采用CNN-SSD的雷达HRRP小样本目标识别方法  

Radar HRRP based few-shot target recognition with CNN-SSD

作　　者：郭泽坤 田隆 韩宁 王鹏辉[1] 刘宏伟[1] 陈渤[1] GUO Zekun;TIAN Long;HAN Ning;WANG Penghui;LIU Hongwei;CHEN Bo(National Laboratory of Radar Signal Processing,Xidian University,Xi’an 710071,China;Unit 32181 of PLA,Xi’an 710032,China)

机构地区：[1]西安电子科技大学雷达信号处理国家重点实验室,陕西西安710071 [2]中国人民解放军32181部队,陕西西安710032

出　　处：《西安电子科技大学学报》2021年第2期7-14,共8页Journal of Xidian University

基　　金：国家杰出青年科学基金(61525105);国家自然科学基金(61771361);高等学校学科创新引智计划(111计划)(B18039)。

摘　　要：雷达高分辨距离像非合作目标识别技术的发展主要受限于两个方面:一是由于非合作目标观测频率极低,导致带标签样本量严重不足,使非合作目标识别成为典型的小样本识别问题,这在学界依然是一个没有定论的开放性的热点和难点问题;二是现有的目标识别方法多基于完备数据集假设,使得其与非合作目标小样本目标识别问题严重失配。针对上述问题,对于非合作目标识别抛开数据集完备假设,提出了一种采用卷积神经网络模型连续自蒸馏的雷达高分辨距离像小样本目标识别方法。该方法首先利用包含45类合作目标的完备的训练数据集训练,得到一个初始的类别无关的特征提取器;基于此,进一步采用模型连续自蒸馏机制得到更具泛化能力的特征提取器;最后,在非合作目标上对所提取特征的泛化能力进行了测试。实验结果表明,对于5类非合作目标,所提方法在仅有1个、5个和10个训练样本的情况下,平均识别率分别达到61.26%,84.69%和92.52%,实现了对库外样本的快速、有效识别。The development of radar high resolution range profile(HRRP)non-cooperative targets recognition technology is mainly limited by two aspects:(1)Due to the low observation frequency of non-cooperative targets,the number of labeled HRRPs is insufficient,making non-cooperative HRRP based target recognition a typical few-shot recognition problem,which is still a hot and difficult issue without definite conclusion in the academia.(2)The existing HRRP based target recognition methods are mostly based on the hypothesis of complete dataset,making them mismatch with non-cooperative target recognition in few-shot setting.In this paper,we put aside the complete hypothesis and propose an HRRP based few-shot target recognition method with CNN-SSD.The proposed method first uses a complete training HRRP containing 45 classes of cooperative targets to learn an initial category-independent feature extractor,on the basis of which we further utilize the model sequential self-distillation mechanism to obtain a more generalized feature extractor.Finally,the generalization ability of the extracted features is evaluated on unseen non-cooperative targets during training.Experimental results on self-simulated HRRP dataset reveal that the proposed method can achieve an average recognition rates of 61.26%,84.69% and 92.52% respectively when only 1,5 and 10 annotated HRRPs of non-cooperative targets are available.

关 键 词：雷达目标识别 小样本学习 特征提取 高分辨距离像 卷积神经网络 连续自蒸馏 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TN959.17[自动化与计算机技术—控制科学与工程]