Video Super-Resolution Reconstruction Based On Time-Domain Deformable Convolution
With the development of technology in recent years,video resources have increased dramatical-ly.In addition to the demand for video resources themselves,the demand for improving video clarity has also gradually increased.Aiming at the problem that the information between reference frames and supporting frames(adjacent frames)is not sufficiently utilized in current video super-resolution reconstruction methods,this paper proposes an algorithm that uses deformable convolution to extract features at different time intervals to complete super-resolution reconstruction.The algorithm divides the video split image frames into different time domain groups,extracts the features of each time domain group by using deformable convolution,fuses the features of each time domain group with separable 3D convolution,and finally fuses the features of each time domain group with spatiotemporal attention module,and improves the resolution through subpixel convo-lution.According to the experimental results on the open data set of video super-resolution reconstruction,the algorithm proposed in this paper is superior to other video reconstruction algorithms such as VSRnet and VESPCN,and can effectively improve the definition of reconstructed video.
video super-resolutiondeep learningdeformable convolution3D convolutionspatiotempo-ral attention mechanism
NETL
NSTL
万方数据
