基于场景先验学习的视频超分辨率重建算法  

Video Super-Resolution Reconstruction Algorithm Based on Scene Prior Learning

作　　者：汤乐 赵必美 蔡卓骏 郑博仑 TANG Le;ZHAO Bimei;CAI Zhuojun;ZHENG Bolun(Hangzhou Dianzi University,Hangzhou 310018,China;Southern Power Grid Artificial Intelligence Technology Co.,Ltd.,Guangzhou 510555,China)

机构地区：[1]杭州电子科技大学,浙江杭州310018 [2]南方电网人工智能科技有限公司,广东广州510555

出　　处：《软件工程》2025年第3期11-17,共7页Software Engineering

基　　金：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)。

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0.03~0.09 dB。实验结果证明,SPL模块可以显著提升VSR方法的性能。Most existing Video Super-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 Signal-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.

关 键 词：视频超分辨率 场景先验学习 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]