基于改进CycleGAN的光学图像迁移生成水下小目标合成孔径声纳图像算法研究  

Optical Image-to-Underwater Small Target Synthetic Aperture Sonar Image Translation Algorithm Based on Improved CycleGAN

作　　者：李宝奇 黄海宁[1,2] 刘纪元[1,2] 李宇 LI Bao-qi;HUANG Hai-ning;LIU Ji-yuan;LI Yu(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：《电子学报》2021年第9期1746-1753,共8页Acta Electronica Sinica

基　　金：国家自然科学基金(No.11904386);国家基础科研计划重大项目(No.JCKY2016206A003);中国科学院青年创新促进会。

摘　　要：针对循环生成对抗网络CycleGAN(Cycle Generative Adversarial Networks)在光学图像迁移生成水下小目标合成孔径声纳图像过程中存在质量差和速度慢的问题,本文提出一种新的特征提取单元SDK(Selective Dilated Kernel),并利用SDK设计了一个新的生成器网络SDKNet.与此同时,提出了一种新的循环一致损失函数MS-CCLF(Multiscale Cyclic Consistent Loss Function),MS-CCLF增加了图像多尺度结构相似性约束.在自建的图像迁移数据集OPT-SAS上,本文SM-CycleGAN(Selective and Multiscale Cycle Generative Adversarial Networks)比原始CycleGAN的图像迁移质量提升4.64%,生成器网络参数降低4.13MB,运算时间减少0.143s.实验结果表明,SM-CycleGAN更适合水下小目标光学图像到合成孔径声纳图像的迁移任务.The original CycleGAN show poor quality and time consuming in optical image to underwater small target synthetic aperture sonar image translation task.To address those problems,a novel convolution building block,SDK(Selective Dilated Kernel),is proposed.By stacking SDK blocks,a generator SDKNet is created.At the same time,Multiscale Cycle Consistent Loss Function(MS-CCLF)is proposed,which add the Multiscale Structural Similarity Index(MS-SSIM)between input images and reconstructed images.On our image translation dataset(OPT-SAS),the classification accuracy of our SM-CycleGAN is4.64%higher than that of original CycleGAN.The generator parameters of SM-CycleGAN is 4.13 MB lower than that of CycleGAN,and the time consuming of SM-CycleGAN is0.143 s less than that of CycleGAN.The experimental results show that SM-CycleGAN is more suitable for the translation task of optical image to small underwater target synthetic aperture sonar image.

关 键 词：光学图像迁移生成合成孔径声纳图像 生成对抗网络 循环生成对抗网络 可选择空洞核网络 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]