Reasearch On Image Restoration Method Based On Ganerative Adversarial Networks

Image degradation can be caused by the shaking of imageing equipment,noise of components and image coding and decoding.Image restoration is to deal with the degraded image to make it closer to the original image.Traditional image restoration algorithms has many disadvantages,such as long calculation time,poor robustness and so on.Recent years,neural network-based restoration method has been gradually proposed.This kind of algorithm can fully acquire the prior knowledge of the image and effectively improved the image processing speed,so it has been widely concerned by scholars.However,most existing algorithms solve the problem of relatively single image degradation.For example,in aspect of image deblurring,only artificial blur kernel is solved,but this kind of blur kernel still has a big gap compared with the blur kernel of real image;in aspect of image super-resolution reconstruction,most of algorithms only solve the reconstruction problem of clear image,and do not consider the image reconstruction including other degradation problems.Therefore,this paper focuses on the more complex problem of image restoration.Firstly,this paper studies the image restoration of natural blurred image.In order to be closer to the real situation,this article uses GOPRO?Large dataset for training,which contains more realistic motion blur pictures.In terms of network design,we apply dense connected convolutional networks(Densenets)to the field of image deblurring,so that the network can make full use of the information in the middle layer.We design a loss function based on perceptual loss which is suitable for deblurring application..In addition,the algorithm alse uses the general adversarial network(GAN),so that the generated image is closer to the real image in vison.The PSNR(peak signal to noise ratio)and SSIM(structural similarity)between the generated image and the clear image and processing time are tested to verify the performance of the algotirhm.The results show that this algorithm can achieve good recovery effects on GOPRO?Large test set and Kohler test set,which means it can effectively recover the details lost due to motion blur,so that the algorithm is robust.In addition,its processing time of the algorithm is alse significantly shorter than the conparision algorithm.In addition to the blurring of the image,there are various reasons for the reduction of image clarity and loss of detailed information.Therefore,when the image has both blur and low definition problems,it needs to joint deblur and super-resolution reconstruction algorithm to restore the image.However,simply concatenation two algorithms is not only inefficient,but alse has poor recovery effect.Therefore,it is very important to solve the problem of super-resolution reconstruction of blurred image.In response to this kind of problem,we downsample the corresponding positions of the clear and blurred images in GOPRO?Large dataset,and forms a group of blurred low-resolution images,clear low-resolution images,and clear high-resolution images to obtain a new dataset for experimentation.In order to avoid the limitation to the network function and the generation of artifacts caused by the normalization,the algorithm does not use any form of normalization.In order to enable the network to avoid the large increase of parameters with the increase of depth,we combines the dense connected network and the residual network to form the dense residual network.In order to make the network extract deblurring features and super-resolution reconstruction features,we adopt a double-branch structure,and combine deblurring loss and super-resolution reconstruction loss to optimization network,and by comparinf the PSNR and SSIM between the restored image and the clear image with related algorithms,the performance and effectiveness of the algorithm in this paper are verified.