Blueprint separable convolution for video super-resolution
Video super-resolution reconstruction technology has always faced the following two issues:it is diffi-cult to achieve content alignment during the alignment of adjacent frames,which is especially likely to occur during the super-resolution reconstruction of video frames with large motion.Secondly,in the process of video super-resolution reconstruction,motion blur and the fusion of multiple motions are challenging to handle effectively.In re-sponse to these two issues,we propose BSCVSR,a high-performance and targeted video super-resolution recon-struction technology algorithm.BSCVSR introduces a pyramid cascade blueprint separable convolution alignment network and a spatiotemporal attention fusion SR network,which aligns video frames at different levels based on dif-ferent information and then addresses the fusion problem of multiple motions and motion blur by incorporating spati-otemporal attention mechanisms.In the alignment network,we introduce blueprint separable convolutions to reduce redundancy.Through research,it is found that the internal correlations within a kernel have a more pronounced and direct effect on the separable operation compared to cross-kernel correlations,which is well supported by the final experimental data.From the final experimental data,we can see that BSCVSR provides a reference solution for video super-resolution reconstruction technology in terms of both performance and model lightweighting.
video super-resolution reconstructionneural networkblueprint separable convolutionspatiotempo-ral attention mechanism
国家科技期刊平台
NSTL
万方数据
