatly increase their screening and registration workload, while removing masks for recognition will increase the potential communication risk.

Based on the industry-leading Reid technology of Tencent Youtu, Tencent Youtu and Tencent Haina use the combination of human features and face recognition to confirm the entry and exit personnel wearing masks that can not be traced under the traditional face recognition mode, so as to improve the efficiency of community staff in arranging and registering outsiders.

At present, relevant technologies have been successively applied in many different areas, and the value of AI has been continuously played in this war of national anti epidemic.

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