基于改进CycleGAN的浑浊水体图像增强算法研究  

Turbid Water Image Enhancement Algorithm Based on Improved CycleGAN

作　　者：李宝奇 黄海宁[1,2] 刘纪元[1,2] 刘正君 韦琳哲 LI Baoqi;HUANG Haining;LIU Jiyuan;LIU Zhengjun;WEI Linzhe(Institute of Acoustics,Chinese Academy of Sciences,Beijing 100190,China;Key Laboratory of Science and Technology on Advanced Underwater Acoustic Signal Processing,Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]中国科学院声学研究所,北京100190 [2]中国科学院先进水下信息技术重点实验室,北京100190

出　　处：《电子与信息学报》2022年第7期2504-2511,共8页Journal of Electronics & Information Technology

基　　金：国家自然科学基金(11904386);国家基础科研计划重大项目(JCKY2016206A003);中国科学院青年创新促进会(2019023)。

摘　　要：针对循环生成对抗网络(Cycle Generative Adversarial Networks,CycleGAN)在浑浊水体图像增强中存在质量差和速度慢的问题,该文提出一种可扩展、可选择和轻量化的特征提取单元BSDK(Bottleneck Selective Dilated Kernel),并利用BSDK设计了一个新的生成器网络BSDKNet。与此同时,提出一种多尺度损失函数MLF(Multi-scale Loss Function)。在自建的浑浊水体图像增强数据集TC(Turbid and Clear)上,该文BM-CycleGAN比原始CycleGAN的精度提升3.27%,生成器网络参数降低4.15MB,运算时间减少0.107s。实验结果表明BMCycleGAN适合浑浊水体图像增强任务。In order to solve the problem of poor quality and slow speed in turbid water image enhancement based on Cycle Generative Adversarial Networks(CycleGAN),a scalable,selective and efficient block Bottleneck Selective Dilated Kernel(BSDK)is proposed,and a new generator network BSDKNet is redesigned by stacking BSDK.At the same time,Multi-scale Loss Function(MLF)is proposed to improve the structural similarity of the clear water image and the generated clear water image.On our turbid water image enhancement dataset Turbid and Clear(TC),the classification accuracy of the proposed BM-CycleGAN is 3.27%higher than that of classical CycleGAN.The generator parameters of BM-CycleGAN is 4.15 MB lower than that of CycleGAN,and the time consuming of BM-CycleGAN is 0.107 s less than that of CycleGAN.The experimental results show that BM-CycleGAN is suitable for turbid water image enhancement.

关 键 词：图像增强 生成对抗网络 循环生成对抗网络 深度可分离空洞卷积 多尺度结构相似性 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]