双重对偶神经网络用于超分辨率任务研究  

Research on dual dual neural networks for super-resolution tasks

作　　者：孙宏伟 廖铃 SUN Hongwei;LIAO Ling(School of Artificial Intelligence,Chongqing Three Gorges Vocational College,ChongQing,China 404155)

机构地区：[1]重庆三峡职业学院人工智能学院,重庆404155

出　　处：《深圳信息职业技术学院学报》2025年第1期43-48,55,共7页Journal of Shenzhen Institute of Information Technology

基　　金：重庆市教育委员会科学技术研究项目(项目编号:KJQN202403514)。

摘　　要：最近的单图像超分辨率研究,大多数模型存在低分辨率图像到高分辨率图像映射关系不准确和泛化能力弱的问题,难以保证稳定的重建效果。针对此问题,提出一种基于对偶回归神经网络的网络架构。这种方案除了构建一个从低分辨率到高分辨率的原始映射之外,还加入了对偶回归算法来学习估计重建低分辨率图像的过程。对偶回归作为一种反馈机制,通过双重任务形成的回路提供了自我校验能力,优化了从低分辨率到高分辨率的映射精度。这种设计减小了模型的映射空间,使得模型更容易学习到准确的高分辨率特征。为了生成更加逼真的纹理,在原始映射中加入多尺度连接模块和重建模块,来融合全局和局部特征,解决长期依赖问题。经过理论和实验证明,所提出方案比基准模型有更加优秀的泛化能力,能避免伪影,重构出清晰的边缘和纹理。In recent single image super-resolution research tasks,most models have problems with inaccurate mapping relationships from low resolution images to high-resolution images and weak generalization ability,making it difficult to ensure stable reconstruction results.A network architecture based on dual regression neural network is proposed to address this issue.This approach not only constructs a raw mapping from low resolution to high resolution,but also incorporates a dual regression algorithm to learn and estimate the process of reconstructing low resolution images.Dual regression,as a feedback mechanism,provides self validation capability through a loop formed by dual tasks,optimizing the mapping accuracy from low resolution to high resolution.This design reduces the mapping space of the model,making it easier for the model to learn accurate high-resolution features.In order to generate more realistic textures,multi-scale connection modules and reconstruction modules are added to the original mapping to fuse global and local features and solve long-term dependency problems.After theoretical and experimental verification,the proposed scheme has better generalization ability than the benchmark model,can avoid artifacts,and reconstruct clear edges and textures.

关 键 词：图像超分辨率 对偶神经网络 多尺度跳跃连接 重建模块 回归网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]