基于生成对抗网络的SAR图像降噪模型  

SAR Image Noise Reduction Model Based on GAN

作　　者：解涛 郭建胜[1] 张晓丰[1] 顾涛勇 赵博欣 XIE Tao;GUO Jiansheng;ZHANG Xiaofeng;GU Taoyong;ZHAO Boxin(Air Force Engineering University,Xi'an 710000 China)

机构地区：[1]空军工程大学,西安710000

出　　处：《电光与控制》2022年第9期48-52,共5页Electronics Optics & Control

基　　金：国家自然科学基金(61703428)。

摘　　要：合成孔径雷达(SAR)是相干成像系统,生成的图像经常会被相干斑噪声污染,继而造成了SAR图像在后续分割、识别中准确率低的问题。针对图像被污染问题,设计了一种结合生成对抗网络(GAN)与残差网络(ResNet)的SAR图像降噪网络模型Re-GAN,其中,GAN中的生成器加入了ResNet中的残差块以增强对SAR图像降噪的能力,模型中的组合损失函数在降噪时可以更好地保留图像细节。在MATAR数据集上,Re-GAN分别与BM3D算法、小波降噪算法进行比较,实验结果证明,Re-GAN在视觉效果和定量分析方面都具有良好的性能。Synthetic Aperture Radar(SAR) is a coherent imaging system.The images generated from SAR are often polluted by speckle noise,which results in low accuracy in subsequent segmentation and recognition of SAR images.Aming at the problem of image pollution,a SAR image denoising network model Re-GAN is designed,which combines Generative Adversarial Network(GAN) with the Residual Network(ResNet).The residual block of ResNet is added into the generator of GAN to enhance the ability of SAR image denoising.The combined loss function in the model can better preserve image details during noise reduction.In MATAR data set,Re-GAN is compared with BM3 D and wavelet denoising algorithm respectively.The experimental results show that Re-GAN has good performance in both visual effect and quantitative analysis.

关 键 词：SAR 图像降噪 生成对抗网络 残差网络 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]