基于BI⁃GM⁃GAMP和U⁃Net的混合电磁反演方法  

A hybrid electromagnetic inversion method based on BI⁃GM⁃GAMP and U⁃Net

作　　者：王芳芳[1,2,3] 周立京 韩婉 汪树典 李开敏 张鑫[1] WANG Fangfang;ZHOU Lijing;HAN Wan;WANG Shudian;LI Kaimin;ZHANG Xin(College of Electronic and Optical Engineering&College of Flexible Electronics(Future Technology),Nanjing University of Posts and Telecommunications,Nanjing 210023,China;National and Local Joint Engineering Laboratory of RF Integration and Micro-Assembly Technology,Nanjing University of Posts and Telecommunications,Nanjing 210023,China;State Key Laboratory of Millimeter Waves,Southeast University,Nanjing 210096,China)

机构地区：[1]南京邮电大学电子与光学工程学院、柔性电子(未来技术)学院,江苏南京210023 [2]南京邮电大学射频集成与微组装技术国家地方联合工程实验室,江苏南京210023 [3]东南大学毫米波国家重点实验室,江苏南京210096

出　　处：《南京邮电大学学报（自然科学版）》2023年第3期28-34,共7页Journal of Nanjing University of Posts and Telecommunications：Natural Science Edition

基　　金：国家自然科学基金(61601245);毫米波国家重点实验室开放课题(K202240);射频集成与微组装技术国家地方联合工程实验室开放课题(KFJJ20210103);江苏省高等学校自然科学研究面上项目(20KJB510008)资助项目。

摘　　要：电磁反演是一种典型的逆散射问题,该问题具有非线性以及不适定性。目前已有一些算法被提出用来解决该问题,但已有算法在效率和精度上无法同时兼顾。为了实时反演高质量图像,文中提出了一种结合玻恩迭代高斯混合广义近似消息传递(Born Iterative Gaussian Mixture Generalized Approximate Message Passing,BI⁃GM⁃GAMP)和卷积神经网络的混合算法,具体为:首先通过BI⁃GM⁃GAMP方法获得初步反演图像,随后利用这些图像训练U⁃Net网络,最后通过训练的U⁃Net网络实现新的电磁反演任务。仿真实验验证了该方法的有效性。Electromagnetic inversion is a typical inverse scattering problem that is nonlinear and ill⁃posed.Some traditional algorithms have been proposed to solve it,but they cannot consider both efficiency and accuracy.Therefore,this paper presents a hybrid method that combines the born iterative Gaussian mixture generalized approximate message passing(BI⁃GM⁃GAMP)algorithm and a convolutional neural network(CNN)to recover high quality images in real⁃time.First,the BI⁃GM⁃GAMP is used to obtain the preliminary images.Second,these images are used to train the U⁃Net.Finally,the new electromagnetic inversion task is realized through the trained U⁃Net network.Simulation results demonstrate the effectiveness of the proposed method.

关 键 词：电磁逆散射 非线性 不适定 卷积神经网络 

分 类 号：TN011[电子电信—物理电子学]