Video Super-Resolution with Recurrent High and Low-Frequency Information Propagation

The utilization of spatiotemporal complementary information of video is essential for video super-resolution (VSR). However, most VSR methods propagate the time sequence information among video frames through the initial feature, ignoring the entanglement of high and low-frequency components, which limits the performance of VSR. To address these issues, we introduce a novel recurrent propagation framework: VSR with recurrent high and low-frequency information propagation (RHLVR). Specifically, we propose a frequency decomposition module (FDM) to decompose the video sequence into high and low-frequency components, so that the network can obtain greater adaptability. In order to achieve the exploitation of high and low frequency complementary information of video sequence, a recurrent high and low-frequency information propagation unit (RHLP) is proposed to model the high and low-frequency information of the input video sequence. We perform ablation studies to confirm our designs and compare the proposed RHLVR with recent advanced VSR methods. Experimental results prove that RHLVR achieves state-of-the-art VSR performance.