二阶对齐与频域先验引导的任意倍率视频超分辨

Key Points

• The study aims to enhance spatial resolution in arbitrary-scale video super-resolution (AVSR) while balancing performance, efficiency, and detail recovery.
• Utilized a recurrent neural network framework with a proactive mechanism.
• Incorporated multi-scale frequency priors, flow-based propagation units, second-order deformable alignment units, and super-upsampling units.
• Evaluated performance on the REDS dataset with multiple resolutions.
• Achieved an average PSNR improvement of 0.16 dB compared to representative AVSR methods on the REDS dataset.
• Demonstrated robust generalization capability across datasets in integer and non-integer scale tests on the Vid4 dataset.
• Ablation studies confirmed that second-order deformable alignment and multi-scale frequency priors effectively enhance reconstruction quality in complex motion scenarios.