Local-Global Dynamic Filtering Network for Video Super-Resolution

Abstract

Video super-resolution (VSR) has been greatly advanced by the use of deep learning techniques, but the challenge of handling motion variability has remained a bottleneck. Many previous methods have treated motions equally, leading to suboptimal alignment. In this paper, we propose a Local-Global Dynamic Filtering Network (LGDFNet) to address this issue. LGDFNet uses a divide-and-conquer strategy to handle motionvarying features, where the overall feature is split into local features and assigned specialized sub-networks to align and fuse them from local to global. To align the features and adaptively aggregate several kernels for calibration, we propose the SelfCalibrated Dynamic Filtering (SCDF) module. Additionally, we introduce the Cross-Attention Feature Fusing (CAFF) module to capture long-range dependencies and fuse each feature. Our extensive experiments on different benchmark datasets demonstrate the effectiveness of LGDFNet, both subjectively and objectively.