| Upscaling low resolution videos to the high space is called video super-resolution,which is a popular research direction in the computer vision.With the development of Artificial Intelligence,researchers begin to pay attention to video super-resolution(VSR)based on deep learning.However,the existing methods only focus on several integer scale factors.Arbitrary magnification is only considered in single image super-resolution and cannot be directly applied to VSR.Because they ignore feature with different scale factors or the multiple inputs information.For the blank of arbitrary video super-resolution,this paper proposes a method that upscales video to arbitrary scale factors,and only a single model will be trained.In order to make full use of information in different scale factors,Scale-Aware Reconstruction Module(SARM)is proposed.This module constructs a position projection relation matrix between low-resolution and high-resolution frames according to scale factors.It uses a kind of attention mechanism to fuse feature extracted from the vector in the matrix,and the feature will be used as additional information in reconstruction module.Position Relation and Image Information Based Meta Upscale Module(PRII-MUM)is proposed to take full advantage of the inputs in video super resolution.The module is composed of position weights prediction branch and feature weights prediction branch.It takes the position relation between input and output frames and aligned features as inputs to dynamically predict the weights of upsampling filter,that is various in different frames.This paper compared the proposed method with many classical image and video super-resolution methods on general datasets of VSR,and shows the quantitative results and upsampled images of multiple different upsampling scales.The experimental results show that the proposed method is better than existing methods in most scale factors,which verify the effectiveness of the proposed method. |
