Research And Implementation Of Short Video Super-resolution Algorithm Based On Deep Learning

With the continuous improvement of network infrastructure and the rapid popularization of mobile devices,the short video industry has developed rapidly in recent years.Watching short videos has become one of the main entertainment methods for people to enjoy online leisure.Since short videos have the characteristics of hand-shot shooting,low production threshold,and simple creation,and users lack professional equipment and shooting knowledge,they are susceptible to factors such as hand shaking,object movement,and lighting environment during the shooting process.As a result,short videos are far from professional videos in terms of resolution,clarity,brightness,and contrast,which affects the viewing experience of users.Therefore,it has important practical significance and academic value for the research of short video super-resolution technology.Based on the above application background,this thesis proposes a short video super-resolution deep neural network for fusion image denoising.Specifically,this thesis designs a short video image denoising model based on the feature phased fusion strategy,which uses multi-scale feature map progressive fusion and feature reconstruction to provide detailed and high-quality video images for subsequent super-resolution.In addition,in the super-resolution stage,a deformable three-dimensional convolution is used to design a feature alignment fusion module that captures spatiotemporal information and adaptive motion compensation.Secondly,in order to improve the ability of the network model to learn related features,this thesis proposes a self-learning method that combines channel attention and spatial attention,using adaptive size local cross-channel interaction strategies and separable convolution to capture the potential dependencies of in the feature channels and feature space.Then this thesis also designs a residual dense connection module based on attention mechanism.Futhermore,in order to verify the effectiveness of the short video super-resolution algorithm,we also construct a short video super-resolution dataset.Finally,we design a prototype system for short video super-resolution conversion.We compare the algorithm on short video dataset and public datasets respectively,and the algorithm has achieved good performance.In addition,we also conduct an experimental analysis on the proposed short video image denoising model and attention method,and then verify the effectiveness of the proposed model for short video super-resolution tasks.