一种基于散射特征增强的SAR目标电磁仿真图像质量提升方法  

A Method of Improving the Quality of SAR Target Electromagnetic Simulation Image Based on Scattering Feature Enhancement

作　　者：张祥辉 封斯嘉 马晓杰 张思乾 孙浩[1] 计科峰[1] 陈珲 ZHANG Xianghui;FENG Sijia;MA Xiaojie;ZHANG Siqian;SUN Hao;JI Kefeng;CHEN Hui(State Key Laboratory of Complex Electromagnetic Environment Effects on Electronics and Information System,College of Electronic Science and Technology,National University of Defense Technology,Changsha,Hunan 410073,China;School of Information Science and Engineering,State Key Laboratory of Millimeter Wave,Southeast University,Nanjing,Jiangsu 210096,China)

机构地区：[1]国防科技大学电子科学学院电子信息系统复杂电磁环境效应国家重点实验室,湖南长沙410073 [2]东南大学信息科学与工程学院、毫米波国家重点实验室,江苏南京210096

出　　处：《信号处理》2023年第9期1573-1586,共14页Journal of Signal Processing

基　　金：国家自然科学基金(62001480,61971426);湖南省自然科学基金(2021JJ40684)。

摘　　要：现阶段深度学习算法在对合成孔径雷达(Synthetic Aperture Radar,SAR)目标识别时,通常面临着实测数据部分样本缺失的情况,利用电磁仿真数据进行辅助识别是有效途径之一。然而,仿真和实测数据存在不可避免的差异,现有仿真图像质量提升方法更关注仿真和实测图像整体风格的相似性,忽略了面向识别的目标散射特征的重要性。针对上述问题,本文提出了一种基于散射特征增强的SAR目标电磁仿真图像质量提升方法。该方法在循环生成对抗网络(Cycle Generative Adversarial Networks,CycleGAN)框架下,改进损失函数,一方面使用最小二乘损失函数替代交叉熵损失函数,避免了梯度消失,实现对目标纹理结构特征的迭代优化;另一方面引入MS-SSIM-L1损失函数,更好地保留生成图像的细节信息和结构轮廓,保持目标整体结构一致性,同时有效避免模型的过度学习。基于4类车辆目标仿真数据集和MSTAR实测数据集,利用目标轮廓、阴影轮廓和目标强度特征相似度指标,验证了本文方法增强了目标细节纹理和结构轮廓等散射特征。在此基础上,结合A-ConvNets网络开展了目标分类识别实验,相较于原始CycleGAN方法,本文方法在不同样本缺失条件下均提高了识别准确率。通过特征可视化,表明生成图像更接近实测图像的目标特征分布,验证了本文方法的有效性。At this stage,the deep learning algorithm usually faces the situation of partial sample missing of the measured data when recognizing the synthetic aperture radar(SAR)target.The use of electromagnetic simulation data for auxiliary recognition is one of the effective ways.However,there are inevitable differences between the simulated and measured data.The existing methods for improving the quality of the simulated image pay more attention to the similarity of the over⁃all style between the simulated and measured data,and ignore the importance of the target scattering characteristics for rec⁃ognition.To solve these problems,this paper proposed a method to improve the quality of SAR target electromagnetic simulation image based on scattering feature enhancement.This method improved the loss function under the framework of cycle generation adversarial networks(CycleGAN).On the one hand,the least squares loss function was used to replace the cross entropy loss function to avoid the gradient disappearing and realize the iterative optimization of the target texture structure features;On the other hand,the MS-SSIM-L1 loss function was introduced to better retain the detail information and structure outline of the generated image,maintain the consistency of the overall structure of the target,and effectively avoid over-learning of the model.Based on four types of vehicle target simulation data set and MSTAR measured data set,by using the target contour,shadow contour and target intensity feature similarity index,it was verified that the method in this paper enhanced the scattering features such as target detail texture and structure contour.On this basis,the target classification and recognition experiment was carried out in combination with A-ConvNets network.Compared with original CycleGAN method,method in this paper improved the recognition accuracy under different sample missing conditions.Through feature visualization,it was shown that the generated image was closer to the target feature distribution of the mea�

关 键 词：合成孔径雷达图像 循环生成对抗网络 仿真图像 质量提升 目标识别 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]