Occlusion Face Recognition Technology Based On Deep Learning

With the rapid development of computer vision,face recognition technology is applied to many fields of people’s life.In recent years,various face recognition algorithms have been continuously proposed,and have good accuracy effects on face recognition datasets.However,in actual scenes,the face may be partially occluded,resulting in the loss of key point information on the face.The existing models are affected by occlusion factors,so that the accuracy rate cannot be improved.This thesis aims to propose two occluded face recognition models based on deep learning from the two technical routes of occlusion aware and occlusion recovery,so as to further improve the accuracy of occluded face recognition.At the same time,the applicability of the face recognition models of the two technical routes in different application scenarios is studied and compared.The main work of the thesis is as follows:1.A face mask occlusion dataset was constructed.In view of the limitations of the current face occlusion dataset,which is small in scale and some datasets use black and white rectangular images to occlude faces,which do not fit the actual application scenarios,this thesis uses the Mask The Face algorithm to build a self-built mask occlusion face dataset,including368,715 occluded face images of 10575 identities.This dataset can better simulate the diverse mask occlusion scenes in life and improve the generalization ability of the occluded face recognition model.2.Based on the occlusion-aware face recognition technology route,an improved face feature rectification network model Inv-CAMAC-i FFRNet is proposed to solve the problem of insufficient feature extraction of unoccluded areas in face images.The spatial branch in the face feature rectification module introduces involution operator to extract richer facial feature information in the spatial range;the channel branch in the face feature rectification module introduces a coordinate attention mechanism to capture cross-channel information to enhance features representation;Meta-ACON is used as a new dynamic activation function of the feature rectification module,and the generalization ability and calculation accuracy of the model can be improved by dynamically adjusting the degree of linearity or nonlinearity.The model is applied to the two scenarios of mask occlusion and glasses occlusion for testing,and the test accuracy respectively reaches 82.50% and 89.75%,which is better than the existing algorithm and verifies the effectiveness of the proposed method.3.Based on the occlusion recovery face recognition technology route,gradually recurrent occluded face recognition model SA-PCA-RRMFR is proposed with the fusion of attention mechanism,which is further optimized on the basis of the RFR-Net image recovery model.The spatial attention mechanism is introduced in the redisual block to seek the relationship in the image space,helping the model focus on the information-rich pixel areas in the image;a redesigned channel self-attention module P-CA is introduced to obtain the channel weight from a global perspective;combined with the Arc Face algorithm for face recognition.Test on the Celeb A dataset,when the occlusion ratio is 10%-20%,30%-40%,50%-60%,the face recognition accuracy respectively reaches 94.8%,91.5% and 89.4%,the proposed method is superior to most mainstream algorithms in terms of image recovery index and accuracy,and can generate high-quality face images with consistent semantics and reasonable vision,which can be applied to face recognition.4.Compare the advantages and disadvantages of two face recognition models in different occlusion scenarios.The advantage of the Inv-CAMAC-i FFRNet model is that it can make full use of the effective information area in the occluded image and maximize the role of the unoccluded area information.The disadvantage is that it cannot infer the full face information from some areas.The advantage of the SA-PCA-RRMFR model is that it can restore the incomplete image to a complete face image and obtain the most likely full face image.The disadvantage is that it will generate semantically inconsistent face images for large-area occlusions,which will affect face recognition process.