DSCVSR: A Lightweight Video Super-Resolution for Arbitrary Magnification

Zixuan Hong, Weipeng Cao, Zhiwu Xu, Zhong Ming, Chuqing Cao, Liang Zheng

Abstract: Video super-resolution, a fundamental task in the field of computer vision, has gained much attention and performance in recent years. However, since deep learning introduces a large number of parameters, which can result in a large resource overhead, the model cannot be deployed on edge devices. Therefore, in this paper, we design a lightweight video super-resolution model, named Depthwise Separable Convolutional Video Super-Resolution (DSCVSR), which utilizes a continuous memory mechanism by constructing a dense depthwise separable convolutional residual block to fuse the deep and shallow feature information in order to enable the network to better learn the details in the video, and also constructs an information-filling module to solve the problem of information loss brought about by the depthwise separable convolution, as well as designing an information-filling module to solve the problem of information loss brought about by the depthwise separable convolution information loss problem caused by deep separable convolution, and a knowledge distillation loss is designed to migrate the knowledge from the teacher’s model to the model to achieve the superscoring results with arbitrary multiplicity. In the experiments, the method is tested on common video datasets, and it is verified that the proposed method can achieve good results with a small number of parameters.