基于多尺度空间自适应注意力网络的轻量级图像超分辨率方法  

Lightweight Image Super-Resolution Reconstruction Method Based on Multi-scale Spatial Adaptive Attention Network

作　　者：黄峰 刘鸿伟 沈英[1] 裘兆炳 陈丽琼 HUANG Feng;LIU Hongwei;SHEN Ying;QIU Zhaobing;CHEN Liqiong(School of Mechanical Engineering and Automation,Fuzhou University,Fuzhou 350108)

机构地区：[1]福州大学机械工程及自动化学院,福州350108

出　　处：《模式识别与人工智能》2025年第1期36-50,共15页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金青年基金项目(No.62405060);福建省自然科学基金项目(No.2022J05113,2024J01245);福建省中青年教师教育科研项目(No.JAT210035)资助。

摘　　要：针对现有图像超分辨率重建方法存在模型复杂度过高和参数量过大等问题,文中提出基于多尺度空间自适应注意力网络(Multi-scale Spatial Adaptive Attention Network,MSAAN)的轻量级图像超分辨率重建方法.首先,设计全局特征调制模块(Global Feature Modulation Module,GFM),学习全局纹理特征.同时,设计轻量级的多尺度特征聚合模块(Multi-scale Feature Aggregation Module,MFA),自适应聚合局部至全局的高频空间特征.然后,融合GFM和MFA,提出多尺度空间自适应注意力模块(Multi-scale Spatial Adaptive Attention Module,MSAA).最后,通过特征交互门控前馈模块(Feature Interactive Gated Feed-Forward Module,FIGFF)增强局部信息提取能力,同时减少通道冗余.大量实验表明,MSAAN能捕捉更全面、更精细的特征,在保证轻量化的同时显著提升图像的重建效果.To address the challenges of high model complexity and excessive parameter counts in existing image super-resolution(SR)reconstruction methods,a lightweight image SR reconstruction method based on multi-scale spatial adaptive attention network(MSAAN)is proposed.First,a global feature modulation module(GFM)is designed to learn global texture features.Additionally,a lightweight multi-scale feature aggregation module(MFA)is introduced to adaptively aggregate high-frequency spatial features from local to global scales.Second,the multi-scale spatial adaptive attention module(MSAA)is proposed by integrating GFM and MFA.Finally,a feature interactive gated feed-forward module(FIGFF)is incorporated to enhance the local feature extraction capability while reducing the channel redundancy.Extensive experiments demonstrate that MSAAN effectively captures more comprehensive and refined features,significantly improving reconstruction quality while maintaining a lightweight structure.

关 键 词：卷积神经网络 Transformer 轻量级图像超分辨率重建 多尺度空间自适应注意力 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]