Deep Compressed Video Super-Resolution With Guidance of Coding Priors

Compressed video super-resolution (VSR) is employed to generate high-resolution (HR) videos from low-resolution (LR) compressed videos. Recently, some compressed VSR methods have adopted coding priors, such as partition maps, compressed residual frames, predictive pictures and motion vectors, to generate HR videos. However, these methods disregard the design of modules according to the specific characteristics of coding information, which limits the application efficiency of coding priors. In this paper, we propose a deep compressed VSR network that effectively introduces coding priors to construct high-quality HR videos. Specifically, we design a partition-guided feature extraction module to extract features from the LR video with the guidance of the partition average image. Moreover, we separate the video features into sparse features and dense features according to the energy distribution of the compressed residual frame to achieve feature enhancement. Additionally, we construct a temporal attention-based feature fusion module to use motion vectors and predictive pictures to eliminate motion errors between frames and temporally fuse features. Based on these modules, the coding priors are effectively employed in our model for constructing high-quality HR videos. The experimental results demonstrate that our method achieves better performance and lower complexity than the state-of-the-arts.