摘要
该文利用深度学习的高维特征泛化学习能力,将卷积神经网络(CNN)用于海上目标微多普勒的检测和分类。首先,在海面微动目标模型的基础上,在实测海杂波背景中分别构建4种类型微动信号的2维时频图,并作为训练和测试数据集;然后,分别采用LeNet, AlexNet和GoogLeNet 3种CNN模型进行二元检测和多种微动类型分类,并进行比较,研究信杂比对检测和分类性能的影响。最后,与传统的支持向量机方法进行比较,结果表明,所提方法能够智能学习微动特征,具有更好的检测和分类性能,可为杂波背景下的雷达动目标检测和识别提供新的技术途径。
In this paper, Convolutional Neural Networks (CNN) are used to detect and classify micro-Doppler effects of maritime targets by using generalized learning ability for high-dimensional features. Based on the micro-motion model of maritime targets, two-dimensional time-frequency maps of four types of micro-motion signals are constructed in the measured sea clutter background. These maps were used as training and test datasets. Furthermore, three types of CNN models, i.e., LeNet, AlexNet, and GoogleNet, are used in binary detection and multiple micro-motion classifications. The effects of signal-to-noise ratio on detection and classification performance are also studied. Compared with the traditional support vector machine method, the proposed method can learn the micro-motion features intelligently, and has performed better in detection and classification. Thus, this study can provide a new technical approach for radar target detection and recognition under a cluttered background.
作者
苏宁远
陈小龙
关键
牟效乾
刘宁波
Su Ningyuan;Chen Xiaolong;Guan Jian;Mou Xiaoqian;Liu Ningbo(Naval Aviation University,Yantai 264001,China)
出处
《雷达学报(中英文)》
CSCD
北大核心
2018年第5期565-574,共10页
Journal of Radars
基金
国家自然科学基金(61871391
61501487
61871392
U1633122
61471382
61531020)
国防科技基金(2102024)
山东省高校科研发展计划(J17KB139)
"泰山学者"和中国科协"青年人才托举工程"(YESS20160115)专项经费~~
关键词
微多普勒
雷达目标检测
深度学习
卷积神经网络(CNN)
海杂波
时频分析
Micro-Doppler
Radar target detection
Deep learning
Convolutional Neural Network (CNN)
SeaClutter
Time-frequency analysis