基于混合注意力模块改进StarGAN的水下图像增强  

Enhancement of underwater images on improved StarGAN by mixed attention module

作　　者：郑浩君 王振 张佳鹏 刘胜男 钱程 涂雪滢 刘世晶[2,4] ZHENG Haojun;WANG Zhen;ZHANG Jiapeng;LIU Shengnan;QIAN Cheng;TU Xueying;LIU Shijing(School of Navigation and Naval Architecture,Dalian Ocean University,Dalian 116023,China;Fishery Machinery and Instrument Research Institute,Chinese Academy of Fishery Sciences,Shanghai 200092,China;College of Marine Sciences,Fujian Agriculture and Forestry University,Fuzhou 350002,China;Sanya Oceanographic Institution,Ocean University of China,Sanya 572011,China)

机构地区：[1]大连海洋大学航海与船舶工程学院,辽宁大连116023 [2]中国水产科学研究院渔业机械仪器研究所,上海200092 [3]福建农林大学海洋学院,福建福州350002 [4]中国海洋大学三亚海洋研究院,海南三亚572011

出　　处：《南方水产科学》2025年第1期185-196,共12页South China Fisheries Science

基　　金：海南省种业实验室资助项目(B23H10004);中国水产科学研究院中央级公益性科研院所基本科研业务费专项资金资助(2024XT0901)。

摘　　要：围绕水下图像色偏和模糊的特点,针对不同浑浊度的水下图像差异较大问题,提出了一种基于混合注意力模块(Convolutional block attention module,CBAM)改进的星型生成对抗网络(Star generative adversarial networks,StarGAN)用于水下多浑浊图像增强。首先使用水下相机采集实验室和养殖平台环境2组水下多浊度图像数据集;其次优化StarGAN,在每个ResidualBlock模块后引入一个由通道注意力模块和空间注意力模块串联组成的CBAM;最后进行消融实验,并与其他方法比较,使用水下图像质量评估(Underwater image quality measurement,UIQM)、水下彩色图像质量评估(Underwater color image quality evaluation,UCIQE)和图像熵作为图像质量评价指标。结果表明,实验室数据集增强后,UIQM达到1.18,UCIQE达到30.13,图像熵达到12.83;养殖平台数据集增强后,UIQM达到0.52,UCIQE达到10.35,图像熵达到9.94。该方法对实验室和养殖平台环境中不同浑浊度的图像增强均有较好的效果,在消融实验及与其他方法的比较中,该方法的得分均为最高。Based on the characteristics of color cast and blur in underwater images,we proposed a StarGAN(Star generative adversarial networks)based on CBAM(Convolutional block attention module)improvement for the underwater multi turbi-dity image enhancement to address the problem of significant differences in underwater images with different turbidity levels.First,we collected two sets of underwater turbidity image datasets from laboratory and aquaculture platform environments by using an underwater camera.Secondly,we optimized StarGAN by introducing a CBAM consisting of a channel attention mo-dule and a spatial attention module in series after each ResidualBlock module.Finally,we conducted ablation experiments and compared them with other methods by using UIQM(Underwater color image quality measurement),UCIQE(Underwater color image quality evaluation)and Image entropy as image quality evaluation indicators.The results show that UIQM reached 1.18,UCIQE reached 30.13 and Image entropy reached 12.83 of the enhanced laboratory dataset.UIQM reached 0.52,UCIQE reached 10.35 and Image entropy reached 9.94 of the enhanced aquaculture platform dataset.The experimental results indicate that in ablation experiments and compared with the other methods,this method has a good effect on enhancing multi turbidity images,with the highest scores.

关 键 词：多浑浊度图像 水下图像增强 注意力模块 星型生成对抗网络 

分 类 号：S951.2[农业科学—水产养殖]