Enhanced Bidirectional Propagation Network for Video Super-Resolution

Abstract

Video super-resolution (VSR) aims to restore a high-resolution (HR) frame from a given corresponding low-resolution (LR) frame. For most VSR models, the pipeline mainly includes feature extraction module, alignment module, fusion module and reconstruction module, among them, the alignment module and the fusion module are the most important. Based on BasicVSR, we propose an enhanced bidirectional propagation network which is both effective and efficient in exploiting previous and next frames to super-resolve the current frame. In the alignment module, in order to obtain a precisely offsets for alignment, we combine optical flow with deformable convolution (DCN). In the fusion module, we propose a channel and double attention module (CDAM) to extract the important features and suppress the unimportant features in channel dimension in addition, it can capture the long-range dependencies too. The experiments demonstrate the proposed method can effectively achieve superior performance compared with state-of-the-art methods.