English

应用科学学报 ›› 2022, Vol. 40 ›› Issue (6): 973-983.doi: 10.3969/j.issn.0255-8297.2022.06.008

• 信号与信息处理 • 上一篇    

基于时频分析与深度学习的高分辨距离像雷达目标识别

聂江华1, 肖永生1,2, 黄丽贞1, 贺丰收3   

  1. 1. 南昌航空大学 信息工程学院, 江西 南昌 330063;
    2. 南京航空航天大学 电子信息工程学院, 江苏 南京 211106;
    3. 中国航空工业集团公司 雷华电子技术研究所, 江苏 无锡 214063
  • 收稿日期:2021-08-12 发布日期:2022-12-03
  • 通信作者: 肖永生，副教授，研究方向为图像处理、雷达自动目标识别。E-mail:xysfly@nchu.edu.cn E-mail:xysfly@nchu.edu.cn
  • 基金资助:
    国家自然科学基金(No.61661035,No.62261040);江西省自然科学基金(No.20192BAB207001);航空科学基金(No.201920056001,No.20200020056001);江西省研究生创新专项资金(No.YC2020-S519)资助

High Resolution Range Profile Radar Target Recognition Based on Time-Frequency Analysis and Deep Learning

NIE Jianghua1, XIAO Yongsheng1,2, HUANG Lizhen1, HE Fengshou3   

  1. 1. College of Information Engineering, Nanchang Hangkong University, Nanchang 330063, Jiangxi, China;
    2. College of Electronic and Information Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 211106, Jiangsu, China;
    3. Leihua Electronic Technology Research Institute, Aviation Industry Corporation of China, Ltd., Wuxi 214063, Jiangsu, China
  • Received:2021-08-12 Published:2022-12-03

PDF

121

摘要： 传统的基于高分辨距离像（high resolution range profile,HRRP）雷达目标识别方法易受噪声影响，为此提出一种基于时频分析与深度学习的HRRP雷达目标识别方法。首先使用生成模型对低信噪比的HRRP数据进行处理，以提高数据的信噪比，生成模型采用深度卷积生成对抗网络（deep convolutional generative adversarial network,DCGAN）和所提出的约束朴素最小二乘生成对抗网络（constrained naive leasts quares generative adversarial network,CN-LSGAN）；其次将处理后的数据分别进行短时傅里叶变换（short-time Fourier transform,STFT）和小波变换（wavelet transform,WT），得到二维时频数据；最后利用卷积神经网络（convolutional neuralnetwork,CNN）进行识别。实验结果表明，CN-LSGAN相对DCGAN能够更好地提高信噪比，WT相比STFT得到的数据更能获取HRRP特征信息，因而基于CN-LSGAN,WT与CNN的HRRP雷达目标识别方法具有更好的识别效果。

关键词: 雷达目标识别, 高分辨距离像, 约束朴素最小二乘生成对抗网络, 深度卷积生成对抗网络, 时频分析

Abstract: In view of the problem that radar target recognition methods based on traditional high resolution range profile (HRRP) are susceptible to noise, an HRRP radar target recognition method employing time-frequency analysis and deep learning is proposed. First, low signal-to-noise ratio HRRP data is processed, and gains an improved signal-to-noise ratio by using a generative model which uses deep convolutional generative adversarial network (DCGAN) and constrained naive least squares generative adversarial network (CNLSGAN) proposed in this paper. Second, the processed data is processed with short-time Fourier transform (STFT) and wavelet transform (wavelet transform, WT) respectively to obtain two-dimensional time-frequency data. Finally, the obtained two-dimensional data is recognized by convolutional neural network (CNN). Experimental results show that the proposed CN-LSGAN performs better in improving signal-to-noise ratio compared to DCGAN, and WT can obtain HRRP feature information more efficiently than STFT. Therefore, the HRRP radar target recognition method based on CN-LSGAN, WT and CNN has higher recognition ability.

Key words: radar target recognition, high resolution range profile (HRRP), constrained naive least squares generative adversarial network (CN-LSGAN), deep convolutional generative adversarial network (DCGAN), time-frequency analysis

中图分类号: