
摘要
大多数视频超分辨率方法依赖于时间滑动窗口内的邻近帧,对单个参考帧进行超分辨率重建,其效率相较于基于循环机制的方法较低。本文提出了一种新型的循环视频超分辨率方法,能够在有效利用历史帧信息的同时,实现高效重建当前帧。该方法将输入图像分解为结构与细节两个分量,并分别送入由多个创新设计的双流结构-细节模块构成的循环单元中进行处理。此外,引入了一种隐藏状态自适应模块,使当前帧能够有选择性地利用隐藏状态中的信息,从而增强对外观变化和误差累积的鲁棒性。大量的消融实验验证了所提模块的有效性。在多个基准数据集上的实验结果表明,与当前最先进的方法相比,所提出的方法在视频超分辨率任务中展现出更优越的性能。
代码仓库
Feynman1999/MgeEditing
GitHub 中提及
junpan19/RSDN
官方
pytorch
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| video-super-resolution-on-msu-super-1 | RSDN + vvenc | BSQ-rate over ERQA: 14.95 BSQ-rate over LPIPS: 4.866 BSQ-rate over MS-SSIM: 9.138 BSQ-rate over PSNR: 14.061 BSQ-rate over VMAF: 10.145 |
| video-super-resolution-on-msu-super-1 | RSDN + x265 | BSQ-rate over ERQA: 13.416 BSQ-rate over LPIPS: 13.232 BSQ-rate over MS-SSIM: 5.682 BSQ-rate over PSNR: 13.403 BSQ-rate over VMAF: 6.467 |
| video-super-resolution-on-msu-super-1 | RSDN + aomenc | BSQ-rate over ERQA: 20.617 BSQ-rate over LPIPS: 14.574 BSQ-rate over MS-SSIM: 11.643 BSQ-rate over PSNR: 15.144 BSQ-rate over VMAF: 10.67 |
| video-super-resolution-on-msu-super-1 | RSDN + x264 | BSQ-rate over ERQA: 6.58 BSQ-rate over LPIPS: 10.775 BSQ-rate over MS-SSIM: 1.023 BSQ-rate over PSNR: 13.348 BSQ-rate over VMAF: 1.5 |
| video-super-resolution-on-msu-super-1 | RSDN + uavs3e | BSQ-rate over ERQA: 18.327 BSQ-rate over LPIPS: 13.844 BSQ-rate over MS-SSIM: 11.643 BSQ-rate over PSNR: 15.144 BSQ-rate over VMAF: 9.796 |
| video-super-resolution-on-msu-vsr-benchmark | RSDN | 1 - LPIPS: 0.819 ERQAv1.0: 0.667 FPS: 1.961 PSNR: 25.321 QRCRv1.0: 0.619 SSIM: 0.826 Subjective score: 5.566 |
| video-super-resolution-on-vid4-4x-upscaling-1 | RSDN | PSNR: 27.92 SSIM: 0.8505 |