基于深度卷积神经网络的极化雷达目标识别  

Polarimetric radar target recognition based on depth convolution neural network

作　　者：盖晴晴 韩玉兵[1] 南华[2] 白振东 盛卫星[1] GAI Qingqing;HAN Yubing;NAN Hua;BAI Zhendong;SHENG Weixing(Nanjing University of Science and Technology,Nanjing 210094,China;Beijing Institute of Space Long March Vechicle,Beijing 100076,China)

机构地区：[1]南京理工大学,南京210094 [2]北京航天长征飞行器研究所,北京100076

出　　处：《电波科学学报》2018年第5期575-582,共8页Chinese Journal of Radio Science

基　　金：国家自然科学基金(No.61471196)

摘　　要：针对宽带多极化雷达,提出将高分辨一维距离像(high resolution range profile,HRRP)与极化信息相结合的算法,获得目标在4种极化组态下的一维距离像并将其组成极化距离矩阵.该算法对目标进行全方位的特征抽取与建模,以适应不同的姿态,有助于减少高分辨一维距离像方位敏感性带来的影响.然后提出了直接基于极化距离矩阵、Pauli分解和Freeman分解三种特征提取方式对极化距离矩阵进行目标特征的提取,并将获得的目标特征向量结合起来送入搭建的深度卷积神经网络进行训练学习.该方法不仅结合了不同的特征提取方式以对极化距离矩阵进行更全面的特征提取,而且深度卷积神经网络的运用又对目标特征向量进行了深层学习,仿真结果验证了该方法的有效性.For broadband multi-polarization radar,this paper proposes an algorithm that combines high resolution range profile(HRRP)and polarization information to obtain the HRRP of target under four polarization configurations which is composed of the polarization distance matrix.This algorithm performs full-scale feature extraction and modeling of the target to adapt to different poses and effectively reduces the impact of HRRP azimuth sensitivity.Then,distance matrix,Pauli decomposition and Freeman decomposition are used to extract the target features of the polarization distance matrix,and the obtained target feature vectors are combined and sent to the constructed deep convolutional neural network for training and learning.This method not only combines different feature extraction methods to extract more comprehensive features of the polarization distance matrix,but also deeply learns the target eigenvectors by using the deep convolution neural network.The simulation results verify the effectiveness of the proposed method.

关 键 词：极化距离矩阵 雷达目标识别 Pauli分解 Freeman分解 深度卷积神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]