Research On Face Recognition Method Of Mask Occlusion Based On Deep Learning

With the gradual improvement of postnatal level,people pay more and more attention to the security of payment methods.Due to the uniqueness of facial biology and the noninvasiveness of the recognition process,facial recognition has gradually become the payment method of choice for most people.However,the conditions for satisfying face payment are relatively limited,and frontal face images with sufficient light and no occlusion are required,and liveness detection is also required.Most of the current face recognition models achieve high recognition rates under restricted conditions,and often fail to obtain ideal recognition results in the face of complex detection situations.Due to the outbreak of the new crown pneumonia,strict epidemic control policies have been implemented in many places,and most people choose to wear masks when traveling.In this case,in order to accurately identify the target identity,using the previous face recognition target can no longer meet the requirements.Based on this situation,this paper proposes a face recognition method based on deep learning that can be occluded,thus providing a solution to the above problems.Aiming at the problem that the occlusion situation affects the recognition accuracy of the face model,this paper improves the training data set,including face area clipping,data cleaning and so on.In the face detection part,the SSD algorithm with better detection performance is selected through experimental comparison.Based on the idea of robust feature extraction,this paper chooses to use neural network to extract features from unoccluded face regions.In order to enhance the feature extraction ability of the network for occluded regions,this paper introduces an attention mechanism into the neural network to enhance the robustness of the model to occlusion.The attention mechanism includes channel attention module and spatial attention module.Since the features of different channel dimensions of the image are interrelated,the channel attention module can learn more channel features while ensuring the range of feature extraction through pooling operations.The spatial attention module uses maximum pooling and average pooling in series to highlight the feature information of unoccluded faces to ensure the robustness of the module to occluded faces.The backbone network in the paper adopts the Inception structure and residual block to ensure the network depth and training accuracy.The experiment inserts the attention mechanism module into the backbone network of Inception ResNet V1,and compares it with other face recognition models on the manually occluded LFW dataset.The results show that the recognition accuracy of the improved network is better than other networks in the case of occlusion,but still can not meet the identification requirements.For the problem of low recognition accuracy,this paper adjusts and improves the model training process.The training dataset is augmented by random cropping,horizontal flipping,random noise,random brightness,random rotation,random modules,etc.Using the trained model to conduct experiments on the LFW model mask dataset,the results show that after adding data augmentation,the model has higher recognition accuracy for occluded pictures.Finally,through the accuracy test under different thresholds on the VGGface2 data machine,the results show that when the threshold is 0.75,the model accuracy and true rate are higher,and the false positive rate is lower,so the model face matching threshold is 0.75.Based on the face recognition model based on the attention mechanism,this paper proposes a real-time face recognition method,which includes a real-time video streaming module,a face detection module,and a face recognition module.Under different occlusion methods and face rotation angles,the target identity can be more accurately identified...