Research On Image Restoration Algorithm Based On Generative Adversarial Network

With the development of digital technology,image restoration technology has been widely used in many fields and has been an important research direction in the field of image processing.Traditional image restoration methods such as inverse filtering,wiener filtering and least square method makes it difficult to solve the problem of function approximation and hard on applying to complex scenes.In this paper,the main tasks are achieving image super-resolution and image inpainting by using Generative Adversarial Networks.In the preprocessing of image super-resolution reconstruction,some important high-frequency details of the image are lost,the reconstructed image is too smooth and the training of the network is unstable.There are some problems in the process of image inpainting,such as vulnerable to artifacts,insufficient consistency of inpainting results,long repair time and so on.Through the research and experimental analysis of the above problems,the improved method of super-resolution of single image based on coupled generative adversarial networks and the improved progressive image restoration method based on multi-scale feature fusion proposed in this paper improve the quality and efficiency of image restoration to a certain extent.The main contents are as follows:(1)We improve the deep generative image models using a laplacian pyramid of adversarial networks based on the conditional generative adversarial networks,adding the training data set labels as the input conditions in the generative networks and the discriminative networks.Constrain the training of generators and discriminators and the methods of spectral norm regularization are used in discriminative networks.The experimental results show that the generated images are clearer,the training process is stable and the networks converge faster.This part of the study for the follow-up image super-resolution and image inpainting research to play a good theoretical and experimental foundation.(2)In this paper,we propose the method of the super-resolution of single image based on coupled generative adversarial networks,applying the coupled generative adversarial networks to image super-resolution reconstruction after adjustment and improvement;introduce attention-augmented convolution into the generative network,which takes into account the details of each position and its far end in the image;balance the learning ability of the generators and the discriminators to and consider the diversity and quality of the reconstructive image.The relative loss functions are used as the objective functions of the generative adversarial networks to make the structure of the networks more stable.Experimental results show the effectiveness of the proposed method in image super-resolution reconstruction.(3)We proposed a progressive image inpainting method based on multi-scale feature fusion.The image inpainting task is completed by curriculum learning through multi-step.Generative networks improve the structure of encoding-decoding network,fuse image features at different scales and use contextual attention mechanism.The use of global and local discriminators in the discriminative networks ensure the authenticity and integrity of the inpainting images.Finally,the task of image semantic inpainting is realized.The experimental results show that the method has good inpainting results.