Multi-Reference-Based Cross-Scale Feature Fusion for Compressed Video Super Resolution

To save transmission bandwidth, there exists an approach to down-sample a video and then up-sample the compressed video to save bit rates. The existing Super Resolution (SR) methods generally design powerful networks to compensate the loss information introduced by down-sampling. But the information of entire video is not fully utilized and effectively fused, resulting in the learned context information that is not enough for the high quality reconstruction. We propose a multi high-quality frames Referenced Cross-scale compressed Video Super Resolution method (RCVSR) that wisely uses past-and-future information, to pursue higher compression efficiency. Specifically, a joint reference motion alignment module is proposed. Low resolution (LR) frame after up-sampling is separately aligned with past-and-future reference frames to preserve more spatial details; at the same time this LR frame is aligned with neighborhood frames to get continuous motion information and similar contents. Then, a reference based refinement module is applied to compensate motion and lost texture details by computing similarity matrix across channel dimensions. Finally, an attention guided dual-branch residual module is employed to enhance the reconstructed result concurrently. Compared with the HEVC anchor, the average gain of Bjontegaard Delta Rate (BD-Rate) under the Low-Delay-P (LDP) setting is 24.86%. In addition, an experimental comparison is made with the advanced SR methods and compressed video quality enhancement (VQE) methods, and the superior efficiency and generalization of the proposed algorithm are further reported.