Research On Key Technologies Of Image Restoration Based On Deep Learning

At present,there is a relatively perfect video monitoring system.However,because the surveillance cameras are far away from the target and the surveillance videos need to be compressed,the targets in the surveillance videos are partially occluded,which lead to the poor quality of the surveillance video images and the inability to fully obtain the effective information in the image.At present,deep learning has been widely used in the field of image processing and computer vision.In order to improve the quality of image,image superresolution restoration and image de-occlusion based on deep learning has become a research hotspot in the current industry and academia.This thesis focuses on the research of image superresolution and image de-occlusion restoration based on deep learning.The main research contents and achievements are as follows.To solve the problem of super-resolution restoration of small face images in long distance,the method of super-resolution restoration of face image based on improved conditional Wasserstein Generative Adversarial Networks(WGAN)is proposed in this thesis.First of all.a new WGAN method with gradient penalty and weight spectrum normalization is proposed to improve the training stability of WGAN.Secondly.face images are aligned and preprocessed.and the progressive sub-pixel up-sampling structure is constructed,to simplify the superresolution of highly structured small-scale face image.At last.the improved residual dense connection network structure with feature reuse is adopted,and the multi constraint target fusion loss functions of pixel level loss,feature level loss and adversarial loss are constructed to retain the detail features of the image.In Celeba face dataset and CNBC face dataset,large number of experiments show that the method proposed in this thesis can restore the small-scale face image eight times the super-resolution of sampling factor,and get better image restoration results.In order to solve the problem of super-resolution restoration of natural scene images with video compression degradation,a lightweight restoration method of super-resolution natural image based on removing video compression artifacts is proposed.Firstly,the video compression degradation process is divided into two parts:frequency-domain coding compression and down sampling degradation,so the above problem is transformed into two sub problems:video compression artifacts removal and image super-resolution,and the overall framework of the artifacts removal super-resolution restoration method is constructed.Secondly,the improved Artifacts Reduction CNN(AR-CNN)method is used to solve the artifacts removal of video compression,and the network model proposed in Chapter 3 is used to solve the problem of image super-resolution restoration.At last,depthwise separable convolution is used to compress the proposed model to reduce the amount of parameters and calculation of video compression image restoration.In addition,to improve the restoration effect of the model in processing the natural image with uneven data distribution,Instance Normalization is used in the generator.In the YOUKU-VSRE 2019 dataset,the experimental results show that the method proposed in this thesis has a good effect on the super-resolution restoration of natural scene image degraded by video compression.Aiming at the problem of partial occlusion in images,this thesis proposes an inpainting de-occlusion method based on feature fusion.First of all,the inpainting de-occlusion model combining edge and depth features is constructed,in which the problem of lacking edge details or generating edge artifacts in the existing methods is improved,and the quality of restored image is improved.Secondly,in the network model construction,the Autoencoders and WGAN which are outstanding in image generation are combined.Beacause WGAN makes up for the lack of details in the image generated by Autoencoders,and Autoencoders makes up for the deviation in the image generated by WGAN.In the construction of loss function,pixel level loss,feature level loss,image style loss and adversarial loss are integrated to enrich the details of the image.In Celeba-256 datasets and Place datasets,a large number of experiments have proved that the proposed method is advanced and effective.