Research On Video Super Resolution Technology Based On Convolutional Neural Network

Video applications have appeared in all aspects of work and life in recent years,such as high-resolution television and online live.However,it is difficult for video systems to provide sufficient real-time high-resolution video sources due to various limitations of processing power,storage capacity and transmission bandwidth.In order to obtain highquality visual content and enhance visual experience,video super-resolution technology has gradually become an important research direction.Video super-resolution is to obtain a corresponding high-resolution video frame from the given low-resolution video frames by increasing the pixel density and recovering the details which lost during the imaging process.Some existing video super-resolution methods use the current low-resolution video frame and its adjacent low-resolution video frames to achieve super-resolution of the current frame.However,both the current frame and the adjacent frames are low-resolution images which lack high-frequency information with detail texture,the high-resolution video frame on this basis is not accurate enough,and even over-smoothing.Further,the degradation method of video frames is often more complicated in actual applications,and the quality of low-resolution video frames are worse,resulting in that the video super-resolution methods may get bad results in super resolution progress.Focus on the above two problems,the related research work as follows:1.Combined with the characteristics of video frame image and convolutional neural network,a new video super-resolution method based on mixed resolution model is proposed in this thesis.Retain some frames as high-quality frames,which remain ground truth image,the input of the network model is the current low-resolution frame and the corresponding high-quality frame,no longer continuous low-resolution frames.Highquality frames provide prior knowledge of high-frequency detail to improve the network model's ability to recover edge details.At the same time,residual learning module is introduced to realize the skipped connection of different network layer features and increase the nonlinear mapping ability of the network model.In this thesis,the effects of mixed resolution model,residual learning module and motion module on network model performance are analyzed through multiple sets of comparison experiments.The experimental results show that the proposed video super-resolution method based on the mixed resolution model can reproduce high-quality content in low-resolution frames and demonstrates better performance over the other methods.2.Most of the existing super-resolution methods only consider bicubic downsampling as the degradation method.Once the degradation mode is more complicated,the super-resolution quality will go down.For this problem,based on the mixed resolution model,a video super-resolution method for multiple degradation factors is proposed in this thesis.In addition to bicubic downsampling,two common degradation factors,blur and noise,are used to make low-resolution video,which is closer to the video in the actual application.In this thesis,the degradation factors are transformed into the degradation feature map,which provides a prior knowledge of the degradation mode,and further improves the estimation model of the degradation feature map to adapt to the video frames with unknown degradation mode.Through experimental comparison,the proposed method can better deal with the video super-resolution problem of multiple degradation factors over the other methods.