Research On Image Restoration Technology Based On Generative Adversarial Nets

With the latest developments in image processing technology,especially the increasing flexibility of image processing tools,traditional image restoration techniques have encountered challenges.In recent years,image restoration technology based on deep convolutional networks has made tremendous progress in the field of image restoration because it can obtain global features.Compared with the method based on the deep convolutional network,the image restoration technology based on the generative adversarial network can restore the original image with higher resolution,so it has become a current research hotspot.Based on generative adversarial networks,this paper first studies face image restoration with a single feature,then studies image restoration of complex areas and cross-category image restoration,and studies face restoration at multi-angle and expression finally.The innovation achievements mainly include:1.Research on face restoration algorithms based on generative adversarial networks.A blurred face restoration algorithm and a masked face restoration algorithm based on generative adversarial networks are proposed.By analyzing the existing blurred image restoration technology,the structure of the generator model is first modified from the process of completely upsampling to the process of first downsampling and then upsampling.Secondly,in order to obtain a clearer recovery result,this paper adds the L1 loss to the cross-entropy loss used by the original generative adversarial networks.Compared with traditional algorithms,the proposed model improves the recognition rate of recovery results by 10%.At the same time,a masked image restoration algorithm based on generative adversarial networks is designed.A secondary discriminator model is added on the basis of the generator model and discriminator model of the original generative adversarial network.The secondary discriminator model is used to reduce the local difference between the masked area and the unblocked area.The discriminator model is used to reduce the overall difference between the masked image and the unblocked image.At the same time,due to the existence of the secondary discriminator model,the loss of the generated model adds four new parameters.Under the premise of local and global restoration,the proposed model restoration results are superior to the local sample block search restoration algorithm.2.A complex image restoration algorithm based on generative adversarial networks.The global mosaic removal algorithm,a visible watermark removal algorithm and a fast masked image restoration algorithm based on generative adversarial networks are proposed.In order to transition from face restoration to complex image restoration beyond the face,two novel parsing networks are added to the original generative adversarial networks with a generator model and a discriminator model to obtain more detail difference between the generated image and the original image.After adding the parsing networks,for the global mosaic removal,a maintenance and restoration algorithm based on mean square error is proposed.In this paper,the maintenance and restoration image algorithm is used to calculate pixel loss and content loss,so as to obtain the highest score on indicators such as peak signal-to-noise ratio.After adding the parsing networks,a solution to adding a self-attention layer to the generator model was proposed for the problem of visible watermark removal.An adversarial network model is generated to establish the mapping between the watermark image and the real image.At the same time,the self-attention layer is used to extract the invariant features of the watermark regions in different watermark images.Finally,a novel structural similarity loss algorithm is proposed to find the pixel difference between the randomly visible watermark image and the real image.The trained model can delete the visible watermark at any position and restore the result close to the real image.Compared with the previous model,the proposed model improves the peak signal-to-noise ratio and other indicators by at least 10%.After adding the parsing networks,a fast generative adversarial networks model is proposed to solve the problem of masked image restoration.Firstly,it is proposed to use the fast marching algorithm to obtain more adjacent pixels before the generator network processing.Then it is proposed to use the square root loss to calculate the pixel loss and content loss,and restore the masked image.Finally,a method for measuring the stability of the generated model is proposed,and the great advantage of square root loss in obtaining pixel differences and content differences is verified mathematically.3.A cross-category image restoration algorithm based on generative adversarial networks.A cross-category complex image restoration algorithm for wild datasets is proposed.In this paper,the potential space between different categories is divided into global and local,and a weakly supervised similarity generative adversarial networks model is designed to capture the local latent space instead of the global latent space,so as to establish a mapping relationship between images of different categories and avoid the collapse of other models that they can not get the correct global potential space,and finally achieve cross-category image restoration.4.Multi-angle and facial expression restoration algorithm based on generative adversarial networks.In view of the infinite number of possible problems of multi-angle and facial expression restoration results,a method of classifying similar facial angles and facial expression features using a classifier is proposed.The classified model only needs to learn a limited number of mappings,thus solving the convergence problem of multi-angles and facial expression restoration in an infinite variety of situations.The generator network of the existing model usually uses low-latitude classification labels as bundles,but it cannot completely constrain the shape and size differences between different faces and lead to erroneous restoration results.This paper proposes to use high dimensions classification features in the generator network as the constraint condition,and the obtained recovery results have good compatibility.Finally,for the problem that the improper sequence control of restoration learning will affect the restoration effect,this paper proposes to use two loss functions with different convergence speeds to learn the angle and expression characteristics of the face in sequence.At last,we achieve better results both in the face angle and the expression restoration.