Core Concepts
The authors propose a collaborative feedback discriminative (CFD) propagation method to effectively explore spatio-temporal information and reduce the influence of artifacts caused by inaccurate feature alignment for video super-resolution.
Abstract
The key insights and highlights of the content are:
Existing video super-resolution (VSR) methods often suffer from inaccurate feature alignment, which leads to artifacts that accumulate during the propagation process and degrade the final video restoration quality.
To address this issue, the authors propose a discriminative alignment correction (DAC) module that adaptively calibrates the inaccurate aligned features using shallow features to suppress the influence of artifacts.
Furthermore, the authors develop a collaborative feedback propagation (CFP) module that leverages feedback and gating mechanisms to jointly propagate different timestep features from forward and backward branches, enabling better exploration of long-range spatio-temporal information.
The proposed DAC and CFP modules are integrated into existing VSR backbones, including BasicVSR, BasicVSR++, and PSRT, resulting in three new models: CFD-BasicVSR, CFD-BasicVSR++, and CFD-PSRT.
Extensive experiments on benchmark datasets demonstrate that the proposed CFD propagation method can significantly improve the performance of existing VSR models while maintaining a lower model complexity and computational cost.
Stats
The authors use the following key metrics and figures to support their approach:
"The runtime is the average inference time on 100 LR video frames with a size of 180×320 resolution."
"Circle sizes indicate the number of parameters."