Research On Masked Face Recognition Technology Based On The Dual Branch Network

With the outbreak of the COVID-19 epidemic around the world,people almost always wear masks in public places to prevent the spread of the virus,which brings a huge challenge to the current face recognition system,and the masked face recognition technology has become a research hotspot.The research on masked face recognition technology in open scenarios faces two major problems.First,there is currently a lack of well-labeled largescale public masked face datasets,and the uneven quality of masked face images in open scenarios increases the difficulty of recognition;Second,the large area of mask occlusion in the masked face image reduces the recognition precision.In order to solve the first problem,the masked face image synthesis algorithm based on image texture fusion technology and face landmarks location technology is implemented respectively.The second synthesis algorithm is improved,and the FAN model is used for face landmarks detection to improve the quality and speed of masked face image synthesis.The masked face test dataset and training dataset are generated on LFW,MS1M-Retina Face and UMDFaces face datasets respectively.The training dataset is expanded by data augmentation methods such as horizontal flip,motion blur,and color jitter,which provides the data basis for improving the generalization ability of the masked face recognition model in open scenarios.To solve the second problem,a dual-branch network for masked face recognition is designed.The global branch network is used to extract the global features of the masked face image,and the Vision Transformer is used as the backbone network.On this basis,improvements are made.A small convolutional neural network is used as the input features embedding layer,the Layer-Scale operation is added to the encoder,and encoder layers are divided into SA layers and CA layers,which is aims to enhance training stability and improve recognition precision;the local branch network is used to extract the eye region features of the masked face image,and an eye region location algorithm is designed,using Res Net as the backbone network,and the first convolutional layer and the final classification head are modified to suit the masked face recognition task.The features extracted by the global branch and the local branch are finally merged in the channel dimension for feature fusion,which further enhances the robustness and discrimination of the dual-branch network against mask occlusion in the process of extracting facial features.On this basis,a masked face recognition system is designed and implemented.The proposed method is tested and evaluated on the synthetic masked face test dataset LFW and the real masked face test dataset MFR2.The evaluation results show that the masked face recognition precision of the method on LFW and MFR2 is 90.49% and 82.18%,the accuracy is 96.73% and 98.13%,respectively,and the forward inference time overhead is 24.25 ms.