系统工程与电子技术 ›› 2022, Vol. 44 ›› Issue (1): 108-116.doi: 10.12305/j.issn.1001-506X.2022.01.15

• 传感器与信号处理 • 上一篇    下一篇

基于卷积神经网络的海面目标全极化高分辨距离像识别技术

但波*, 付哲泉, 高山, 简涛   

  1. 海军航空大学, 山东 烟台 264001
  • 收稿日期:2020-10-16 出版日期:2022-01-01 发布日期:2022-01-19
  • 通讯作者: 但波
  • 作者简介:但波(1985—), 男, 讲师, 博士, 主要研究方向为人工智能与机器学习、精确制导与目标识别|付哲泉(1992—), 男, 博士研究生, 主要研究方向为精确制导技术及其智能化|高山(1978—), 男, 副教授, 硕士,主要研究方向为目标识别与选择技术|简涛(1980—), 男, 教授, 博士,主要研究方向为雷达信号处理与目标识别
  • 基金资助:
    国家自然科学基金(61971432);国家自然科学基金(61790551);山东省泰山学者工程(tsqn201909156);国防科技项目基金(2019-JCJQ-JJ-060)

Full-polarization high resolution range profile recognition technology for sea surface target based on convolutional neural network

Bo DAN*, Zhequan FU, Shan GAO, Tao JIAN   

  1. Naval Avaition University, Yantai 264001, China
  • Received:2020-10-16 Online:2022-01-01 Published:2022-01-19
  • Contact: Bo DAN

摘要:

针对雷达目标全极化高分辨距离像(high resolution range profile, HRRP)提取可分性特征时, 利用全部距离单元作为度量尺度无法保留各距离单元具体特征的问题, 在综合利用4个极化通道的舰船目标HRRP信息时选择单个距离单元作为度量尺度。在此基础上, 提出基于Pauli分解, HαAα1分解和结构相似性参数的特征提取方法对目标极化散射矩阵进行特征提取, 并将提取得到的特征与基于卷积神经网络(convolutional neural network, CNN)的舰船目标HRRP识别方法结合, 利用改进残差结构CNN从极化特征中进一步提取深层可分性特征进行目标识别。实验结果表明, 所提方法能够保留目标全极化HRRP更多特征, 提高目标识别的准确率。

关键词: 卷积神经网络, 全极化高分辨距离像, 可分性特征, Pauli分解

Abstract:

Aiming at the problem that when extracting separability features from radar target full polarization high resolution range profile (HRRP), using all range units as the measurement scale cannot retain the specific characteristics of each range unit, a single range unit is selected as the measurement scale when comprehensively using the ship target HRRP information of four polarization channels. On this basis, a feature extraction method based on Pauli decomposition, HαAα1 decomposition and structural similarity parameters is proposed to extract the features of the target polarization scattering matrix, and the extracted features are combined with the ship target HRRP recognition method based on convolutional neural network (CNN). The improved residual structure CNN is used to further extract deep separability features from polarization features for target recognition. Experimental results show that the proposed method can retain more features of target full polarization HRRP and improve the accuracy of target recognition.

Key words: convolutional neural network (CNN), full-polarization high resolution range profile (HRRP), separable characteristics, Pauli decomposition

中图分类号: