系统工程与电子技术 ›› 2022, Vol. 44 ›› Issue (12): 3652-3660.doi: 10.12305/j.issn.1001-506X.2022.12.08

• 传感器与信号处理 • 上一篇    下一篇

基于改进ResNet网络的复数SAR图像舰船目标识别方法

雷禹, 冷祥光*, 周晓艳, 孙忠镇, 计科峰   

  1. 国防科技大学电子科学学院, 湖南 长沙 410073
  • 收稿日期:2021-06-28 出版日期:2022-11-14 发布日期:2022-11-24
  • 通讯作者: 冷祥光
  • 作者简介:雷禹(1997—), 男, 硕士研究生, 主要研究方向智能电子对抗与评估、SAR图像目标智能解译|冷祥光(1991—), 男, 讲师, 博士, 主要研究方向为遥感信息处理、SAR图像智能解译、机器学习|周晓艳(1997—), 女, 硕士研究生, 主要研究方向为SAR图像目标智能识别、张量分解|孙忠镇(1996—), 男, 硕士研究生, 主要研究方向为智能电子对抗与评估、SAR图像目标识别|计科峰(1974—), 男, 教授, 博士, 主要研究方向为SAR图像解译、目标检测与识别、特征提取、SAR和AIS匹配
  • 基金资助:
    国家自然科学基金(62001480);湖南省自然科学基金(2021JJ40684)

Recognition method of ship target in complex SAR image based on improved ResNet network

Yu LEI, Xiangguang LENG*, Xiaoyan ZHOU, Zhongzhen SUN, Kefeng JI   

  1. College of Electronic Science and Technology, National University of Defense Technology, Changsha 410073, China
  • Received:2021-06-28 Online:2022-11-14 Published:2022-11-24
  • Contact: Xiangguang LENG

摘要:

合成孔径雷达(synthetic aperture radar, SAR)采用微波相干成像, 因此SAR图像本质上是复数的。传统基于神经网络的SAR图像目标识别方法, 通常只处理SAR图像的幅度信息, 无法有效利用SAR图像特有的复数信息。本文面向SAR图像中的舰船目标识别应用, 从SAR图像的本质出发, 首先通过组合SAR图像的实部、虚部和幅度三通道信息, 隐式地提供了输入数据的复数信息表示; 然后在ResNet18网络及其结构基础上引入通道注意力机制, 使网络能自适应学习实部、虚部和幅度三通道之间包含的复数信息; 最后引入标签平滑正则化, 解决因复数数据集样本较少出现的过拟合现象。基于OpenSARShip数据集的实验结果表明, 所提方法可以较好利用SAR图像本身的复数信息, 在一定程度上提升了基于深度神经网络的舰船目标识别效果。

关键词: 合成孔径雷达, 神经网络, 复数信息, 舰船目标识别

Abstract:

Synthetic aperture radar (SAR) uses microwave coherent imaging, so SAR images are complex in essence. Traditional SAR image target recognition methods based on neural networks usually only process the amplitude information of SAR images, but cannot effectively use the unique complex information of SAR images. This paper aims at the application of ship target recognition in SAR image. Starting from the essence of SAR image, the complex information representation of input data is implicitly provided by combining the real part, imaginary part and amplitude information of SAR image. Then the channel attention mechanism is introduced on the basis of the ResNet18 network and its structure, so that the network can adaptively learn the complex information contained in the three channels of real part, imaginary part and amplitude. Finally, label smoothing regularization is introduced to solve the over-fitting phenomenon due to the lack of samples in complex data sets. The experimental results based on the OpenSARShip data set show that the method proposed in this paper can make better use of the complex information of the SAR image itself, and to a certain extent improve the effect of ship target recognition based on neural network.

Key words: synthetic aperture radar (SAR), neural networks, complex information, ship target recognition

中图分类号: