Research On Key Technologies Of Face Recognition Based On Mask Constraint

In the field of artificial intelligence and computer vision,face recognition is a basic and important task.It has a wide range of application scenarios such as security check of the station,work sign-in,ticket check of the scenic spot,hotel check-in,gate control of the campus,etc.It provides great convenience and safety guarantees for people's lives.However,due to the recent large-scale outbreak of the COVID-19,people need to wear masks when traveling.The widely deployed face recognition system is facing the recognition challenge represented by facial occlusion.In crowded transportation hubs,people have to take off their masks and pass the verification of the face recogni-tion system.It not only increases the spread risk of the epidemic but also reduces the verification speed of face recognition.Mask Face recognition has seriously interfered with people's daily life and has become an urgent problem to be solved.As an emerging problem,there are currently no large-scale mask face datasets for research and use,and it is difficult to produce qualified large-scale mask face datasets from real scenes in the short term.In previous studies,masks are usually integrated into the problem of facial occlusion for the research,so there is a lack of targeted solutions for mask face recognition.To solve the above problems,this paper conducts detailed research on the construction method of mask face datasets and mask face recognition.The specific research content and contribution points include:(1)Introduced a method for constructing mask face datasets based on keypoint detection.This paper considers constructing the mask face by ”fitting” the mask template on the face image.Due to the different face positions,posture,and size in different images,the position,shape,and size of the mask fitting area are vari-ous.To solve the above problem,this paper introduces a mask face construction method based on keypoint detection.It can adaptively locate the fitting area of the mask by detecting the key points of the face.This paper uses the method to create a large-scale mask train dataset Masked CASIA-Web Face and mask test datasets Masked LFW,Masked CFP,and Masked Age DB.The experimental results on the mask test datasets show that the basic model of face recognition Res Net has a recognition degradation ranging from 7 ? 14%.The experimental results on the mask train dataset show that the basic model of face recognition Res Net has a recognition improvement ranging from 2 ? 4%.It is in line with the law of face recognition,which proves that the mask face datasets produced in this paper can be used in experiments.(2)Proposed a method of mask face recognition based on dual attention.Inspired by the character of the attention mechanism which can help the model re-focus on the key features,this paper proposes a dual attention module to help the face recognition model to re-focus on the key features of the mask face.Specifically,the dual attention module is divided into two parts: channel attention and spa-tial attention.Different from the previous methods,this paper divides attention learning into two parts: information aggregation and weight activation.In infor-mation aggregation,this paper designs an attention pooling network to aggregate feature information losslessly.In the weight activation,this paper designs the ac-tivation network of the large receptive field to strengthen the learning of spatial attention.The experimental results on the mask test datasets show that the dual attention module is better than the popular attention modules and helps the basic model Res Net improve the recognition performance ranging from 1 ? 4%.In addition,dual attention has also obtained the best recognition decision basis in the visualization experiment.(3)Proposed a method of mask face recognition based on spatial guide learn-ing.Based on the prior information that the mask is usually in the lower half of the face,this paper designs a loss function to guide the learning of the model in the feature space,reducing the interference of the mask information in the face features.Specifically,based on the attention module,this paper designs a loss function to constrain the learning of feature space attention,which can strengthen the weight of the upper half of the spatial attention map and weaken the weight of the lower half of the spatial attention map,blocking the flow of mask informa-tion in the network.The experimental results on the mask test datasets show that the spatial guide learning method is better than the popular mask face recognition methods and helps the basic model Res Net improve the recognition performance ranging from 3 ? 5%.In addition,spatial guide learning has also obtained the best recognition decision basis in the visualization experiment.