Research On Face Image Restoration Method Based On GAN Networ

The objectivity of image restoration is mainly to restore reasonable pixels in the damaged area.Face restoration is an interesting and challenging task in the field of computer vision and image processing.Compared with the rich and diverse natural scenes,facial images are familiar and pure.Because the appearance changes greatly,such as different postures,expressions and occlusion.An excellent face repair algorithm should ensure that the algorithm can generate the topology between different parts of the face,as well as the generation similarity of attributes such as posture,expression and character.Therefore,any tiny defect can be easily found in this type of image.Therefore,repairing facial images has always been a difficult task.The traditional image filling technology generally uses the diffusion method to transfer the low-level features from the intact area to the damaged area.In recent years,the development of deep learning technology and convolutional neural network has greatly promoted the progress of image restoration.At present,the mainstream image generation and restoration algorithms are designed based on the generation of countermeasure network.Obviously,a complete facial image may contain a lot of details,so it is difficult to reconstruct directly when there is serious damage.In order to solve this problem,we choose to restore the overall image structure represented by the analytical image.Compared with the details of the original image,the analytical image is very simple.Compared with the direct whole face restoration,the restoration of analytical image is much easier,because its structure is obviously simpler and contains much less detailed information.Face semantic analysis contains the structural information of the face and succinctly expresses various attributes of the face.Face analysis is a good indicator.It is neat,sufficient and robust,and can reflect the structure of the face.Therefore,this paper proposes a new idea of face image restoration.Firstly,the face semantic analysis image is reconstructed,and the reconstructed image can clearly reflect the structure of the whole face.Compared with the direct whole face restoration,the restoration of analytical image is much easier,because its structure is obviously simpler and contains much less detailed information.On this basis,we design a generative face analysis guided repair network framework based on generative confrontation network,namely GAN network.The two-stage structure is a face restoration framework composed of face analysis and restoration subnet and image restoration subnet.The first stage is the prediction subnet of face analysis,and the second stage is the compensation subnet of face restoration.We designed two new modules at the same time.The semantic compensation module is added to the face analysis and reconstruction network.The module is designed based on void convolution to ensure the effective aggregation of context information.On the other hand,in the second stage,a new module structure is introduced,which improves the self attention module and introduces the features of the coding layer to select fine-grained features to ensure the consistency of the image.A large number of experiments are carried out on the public Celeb A-HQ data set to verify the effectiveness of the proposed method.