Video Super-Resolution Based On Deep Recursive Network

With the continuous updating of display devices,the existing low resolution(LR)videos can no longer meet its demand for high resolution(HR),which has become one of the urgent technical challenges to be overcome in the industrial field related to video.Video superresolution(VSR)algorithm is an important technology to solve such problems.In recent years,scholars have proposed various VSR models to solve the LR video degradation problem and achieved impressive performance.However,because the existing methods either do not do alignment of adjacent video frames or explicitly perform image alignment using optical flow methods,the consistency and subjective effect of the reconstructed videos are difficult to meet the practical requirements.In view of the above,this thesis intends to investigate the super-resolution(SR)algorithm for LR images and videos,and propose an improved model for recovering and reconstructing LR video contents.The main research contents and contributions of this thesis are as follows.(1)In image and video SR algorithms,feature extraction is one of the most critical steps.To address the problem that Res Net networks do not differentiate feature importance,which makes the feature extraction capability insufficient,we adopt the Res Ne St network that introduces an attention mechanism and modify the split attention block(SAB)in it.We remove the batch normalization(BN)layer in SAB and design the feature extraction backbone network based on the modified SAB.In addition,jump connections are used between each convolutional layer to better transfer information to the subsequent convolutional layers so that the network can extract rich and effective feature information.2)This thesis proposes a method for image feature alignment using dynamic convolution.Video inter-frame alignment is an important step in VSR.Previous VSR methods usually implement motion estimation(ME)and motion compensation(MC)for adjacent frames to perform inter-frame alignment,and ME usually uses optical flow method,however,it is very difficult to estimate optical flow accurately,and inaccurate optical flow will bring artifacts to the VSR results.To solve the above problem,this thesis proposes a novel end-to-end deep convolutional network that dynamically generates a spatially adaptive filter for alignment,which consists of local pixels in each channel.Our approach avoids explicit MC and instead uses the adaptive filter to implement the alignment operation,which effectively fuses multiframe information and improves the temporal consistency of the video.(3)Integrating the above network model,for the problem of low-resolution images,this thesis first designs an single image super-resolution(SISR)algorithm based on modified SAB.Since the principles between SISR and VSR methods are the same,naturally,this thesis extends the SR algorithm from image to video and designs an end-to-end deep recursive network-based SR model.Given the input of multiple consecutive frames,after image feature extraction,alignment and fusion,the SR reconstruction result of the intermediate frames is output by up-sampling reconstruction after reusing the adjacent frame information.The evaluation results on datasets covering many types videos show that the proposed method outperforms some existing methods in terms of texture details and video consistency.