Video Super-resolution Method Based On Deep Learning Prior

Video super-resolution aims to reconstruct high-resolution video from low-resolution video sequences.In recent years,with the continuous development of deep learning,video super-resolution based on deep learning has made outstanding achievements.However,the existing video super-resolution algorithms still have some problems.In this thesis,two issues with existing video super-resolution algorithms are studied,and corresponding solutions are proposed.At the same time,this thesis develops a video super-resolution system.The main contents of this thesis are as follows:(1)Existing back-projection algorithms are not discriminative in the feedback process,and cannot effectively distinguish between useful and useless feature information,which affects the result of video reconstruction.To address the shortcomings of the existing back projection,this thesis proposes a video super-resolution algorithm based on adaptive back projection.Under the constraints of the visual attention model,the feedback process of back-projection can effectively extract useful feature information between adjacent video frames,thereby effectively improving the reconstruction results of video super-resolution.(2)Most existing video super-resolution algorithms are based on ideal assumptions and are difficult to employ in real-world scenarios.To solve the problem of video super-resolution in real-world scenarios,this thesis proposes a blind video super-resolution algorithm based on deep learning priors.By modeling the image degradation process,an effective gradient iteration formula is obtained,in which the gradient learning rate can be learned adaptively through the network.Experimental results show that the gradient formula combined with deep learning can recover more detailed information,which is important for solving the video super-resolution problem in real-world scenarios.(3)This thesis develops a video super-resolution system based on the algorithm proposed in this thesis.The user only needs to select the low-resolution video and super-resolution algorithm to get the corresponding high-resolution video.