Research On Image Restoration Based On Generative Adversarial Network

Since the beginning of the 21 st century,due to the rapid development of computers and the continuous popularization of image processing tools,image inpainting technology has attracted more and more attention.In the context of computer vision and graphics,image inpainting refers to restoring pixels in missing regions in the corrupted images,and it plays an irreplaceable role in computer vision tasks.Image inpainting technology is not only one of the basic technologies in the field of computer vision,but also one of the more widely used technologies at present.The image inpainting task can also be regarded as a special image generation task,which generates the missing pixels of the image to make the image complete and realistic.The generative adversarial network model has a powerful ability to generate content,and researchers have applied it to image inpainting tasks and achieved exciting results.However,the generative adversarial network model still has problems such as model instability and long training time,which leads to semantically inconsistent sample content in the generated images.And the model cannot obtain sufficient deep semantic information.In addition,common image inpainting methods often do not consider the spatial information and semantic information of the image at the same time.Therefore,in view of the above shortcomings,this thesis mainly studies the image inpainting task based on generative adversarial networks,and the specific work is as follows:(1)Aiming at the problem of model instability and insufficient access to deep semantic information of image,an image inpainting method based on the combination of generative adversarial network and residual pyramid is proposed.Firstly,a multi-layer residual pyramid generator is introduced into the generated adversarial network to solve the precision reduction caused by the increase of the depth of the neural network.At the same time,instance normalization is used to accelerate the training speed of the model.Secondly,the dual discriminator is used to solve the deficiency of semantic inconsistency between the global region and the local regions of the image,and the problems such as discontinuity and blur at the boundary of the missing region of the image are eliminated.Finally,the pyramid L1 loss and the average adversarial loss are used to optimize the model.Experiments on the four datasets of face,texture,architecture and scene show that this method not only outperforms the comparison method in terms of experimental data,but also can generate the inpainted image with more consistent semantics with the original image.(2)Aiming at the inadequacy that the model cannot simultaneously consider the image spatial information and semantic information,puts forward a kind of based on feature fusion gating adversary image inpainting method.Firstly,two different gated feature fusion modules are proposed,namely channel feature fusion module and spatial feature fusion module.Among them,the channel feature fusion module is used to obtain the semantic information of different channel feature maps;the spatial feature fusion module is used to obtain more spatial and location information.Secondly,a new multi-scale discriminator is proposed to identify the global consistency of generated images,which consists of four identical discriminators to judge the consistency between images with different resolutions and their corresponding real images respectively.Finally,the pyramid L1 loss,the perception loss and the average adversarial loss are used to optimize the model.Experiments on face,texture,architecture and scene datasets show that the proposed method is superior to popular inpainting methods in both experimental data and visualization results,and has strong generalization.