基于生成对抗网络的SAR图像去噪  

SAR image denoising based on generative adversarial networks

作　　者：刘帅奇[1] 雷钰 庞姣[1] 赵淑欢 苏永钢 孙晨阳 LIU Shuaiqi;LEI Yu;PANG Jiao;ZHAO Shuhuan;SU Yonggang;SUN Chenyang(Machine Vision Technology Innovation Center of Hebei Province,College of Electronic and Informational Engineering,Hebei University,Baoding 071002,China)

机构地区：[1]河北大学电子信息工程学院,河北省机器视觉技术创新中心,河北保定071002

出　　处：《河北大学学报（自然科学版）》2022年第3期306-313,共8页Journal of Hebei University(Natural Science Edition)

基　　金：国家自然科学基金资助项目(62172139);河北大学研究生创新资助项目(HBU2021ss002);河北省自然科学基金资助项目(F2020201025);河北省高等学校科学技术研究资助项目(BJ2020030);国家重点实验室开放课题基金资助项目(20220007);河北大学校长基金资助项目(XZJJ201909)。

摘　　要：合成孔径雷达(synthetic aperture radar,SAR)图像是一种能够全天时、全天候产生高分辨率图像的主动式对地观测系统,在农业和军事等方面得到了广泛应用.然而,由于相干成像机制受到相干斑噪声的影响,因此提出了一种基于生成式对抗网络的SAR图像盲去噪算法,构造了基于残差结构的深度卷积神经网络(deep convolutional neural network,DCNN)作为生成网络,可以加速训练过程,提高去噪性能.本文还利用峰值信噪比(peak signal to noise ratio,PSNR)和结构相似指数(structural similarity index measure,SSIM)定义一种新的损失函数,使得去噪后的图像更符合人眼的视觉感知要求.实验结果表明,本文算法可以有效地抑制SAR图像中的相干噪声,获得良好的去噪效果.Synthetic aperture radar(SAR)is a kind of active earth-observation system,which can produce high-resolution image all day and it has been widely used in agriculture and military.However,SAR image is affected by speckle due to the coherent imaging mechanism.In this paper,we propose a SAR image denoising algorithm based on generative adversarial network.We constructed the generated network by the structure of deep convolutional neural network(DCNN)with residual structure.The new generated network can accelerate the training process and improve the denoising performance.We also define a loss function by combining PSNR and SSIM,which makes the denoising result of SAR noisy image more in line with the visual perception requirements of human eyes.The experiment results show that the proposed algorithm can suppress speckle in SAR images and get good denoising result.

关 键 词：SAR图像去噪 生成对抗网络 深度卷积神经网络 损失函数 

分 类 号：TN95[电子电信—信号与信息处理]