Research On Video Super Resolution Based On Motion Compensation

Video super-resolution technology is a technology that processes low-resolution videos with rough details to obtain corresponding high-resolution videos with better visual quality and detailed information.Motion compensation technology is a commonly used method for processing low-resolution video frames in video super-resolution technology.Its purpose is to help super-resolution technology make better use of similar areas in adjacent video frames to obtain better video quality.The way to effectively extract and use the information of adjacent frames in lowresolution video is the focus of video super-resolution technology,and it is also a research hotspot at home and abroad.To solve this problem,this paper explored technologies with the help of deep learning and convolutional neural network,and our research works are described below :1.The main technical route and research history of video super-resolution are analyzed.With the help of a large amount of papers,combined with the convolutional neural network that has been the most concentrated in the field of image processing in recent years,a multi-frame fusion network is introduced to extract feature from adjacent frames.The multi-frame fusion network boosted the performance of the video super-resolution network.2.In view of the characteristics that large-scale objects and small-scale objects often coexist in video frames,a multi-scale residual network is proposed.This network can enhance the performance of the video super-resolution network for large-scale objects and small-scale objects,and improve the overall performance of the super-resolution network.3.In view of the problem that the conventional video super-resolution method uses simple convolutional operations at the initial feature extraction stage,which leads to a single network receptive field and cannot extract multi-scale features well,a multi-branch initial perception module is proposed to enhance the capability of image feature initial extraction in super-resolution neural network.4.In view of the phenomenon that the motion estimation method is not applicable to the adjacent frames of the video with a large time span,which affects the super-resolution performance,a progressive motion compensation strategy is proposed to effectively reduce the motion compensation error.5.In view of the phenomenon that convolutional neural networks can hardly extract redundant information of similar targets with large spatial spans,a non-local neural network is introduced in the initial feature extraction stage to achieve efficient spatial feature extraction.6.In view of the case where traditional convolutional neural networks simply cascade the building blocks,making the model have poor multi-scale mapping capabilities and a single receptive field,a multi-scale network architecture is proposed as a composition framework in the network.Without changing the structure of the building block,the multi-scale mapping capability is given to the neural network,which further improves the performance of the video super-resolution network.Based on the techniques raised above,this paper constructs an end-to-end convolutional neural network.Experiments show that the video super-resolution performance of the overall neural network reaches the current advanced level.