端到端全复数域SAR目标分类神经网络  

End-to-End Full Complex-Valued Domain SAR Target Classification Neural Network

作　　者：方澄 管方恒 李天驰 邹政峰 杨磊 FANG Cheng;GUAN Fang-heng;LI Tian-chi;ZOU Zheng-feng;YANG Lei(College of Electronic Information and Automation,Civil Aviation University of China,Tianjin 300300,China)

机构地区：[1]中国民航大学电子信息与自动化学院,天津300300

出　　处：《电子学报》2024年第7期2449-2460,共12页Acta Electronica Sinica

基　　金：国家自然科学基金(No.62271487);中国民用航空局安全能力建设基金(No.KLZ49420200019,No.KLZ49420200023)~。

摘　　要：合成孔径雷达(Synthetic Aperture Radar,SAR)图像探测经常遇到误差敏感和计算量大的问题,给SAR图像目标识别带来困难,为此研究人员针对SAR数据提出很多新颖高效的深度学习方法,但面向SAR目标识别的深度学习网络大多使用与光学实值处理相同的方法,直接将实值深度神经网络应用于SAR图像.实值的神经网络都不同程度丢失了相位信息,不能充分利用SAR数据复数特性.相位信息是SAR图像独有的数据特征,在SAR干涉测量、信息检索、目标识别等应用中起到至关重要作用.本文为了使网络更适合SAR复数数据特征提取,打破传统神经网络架构,将整体网络端到端复数化,提出一种新型的端到端全复数域多层级神经网络(Complex-valued mUltI-Stage convolutIonal Neural nEtworks,CUISINE)架构,从SAR复数图像数据的输入到卷积计算,再到分类标签,实现全网络复数域下的计算.通过在MSTAR(Moving and Stationary Target Acquisition and Recognition)公开数据集上实验对比表明,本文的方法在SAR数据目标分类上有很好表现,在相位误差为0 rad的测试集上正确率达到99.42%,在相位误差为50 rad的测试集上正确率达到88.05%.Synthetic aperture radar(SAR)image detection often encounters problems such as error sensitivity and high computational complexity,which pose challenges to SAR target recognition.Researchers have proposed many novel and efficient deep learning methods for SAR data.However,most of these deep learning networks for SAR target recogni⁃tion use the same methods as optical real-valued processing,directly applying real-valued deep neural networks to SAR im⁃ages.Real-valued neural networks to some extent lose the phase information,which cannot fully utilize the complex charac⁃teristics of SAR data.As phase information is a unique data feature in SAR images,it plays a crucial role in applications such as SAR interferometry,information retrieval,and target recognition.In order to make the network more suitable for ex⁃tracting complex data features from SAR,breaking the architecture of traditional neural networks,this paper proposes a nov⁃el end-to-end fully complex-valued multi-stage convolutional neural network(Complex-valued mUltI-Stage convolutIonal Neural nEtworks,CUISINE)architecture.It realizes the computation in the full complex-valued domain from the input of SAR complex image data to convolutional calculations,and finally to classification labels.Experimental comparisons on the publicly available MSTAR dataset show that our method performs well in SAR target classification.The accuracy reach⁃es 99.42%on the test set with a phase error of 0 rad,and 88.05%on the test set with a phase error of 50 rad.

关 键 词：合成孔径雷达 相位信息 全复数域 端到端 神经网络 

分 类 号：TN957.52[电子电信—信号与信息处理]