D2SE-CNN:改进的SAR图像相干斑抑制算法  

D2SE-CNN:An improved SAR image despeckling network

作　　者：张一铭 赵生福 郑鑫 王艺博 丁辉[1,2] ZHANG Yiming;ZHAO Shengfu;ZHENG Xing;WANG Yibo;DING Hui(College of Information Engineering,Capital Normal University,Beijing 100048,China;Beijing Engineering Research Center of Highly Reliable Embedded System,Beijing 100048,China)

机构地区：[1]首都师范大学信息工程学院,北京100048 [2]高可靠嵌入式系统技术北京市工程研究中心,北京100048

出　　处：《兵器装备工程学报》2022年第11期103-111,共9页Journal of Ordnance Equipment Engineering

基　　金：国家自然科学基金项目(61876112)。

摘　　要：合成孔径雷达(SAR)的相干成像时,由于存在相干斑噪声,导致图像细节模糊,影响SAR图像的解译等后续应用。结合注意力机制,提出一种改进的下采样卷积神经网络D2SE-CNN。该方法在ID-CNN模型的基础上,去除估计噪声的残差连接;引入下采样,使原图重新排列成四个子图,扩大感受野;并添加挤压与激励块(SE)注意力模块,从而实现相干斑的抑制。为了验证算法的有效性,在BSDS500及NWPUVHR-10数据集和真实SAR图像上与主流方法进行了比较,实验结果表明,所提模型在PSNR、SSIM、ENL、Cv多个评价指标上得到较好的提升。Synthetic Aperture Radar(SAR)is a coherent imaging system and as such it strongly suffers from the presence of speckles.The presence of speckles degraded the image quality and makes SAR images difficult to interpret,such as image segmentation,detection,and recognition.To achieve promising results in removing noise from real-world images,we proposed an improved deep learning-based method called D2SE-CNN,despeckling convolutional neural networks combine with down sampling and squeeze-and-excitation(SE)block.The proposed D2SE-CNN works on down sampled subimages,rearranged into four subimages to enlarge the receptive field,and combine with the SE block,achieving a better denoising performance.Compared with the current mainstream method,our experimental results show that the proposed network achieved more competitive performances on the BSDS500,NWPU VHR-10 datasets,and real SAR images in PSNR,SSIM,metrics.

关 键 词：SAR图像 卷积神经网络 相干斑抑制 挤压与激励块 图像质量增强 

分 类 号：TJ02[兵器科学与技术—兵器发射理论与技术] TP75[自动化与计算机技术—检测技术与自动化装置]