Multi-Directional Information Channel for Compressed Video Super-Resolution

Abstract

In the last few years, the compressed video super-resolution(CVSR) method has emerged that downsampling videos and then compressing them to meet limited bandwidth requirements. Among them, high-quality recovery and resolution recovery for low resolution compressed videos are key to success. How to use temporal and spatial information is essential for the quality recovery of low resolution compressed videos. The existing methods place more emphasis on designing powerful networks to fuse the spatiotemporal information of videos, leading to the problem of one-way information propagation, which undermines the spatiotemporal information of the entire video and results in a lack of close relationship between the learned contextual information and the target frame. In this paper, we integrate the multi-directional information channel framework into CVSR, utilizing past, present, and future information to fully utilize the spatiotemporal information of the entire video. This is the work of the first multi-directional information channel framework on CVSR. Our experiment demonstrates the practicality of the multi-directional information channel framework on CVSR.