| 摘 要: 现有的大多数视频超分辨率(Video Super-Resolution,VSR)方法只关注临近几帧间的信息,忽视了长时间段内场景中包含的先验信息。针对这一问题,提出了一种场景先验学习(Scene Prior Learning,SPL)模块。SPL模块是一个轻量级的插件模块,可以在模型推理阶段显式地对长时间段内场景中包含的先验信息进行建模,并利用先验信息增强原始网络中的注意力机制。将SPL模块集成到不同 VSR方法中后,其PSNR(峰值信噪比)提升了0.03~0.09dB。实验结果证明,SPL模块可以显著提升 VSR方法的性能。 |
| 关键词: 视频超分辨率;场景先验学习;注意力机制 |
|
中图分类号: TP391
文献标识码: A
|
| 基金项目: 浙江省重点研发计划(2023C01044);国家自然科学基金面上项目(62371175);浙江省属高校基本科研业务费(GK239909299001-013) |
|
| Video Super-Resolution Reconstruction Algorithm Based on Scene Prior Learning |
|
TANG Le1
, ZHAO Bimei2
, CAI Zhuojun1
, ZHENG Bolun1
|
(1.Hangzhou Dianzi University, Hangzhou 310018, China;
2.Southern Power Grid Artificial Intelligence Technology Co., Ltd., Guangzhou 510555, China)
tangle@hdu.edu.cn; zhaobm@csg.cn; caizj1@csg.cn; blzheng@hdu.edu.cn
|
| Abstract: Most existing Video Supe-r Resolution (VSR) methods focus only on the information within the next few frames, ignoring the prior information contained in the scene over a long period. To solve this problem, a Scene Prior Learning (SPL) module is proposed. The SPL module is a lightweight plug-in module that can explicitly model the prior information contained in the scene over a long period during the model inference phase, and use the prior information to enhance the attention mechanism in the original network. After integrating the SPL module into different VSR methods, the PSNR (Peak Signa-l to-Noise Ratio) improves by 0.03 to 0.09 dB. Experimental results demonstrate that the SPL module can significantly enhance the performance of VSR methods. |
| Keywords: Video Supe-r Resolution (VSR); Scene Prior Learning (SPL); attention mechanism |
