A Research On Video Repair Algorithms Based On Deep Learning

This thesis studies the video de-noising,inpainting and texture generation without reference,based on the requirement of intelligent restoration of VHS video.The main contents of this thesis are as follows:First,a blind de-noising model based on inter-frame regression network is proposed to solve the problem that there are no high definition reference samples and the noise distribution is complex.Secondly,aiming at the blurring phenomenon caused by the motion offset between frames,the following measures are proposed in this paper: 1.Inter-frame feature fusion model is applied to deal with the pre-frame and post-frame offsets.2.Inter-frame alignment algorithm based on block matching is adopted to achieve the alignment between frames in low signal-to-noise ratio(SNR).Considering the difficulty caused by the high time complexity of matching algorithm,this paper transforms it through matrix operation and convolution,and deploys it on GPU platform.Aiming at the blurring problem caused by the loss of texture details in the original VHS video,this paper constructs a texture generation model based on adversarial loss and perceptual loss.Relative discriminator and gradient penalty are adopted to make the training process of generating confrontation more stable.Perceptual loss including semantic information is applied to generate more realistic texture.Experiments show that the video quality is improved significantly after restoration.