Research On Face Image Inpainting Based On Deep Learning

Image repair refers to the restoration of the damaged area by analyzing the information of the undamaged area,and then infer the content of the damaged area.Face repair as a branch in the field of image repair,it often contains more features,such as eyebrows,eyes,mouth,nose,etc.,the traditional method of face image repair although has a good repair effect,but the repaired image often fuzzy and missing details and texture features,and traditional method only applies to the image damage area smaller,also cannot guarantee the overall repair results semantic and structure of the consistency.In the field of deep learning with the rise of GAN,GAN is widely studied in face image repair,this is because it has a powerful image generation ability and feature learning ability,not only can repair more complex missing type and damaged area larger damaged image,at the same time on the semantic and structure also maintains the consistency with the original image.However,there are often problems such as gradient disappearance,training instability in the image repair tasks,which exist in the generative adversarial network.On the basis of the original generated adversarial network,this thesis proposes the face repair model with the dual discriminant generated adversarial network,and then improves the problems of the dual discriminant face repair model with slow training speed and unable to understand the image context information better.The main work of this thesis is as follows:(1)The face image inpainting network based on the double-discriminator GAN is used;the discriminator consists of a local discriminator and a global discriminator,the former ensures the consistency of the repair results with the surrounding area,the latter ensures the overall consistency of the repaired image and the real image;generator uses a convolutional neural network,and added jump connections to it,to improve the predictive ability of structural information;in order to improve the model stability during the training process,introduction of the WGAN-GP,thus completely solve the problem of gradient explosion and pattern collapse,at the same time,the model can also obtain rich generated samples;in the discriminant,to avoid the interdependence of the samples in the same batch,layer normalization instead of batch normalization is used in the discriminator.(2)Using a two-stage face image inpainting approach based on context attention,the method uses a two-phase repair network based on generative adversarial network,divide the generator into two stages: coarse repair and fine repair,during the first stage of the crude repair,using expansion convolution to increase the receptive field while introducing context attention,during the second phase of the fine repair,contextual attention is also introduced to enhance the face image repair network to understand contextual information,capture more information around the broken area,expansion convolution was also used to increase the receptive field,improve the repair effect of the damaged area;using a discriminant of the Patch GAN structure,to get clearer texture features;in the loss function,in addition to using adversarial losses and reconstruction losses,perceived losses and style losses were also introduced for better remediation results.