Research On Video Super-resolution Reconstruction Technology Based On Deep Neural Networ

With the development of science and technology and the emergence of artificial intelligence technology,people have a higher pursuit of video resolution.Compared with low resolution video sequence,high resolution video sequence often contains more information and usually has better visual experience.Due to the continuous development of deep neural networks in recent years,the video super resolution model built on the basis of deep neural networks can obtain better reconstruction performance and visual effects than the traditional methods.The utilization of spatial and temporal information between video frames,especially time information,is the key to determining the effect of video super resolution reconstruction.However,many networks,especially those based on deep neural networks,are unable to fully process the motion relationship between video frames,resulting in insufficient utilization of video temporal and spatial information,which affects the resolution and fluency of video reconstruction.Therefore,two new implicit motion compensation methods are proposed in this paper to process interframe information of video.On this basis,deep neural network is used to build a suitable network model and carry out in-depth research on video super resolution reconstruction technology,including the following contents:1)In view of the insufficient use of temporal and spatial information,especially time information,of video sequences in the current video super resolution network when processing motion videos,this paper proposes a video super resolution network structure based on D3 D convolution intra-group fusion.Firstly,input sequences are grouped according to different frame rates,which can effectively integrate time information in a hierarchical way.Then,D3 D convolution is used for intra-group feature fusion for each group of videos to preserve more spatio-temporal correlation of video sequences.Finally,time attention mechanism is introduced and integrated with intra-group fusion module to fuse supplementary information provided by each group,so as to recover missing details in video sequences.When training and testing on classic Vid4 standard video data set,The PSNR and SSIM values of the generated high-resolution video frames are 27.39 and 0.8266,respectively,showing superior performance to the current advanced algorithms in terms of visualization effect and time consistency.2)Feature fusion of existing VSR methods is usually carried out in a one-stage manner,and the fused features may have a large deviation from the visual information in the original LR reference frame.In order to take advantage of the rich complementary information of adjacent frames,we propose a multi-stage feature fusion network based on D3 D convolution,which consists of time alignment,auxiliary loss and reallocation modules.The feedforward neural network structure is divided into three stages.In the first stage,D3 D convolution is used to fuse the time alignment features of the support frame and the spatial features of the original reference frame,and the time alignment features of the support frame are generated.In the second stage,an auxiliary loss module is added to make the features obtained in the first stage retain more HR space-related information.In the third stage,the re-alignment module is introduced to make full use of the feature information of the previous stage to improve the visual effect of the reconstructed video.A large number of experiments show that this method has good reconstruction performance on three standard data sets,and the number of parameters can be reduced to a certain extent.