Research On Super-resolution Of Face Image Based On Generative Adversarial Networks

In recent years,the study of super-resolution of face images is of great importance and has received more and more attention.Traditional super-resolution of face images often uses interpolation methods that easily lead to rough texture structures and even distorted face structures.Although the generative adversarial network-based face image super-resolution method has powerful data generation capability,it cannot achieve complex mapping from low-resolution input to super-resolution results due to the lack of feedback,and the reconstruction results have problems such as local blurring or distortion.Therefore,this paper proposes an improved method for super-resolution of face images based on generative adversarial networks,with the following main contributions.(1)Traditional face image super-resolution methods use only feed-forward structures,implement one or more upsampling operations on the extracted features,scale the image to the desired size,and use non-accurate face keypoint information and weak constraints to limit the size of the solution space,so the model results are blurred and the visual perception is poor.To this end,this paper proposes a face super-resolution generative adversarial network model based on face keypoint information and error feedback.The model firstly adopts an iterative up-and-down sampling alternation module based on error feedback to double-feedback correct intermediate features through a circular iterative feedback structure,and an error formation and feedback mechanism inside the sampling block.Second,a recurrent network structure is designed in which face super-resolution reconstruction and face key point extraction promote each other.The reconstruction results are improved by extracting accurate face keypoint information from the results of face super-resolution reconstruction,and then inputting them into face super-resolution reconstruction.Finally,a combined optimization objective is designed.By focusing on the extracted face keypoint information at pixel level and channel level separately,the edge information of relative discriminative adversarial loss focused face sharpening is also introduced.(2)To address the problems of roughness,blurring or artifacts of facial features in super-resolution reconstruction using only face key point information.In this paper,we propose a super-resolution model of face that combines face attribute information and face key point information.Face attribute information as a kind of semantic information can overcome the problems such as local attribute distortion and texture roughness.Traditional methods directly stitch non-identical low-resolution face images and attributes,which weaken the expression of face attributes.However,this model extracts the features of the upsampled image by coding structure,fuses the attribute information with the extracted features,and then by decoding structure,the upsampled image reconstructed based on the attribute information is obtained.Finally,the upsampling results are used in the reconstruction of the previous model.In addition,in order to make full use of the face attribute information,the attribute-related loss function is added to the optimization objective and the training parameters are adjusted by back propagation,which helps the network to extract more attribute-related features and further improve the quality of image generation.To verify the effectiveness of the proposed method,this paper conducts several experiments on two standard face datasets,Celeb A and Helen.The final experimental results show the effectiveness of the proposed method compared with the typical methods.