基于GAN的半监督低分辨雷达目标识别算法  

Semi-supervised Low-resolution Radar Target Recognition Technology Based on Generative Adversarial Network

作　　者：朱克凡 王杰贵 吴世俊 ZHU Kefan;WANG Jiegui;WU Shijun(Electronic Countermeasure Institute of National University of Defense Technology,Hefei 230037,China;Unit 96713 of PLA,Shangrao 334100,China)

机构地区：[1]国防科技大学电子对抗学院,安徽合肥230037 [2]中国人民解放军96713部队,江西上饶334100

出　　处：《探测与控制学报》2019年第6期57-63,共7页Journal of Detection & Control

基　　金：国防预研基金项目资助(9140C100404120C1003)

摘　　要：针对雷达侦察过程中数据库标签样本不足导致目标识别率难以提高的问题,提出了基于生成对抗网络(GAN)的半监督低分辨雷达目标识别算法。该算法将现有的少量标签样本和接收到的大量无标签样本作为样本集,使用卷积神经网络(CNN)替代GAN的判别器部分,利用GAN强大的对抗生成能力,提高小标签样本条件下对低分辨雷达目标的分类识别能力。仿真实验结果表明,该算法较传统半监督雷达目标识别方法具有更短的识别时间和更好的识别效果,证明了算法的优越性。Modern radar target recognition usually encountered the problem of receiving a large number of target echo signals,but the recognition rate was difficult to improve due to insufficient label samples.To achieve low-resolution radar target recognition under small training samples,this paper proposed a semi-supervised low-resolution radar target recognition algorithm based on Generative Adversarial Network(GAN).The algorithm used Convolutional Neural Networks(CNN)instead of the discriminator part of GAN,and used a small number of existing label samples and a large number of unlabeled samples as sample sets,which improved the CNN s low-resolution radar target under small training sample conditions.The simulation experiment proved that the GAN-based semi-supervised low-resolution radar target recognition method had shorter recognition time and better recognition effect than the traditional semi-supervised radar target recognition method.

关 键 词：低分辨雷达目标识别 深度学习 生成对抗网络 卷积神经网络 

分 类 号：TN959.1[电子电信—信号与信息处理]