系统工程与电子技术 ›› 2020, Vol. 42 ›› Issue (11): 2426-2433.doi: 10.3969/j.issn.1001-506X.2020.11.03

• 电子技术 • 上一篇    下一篇

基于DCNN的弹道中段目标HRRP图像识别

向前(), 王晓丹(), 李睿(), 来杰(), 张国令()   

  1. 空军工程大学防空反导学院, 陕西 西安 710051
  • 收稿日期:2020-03-06 出版日期:2020-11-01 发布日期:2020-11-05
  • 作者简介:向前(1995-),男,硕士研究生,主要研究方向为智能信息处理。E-mail:qianxljp@126.com|王晓丹(1966-),女,教授,博士,主要研究方向为智能信息处理。E-mail:afeu_wang@163.com|李睿(1992-),男,博士研究生,主要研究方向为智能信息处理。E-mail:lazy136200@163.com|来杰(1994-),男,博士研究生,主要研究方向为智能信息处理。E-mail:531418867@qq.com|张国令(1995-),男,硕士研究生,主要研究方向为智能信息处理。E-mail:823721228@qq.com
  • 基金资助:
    国家自然科学基金(61876189);国家自然科学基金(61503407);国家自然科学基金(61703426);国家自然科学基金(61806219)

HRRP image recognition of midcourse ballistic targets based on DCNN

Qian XIANG(), Xiaodan WANG(), Rui LI(), Jie LAI(), Guoling ZHANG()   

  1. Air and Missile Defense College, Air Force Engineering University, Xi'an 710051, China
  • Received:2020-03-06 Online:2020-11-01 Published:2020-11-05

摘要:

针对弹道中段目标识别问题,现有的基于高分辨距离像(high resolution range profile, HRRP)的识别方法直接提取一维HRRP(1-dimension HRRP, 1D-HRRP)的整体特征,对局部特征提取能力较弱,且由1D-HRRP数据提取的特征的表达能力有限,为此提出了一种基于深度卷积神经网络(deep convolutional neural network, DCNN)的弹道中段目标HRRP图像识别方法。首先,将1D-HRRP转化为0-1二值图像,从而把数值变化特征转化为图像结构特征;然后,构建DCNN逐层提取图像的局部特征和共性特征并进行识别;最后,结合Dropout和L2正则化双重机制缓解DCNN过拟合问题,利用AdaBound算法提高DCNN训练收敛速度和识别正确率。实验结果表明,所提出的基于HRRP图像的弹道中段目标识别方法比其他12种基于1D-HRRP或二维HRRP(2-dimension HRRP, 2D-HRRP)的识别方法正确率更高,在测试数据集上达到了96.28%,实验结果验证了该方法的有效性。

关键词: 弹道导弹, 目标识别, 高分辨距离像, 深度卷积神经网络, AdaBound算法

Abstract:

To solve the problem of midcourse ballistic target recognition, the existing recognition methods based on high resolution range profile (HRRP) directly extract the overall features of 1-dimension HRRP (1D-HRRP), which has a weak ability to extract local features, and the expression ability of features extracted from 1D-HRRP is limited. Therefore, a recognition method for midcourse ballistic targets of HRRP images based on the deep convolutional neural network (DCNN) is proposed. Firstly, 1D-HRRP is transformed into 0-1 binary images, so that the numerical variation features are transformed into the graph structure features. Then, DCNN is constructed to extract the local and common features of HRRP images layer by layer and recognize them. Finally, Dropout and L2 regularization are combined to alleviate the over-fitting problem of DCNN, and AdaBound algorithm is used to improve the convergence speed and accuracy of DCNN training. The experimental results show that the recognition method based on HRRP images is more accurate than other 12 recognition methods based on 1D-HRRP or 2-dimension HRRP (2D-HRRP), and achieves an accuracy of 96.28% in the test dataset, verifying the effectiveness of the proposed method.

Key words: ballistic missile, target recognition, high resolution range profile (HRRP), deep convolutional neural network (DCNN), AdaBound algorithm

中图分类号: