Research And Application Of Video Super-resolution Algorithm Based On Deep Learnin

Video is one of the most common multimedia in daily life and is an important way for people to access,express and disseminate information.With the development of various information technologies such as mobile devices and multimedia,the demand for video quality has been increasing.Therefore,super-resolution reconstruction of low-resolution videos has become very important.The goal of video super-resolution technology is to reconstruct low-resolution video into sharper,high-resolution video,which is used to improve the picture quality of video.This technology is widely used in many fields such as entertainment,video surveillance and medical image analysis.With the development of deep learning technology,a large number of video super-resolution algorithms based on deep learning have been proposed,and although great improvements in performance have been achieved,there are still many problems.For example,it is not possible to make full use of adjacent frame information and the problem of poor video reconstruction time consistency.Meanwhile,some low-resolution videos have low frame rates,and frame rate is an important factor affecting video picture quality.However,most of the video super-resolution algorithms only target the spatial resolution of the video,and the algorithms cannot enhance the temporal and spatial resolution of the video at the same time.In this paper,we conduct an in-depth study of the above problems and propose targeted solutions.The main research work includes:(1)To address the problems of existing video super-resolution methods that cannot fully utilize the adjacent frame information and poor temporal consistency of video reconstruction results,the paper proposes a non-local deformable aligned frame recursive progressive fusion network that uses non-local operations to align sequential frame features,and then applies recursion to temporally model the hidden information and aligned features at the previous moment,so that better temporal consistency can be obtained.The recursive progressive fusion unit is used to fully fuse the hidden information from the previous moment and the currently aligned features,so that the temporal information in the adjacent frames can be fully utilized,resulting in higher super-resolution reconstruction results.(2)To address the problem that the simple video super-resolution method can only improve the spatial resolution of low-resolution video but not its temporal resolution,and the existing spatio-temporal video super-resolution method does not fully consider the spatio-temporal correlation between consecutive video frames,which makes the video frame reconstruction results unsatisfactory,this paper proposes a spatio-temporal video super-resolution method based on multi-scale feature interpolation and temporal feature fusion(MSITF).First,feature interpolation is performed in the low-resolution feature space to obtain features corresponding to missing frames.Secondly,deformable convolutional sampling is used to learn enhanced missing frame features to obtain more accurate missing frame features.Finally,temporal alignment and global context learning of sequence frame features are performed by a temporal feature fusion module to fully extract and utilize useful spatio-temporal information in adjacent frames,resulting in better quality of the reconstructed video frames.(3)Based on the above proposed algorithm,this thesis designs and implements a video super-resolution Web system.The system has functional modules such as login/registration,video super-resolution,reconstructed video display and message notification,etc.The user logs into the system through the web terminal and uploads the video for super-resolution reconstruction,and the original video and the reconstructed video can be displayed against each other after the video is reconstructed,and users will be notified by email to view and download the video.