电子学报 ›› 2022, Vol. 50 ›› Issue (5): 1042-1049.DOI: 10.12263/DZXB.20211255

• 学术论文 • 上一篇    下一篇

基于混合型复数域卷积神经网络的三维转动舰船目标识别

张云, 化青龙, 姜义成, 徐丹   

  1. 哈尔滨工业大学电子与信息工程学院,黑龙江 哈尔滨 150001
  • 收稿日期:2021-09-13 修回日期:2022-02-20 出版日期:2022-05-25 发布日期:2022-06-18
  • 通讯作者: 姜义成
  • 作者简介:张 云 女,1975年11月出生于黑龙江省虎林市.现为哈尔滨工业大学教授、博士生导师.主要研究方向为雷达信号处理、SAR成像、机器学习和遥感模式分析.E-mail: zhangyunhit@hit.edu.cn
    化青龙 男,1995年2月出生于安徽省阜阳市.现为哈尔滨工业大学博士.主要研究方向为雷达图像处理和深度学习网络框架.E-mail: huaqinglong_hit@163.com
    姜义成 男,1964年11月出生于黑龙江省哈尔滨市.现为哈尔滨工业大学教授、博士生导师.主要研究方向为雷达信号处理.E-mail:jiangyc@hit. edu. cn
    徐 丹 女,1996年4月出生于黑龙江省七台河市.主要研究方向为数据分析与智能图像处理.E-mail: xudanhit@hit.edu.cn
  • 基金资助:
    国家自然科学基金(61201304)

Recognition of 3D Rotating Ship Based on Mix-CV-CNN

ZHANG Yun, HUA Qing-long, JIANG Yi-cheng, XU Dan   

  1. School of Electronics and Information Engineering,Harbin Institute of Technology,Harbin,Heilongjiang 150001,China
  • Received:2021-09-13 Revised:2022-02-20 Online:2022-05-25 Published:2022-06-18
  • Contact: JIANG Yi-cheng

摘要:

在较高海情下,由于舰船目标处于随机摆动的非平稳运动状态,常规合成孔径雷达(Synthetic Aperture Radar, SAR)成像处理会使得目标散焦、方位模糊,从而导致三维转动舰船目标识别准确率低.本文提出一种混合型复数域卷积神经网络(Mix-type Complex-Valued Convolutional Neural Network,Mix-CV-CNN),并推导Mix-CV-CNN前向传播与反向传播算法.三维转动舰船目标经过SAR成像处理后存在剩余相位信息,Mix-CV-CNN能充分利用SAR复数域图像的幅度和相位信息,在不进行目标重聚焦的情况下,较好完成SAR复杂运动舰船目标的识别.实验表明,Mix-CV-CNN相较于具有相同自由度的实数域卷积神经网络(Real-Valued Convolutional Neural Network,RV-CNN)识别性能有所提高,实测数据识别平均准确率提高3.85%.

关键词: 合成孔径雷达, 复数域卷积神经网络, 三维转动, 目标散焦, 舰船目标识别, 混合型复数域卷积神经网络

Abstract:

Because the ship targets are in a non-stationary motion state of random swing, conventional synthetic aperture radar(SAR) imaging processing will make the targets defocused and azimuth blurred, resulting in the recognition accuracy of three-dimensional rotating ship. This paper proposes a mixed-type complex-valued convolutional neural network(Mix-CV-CNN) and derives the Mix-CV-CNN forward propagation and backpropagation algorithms. The three-dimensional rotating target has residual phase information after SAR imaging processing. The Mix-CV-CNN could make full use of the amplitude and phase information of the complex SAR image and could better complete the recognition of SAR three-dimensional rotating targets without target refocusing. The experimental results show that Mix-CV-CNN has improved recognition performance compared with the real-valued convolutional neural network(RV-CNN) with the same degree of freedom. The average accuracy is increased by 3.85%.

Key words: synthetic aperture radar(SAR), complex-valued convolutional neural network(CV-CNN), three-dimensional rotation, target defocus, ship target classification, mix-type complex-valued convolutional neural network(Mix-CV-CNN)

中图分类号: