Face Occlusion Area Recovery And Posture Correction Based On Generative Adversarial Network

Face recognition refers to a technology that uses a computer system to analyze the input face and compare its effective feature information to distinguish the identity of the input person.Among them,dealing with face changes is the key to face recognition.The traditional face recognition technology tends to mature in a restricted environment,but in an unrestricted environment,due to factors such as occlusion,illumination,posture,etc,the face will be different,so the performance of face recognition will be obvious Decline.It is of great significance to study the difference of face in unrestricted environment for practical application.The main work of the thesis aims at the occlusion and attitude factors in the unrestricted environment.By combining the generative adversarial network framework,the following two face generation algorithms are proposed.(1)Aiming at the problem of face occlusion in an unrestricted environment,the thesis proposes a face image repair algorithm based on generating an adversarial network to achieve the purpose of restoring missing areas and complementing identity features.The algorithm introduces a model for predicting face analysis to improve the authenticity and clarity of the repaired face image,and at the same time introduces a cross-sensing attention mechanism to improve the feature correlation between the foreground area and the background area.In order to improve the repair effect of face images under occlusion of any shape,the algorithm also introduces gate convolution and Patch-GAN discriminator.Experimental results show that the algorithm proposed in the thesis is improved in both visual effects and evaluation indicators compared to current image restoration algorithms.(2)Aiming at the problem of multi-pose in an unrestricted environment,the thesis proposes a face pose conversion model(FPM)based on generative confrontation network from the perspective of style transfer.The algorithm uses the ability to generate domain migration against the network,and transforms the gesture of the input facial image to the specified gesture to achieve the purpose of eliminating facial differences and purifying facial identity.FPM introduces the loss of identity perception based on the face recognition model,and uses its a priori knowledge to maintain the face identity during posture transformation.In order to stabilize the training process and training effect of the network,FPM also introduces pixel-level loss by generating face pose restoration.Experimental results show that the FPM proposed in the thesis can generate high-quality,high-confidence face with specified pose,which improves the performance of face recognition.