Research On Face Repair And Recognition Method Under Partial Occlusion

In recent years,face recognition technology has been successfully applied in various application scenes,such as security monitoring,face-scan payment,intelligent access control and so on.However,in real applications,it may face serious challenges: the criminal can cover his face to avoid sky eye system,the face captured by a traffic camera is unable to be identified and the face recognition system can’t recognize the masked faces which exist widely during the COVID-19 pandemic.Thus,there is an urgent need and important application value to study face recognition methods with partial occlusion.In this dissertation,methods of face repairs and recognition under partial occlusion are studied.The main works are as follows:To repair face occlusion,this dissertation proposes a partial occlusion face repair algorithm,named STRDE-GAN.Some facial parts are prominent and unique in appearance,the occlusion of which will degrade the performance of the face recognition algorithm.For the purpose of reconstructing the missing parts of a face,a generator model based on the noise-removing autoencoder is introduced into the generative adversarial network framework to remove the noise caused by the reconstruction of the face.In addition,a new dual-mode training algorithm is designed.In the training process,mode 1 generates the face with ambient illumination,and mode 2 eliminates the noise generated by mode 1.Based on this,a unique training algorithm is proposed,which has a fast convergence speed.Finally,to maintain the overall quality of the face image,an adversarial "structural" loss is also proposed,which includes a fragmented mean square error focusing on the structural similarity of the overall facial features and the pixel difference.The proposed STRDE-GAN model is tested on the AR dataset and the Celeb A dataset.Experimental results show that the model has high accuracy and can accurately restore the important features compared with the original image without occlusion.Aiming at the problem of low face recognition accuracy with occlusion,a partial occlusion face recognition algorithm is proposed,named YOLOV5-Tripatt.The deep network of YOLOV5-Tripatt model has three important components:backbone network PO-Res2net-101,bidirectional feature fusion network Tri-PAN and classification regression network.The backbone PO-Res2net-101 is divided into five stages.The first two stages only need a few parameters and can complete the maximum pooling operation,and the last three stages introduce non-local operation and deformable convolution.Tri-PAN,a bidirectional feature fusion network,introduces a triple attention mechanism to extract more profound features useful for occluded face recognition through cross-dimensional interaction.In addition,in the YOLOV5-Tripatt model,the localization loss is innovatively proposed to reflect the localization accuracy.The proposed YOLOV5-Tripatt model is tested on the Celeb A and RMFD datasets and experimental results show that the proposed algorithm outperforms the mainstream face recognition algorithms.