多尺度注意力网络的水下图像增强算法  

Underwater Image Enhancement Algorithm with Multi-scale Attention Networks

作　　者：陈海秀[1,2] 陆康 何珊珊[1] 刘磊[1] 颜秋叙 CHEN Haixiu;LU Kang;HE Shanshan;LIU Lei;YAN Qiuxu(School of Automation,Nanjing University of Information Science&Technology,Nanjing 210044,China;Jiangsu Collaborative Innovation Center of Atmospheric Environment and Equipment Technology(CICAEET),Nanjing University of Information Science&Technology,Nanjing 210044,China)

机构地区：[1]南京信息工程大学自动化学院,南京210044 [2]南京信息工程大学江苏省大气环境与装备技术协同创新中心,南京210044

出　　处：《机械科学与技术》2025年第3期505-512,共8页Mechanical Science and Technology for Aerospace Engineering

基　　金：国家自然科学基金项目(61302189);教育部产学合作协同育人项目(202101159003)。

摘　　要：针对水下环境中的图像存在严重的偏色、模糊等问题,该文提出了一种新型的生成对抗网络。采用U-Net作为生成网络的基础模型并对其进行了改进,首先将注意力机制引入到网络中,并设计多尺度特征提取模块,来提取不同层次的特征。其次通过预处理操作输入白平衡图像提升模型的鲁棒性。为解决单一损失造成图像细节恢复不均匀的问题,在传统的对抗损失函数中联合L1损失与内容损失。实验结果表明:此方法在水下图像的颜色恢复和提高清晰度方面具有很好的效果,其中结构相似度、峰值信噪比、水下彩色质量评估和水下图像质量度量的平均值分别为0.8906、29.0761、0.4454和3.1810。在主观评价和客观评价指标上,综合来说该文算法实验结果均优于对比算法。Aiming at the problems of serious color cast and blurring of images in underwater environment,this paper proposes a novel generative adversarial network algorithm.U-Net is used as the basic model of the generative network and improved.Firstly,the attention mechanism is introduced into the network,and a multi-scale feature extraction module is designed to extract features at different levels.Secondly,the robustness of the model is improved by preprocessing the input white balance image.In order to solve the problem of uneven restoration of image details caused by a single loss,L1 loss and content loss are combined in the traditional adversarial loss function.The experimental results show that this method has a good effect on color recovery and sharpness improvement of underwater images,where the average value of structural similarity,peak signal-to-noise ratio,underwater color quality assessment,and underwater image quality metric is 0.8906,29.0761,0.4454 and 3.1810 respectively.In terms of subjective evaluation and objective evaluation indicators,the experimental results of the algorithm in this paper are better than the comparison algorithms.

关 键 词：水下图像增强 生成对抗网络 注意力机制 多尺度 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]