摘要
针对水下图像出现色偏、模糊、对比度低等退化现象,提出基于动态拓展注意力的多尺度筛选融合水下图像增强网络(MSF-Net)。提出拓展通道注意力模块,增强单一通道表征能力,精细建模通道间关系,提升全局色彩恢复效果。引入四边形注意力机制,采用动态可形变窗口划分策略,捕获更丰富的局部与全局上下文信息,增强复杂场景下多样化对象的纹理细节。结合特征金字塔结构,提出特征筛选融合模块,通过自适应权重分配机制实现多尺度特征的差异化筛选与深度融合,有效保留多尺度关键信息,抑制退化因素,增强融合效果。在UIEB数据集上的实验结果表明,MSF-Net的PSNR与SSIM指标分别提升了1.00 dB与0.005,取得了优秀的量化指标,在色彩、亮度和细节纹理等方面的视觉表现均具有优异的增强效果,增强性能优于现有方法。
Underwater images frequently exhibit degradation including color casts,blur,and low contrast.To address these challenges,we propose MSF-Net:a Multi-scale Screening Fusion Network based on dynamic and extensive attention for underwater image enhancement.The core of MSF-Net includes an Extensive Channel Attention module that enhances individual channel representations and explicitly models inter-channel dependencies,significantly improving global color restoration.We further introduce a Quadrilateral Attention Mechanism,utilizing deformable window partitioning to capture rich local and global contextual information,thereby effectively enhancing texture details across diverse objects in complex underwater scenes.Leveraging a feature pyramid,a Feature Screening and Fusion module employs adaptive weight allocation for differentiated screening and deep fusion of multi-scale features,optimally preserving critical information while suppressing degradation factors.Extensive experiments on the UIEB benchmark demonstrate MSF-Net’s superiority,achieving significant improvements of 1.00 dB in PSNR and 0.005 in SSIM over state-of-the-art methods,confirming exceptional quantitative performance.Qualitative evaluations further reveal superior visual enhancements in color fidelity,illumination uniformity,and fine-grained texture detail.Collectively,MSF-Net surpasses existing approaches in both quantitative metrics and visual quality for underwater image enhancement.
作者
曲元明
魏德宾
袁国豪
潘成胜
闫甜甜
QU Yuanming;WEI Debin;YUAN Guohao;PAN Chengsheng;YAN Tiantian(College of Information Engineering,Dalian University,Dalian 116622,China;Nation Local Joint Engineering Research Center for Communitation and Network Technology,Nanjing University of Posts and Telecommunications,Nanjing 210023,China;Nation and Local Joint Engineering Laboratory of Computer Aided Design,Dalian University,Dalian 116622,China)
出处
《兵器装备工程学报》
北大核心
2026年第2期221-230,278,共11页
Journal of Ordnance Equipment Engineering
基金
辽宁省自然科学基金-博士科研启动项目(2024-BSLH-017)
辽宁省教育厅高校基本科研项目-青年项目(JYTQN2023101)。
关键词
水下图像增强
拓展通道注意力
四边形注意力
多尺度特征融合
混合对比学习正则化
underwater image enhancement
extensive channel attention mechanism
quadrilateral attention mechanism
multi-scale feature fusion
hybrid contrastive learning regularization
