系统工程与电子技术 ›› 2022, Vol. 44 ›› Issue (10): 3096-3103.doi: 10.12305/j.issn.1001-506X.2022.10.13

• 传感器与信号处理 • 上一篇    下一篇

复杂背景下SAR图像近岸舰船目标检测

李永刚1,2,*, 朱卫纲1, 黄琼男1, 李云涛1, 何永华1   

  1. 1. 航天工程大学电子光学工程系, 北京 101416
    2. 电子信息系统复杂电磁环境效应国家重点实验室, 河南 洛阳 471032
  • 收稿日期:2021-04-01 出版日期:2022-09-20 发布日期:2022-10-24
  • 通讯作者: 李永刚
  • 作者简介:李永刚(1995—), 男, 硕士研究生, 主要研究方向为SAR图像目标检测与识别|朱卫纲(1973—), 女, 教授, 博士, 主要研究方向为SAR图像目标检测与识别、空间信息对抗、认知电子战|黄琼男(1995—), 男, 硕士研究生, 主要研究方向为SAR舰船目标检测和识别数据集构建、SAR图像生成|李云涛(1984—), 男, 讲师, 博士, 主要研究方向为雷达信号处理与目标探测技术|何永华(1980—), 女, 副教授, 硕士, 主要研究方向为雷达目标检测与识别
  • 基金资助:
    复杂电磁环境效应国家重点实验室项目(CEMEE2020Z0203B)

Near-shore ship target detection with SAR images in complex background

Yonggang LI1,2,*, Weigang ZHU1, Qiongnan HUANG1, Yuntao LI1, Yonghua HE1   

  1. 1. Department of Electronic and Optical Engineering, Space Engineering University, Beijing 101416, China
    2. National Key Laboratory of Complex Electromagnetic Environmental Effect of Electronic Information System, Luoyang 471032, China
  • Received:2021-04-01 Online:2022-09-20 Published:2022-10-24
  • Contact: Yonggang LI

摘要:

针对合成孔径雷达(synthetic aperture radar, SAR)图像近岸舰船目标易受背景杂波的影响, 造成SAR图像近岸舰船目标检测检测率低、虚警率和漏检率高的问题,提出一种适用于复杂背景下SAR图像近岸舰船目标检测的DFF-Yolov5(deformable feature fusion you only look once 5)算法。构建了一个专门用于SAR图像复杂背景近岸舰船目标检测的数据集, 基于Yolov5目标检测算法, 在特征提取网络中进行特征细化和多特征融合两个方面的改进。在特征提取网络中利用可变形卷积神经网络改变卷积对目标采样点的位置, 增强目标的特征提取能力, 提高复杂背景下SAR图像舰船目标的检测率。在多特征融合网络结构中采用级联和并列金字塔, 进行不同层级的特征融合。同时,使用空洞卷积扩大特征提取的视觉感受野, 增强网络对复杂背景近岸多尺度舰船目标的适应性, 降低复杂背景下SAR图像舰船目标检测的虚警率。通过在构建的复杂背景近岸舰船检测数据集上的测试实验, 结果表明: DFF-Yolov5的平均准确率为85.99%, 相比于原始的Yolov5, 所提方法平均准确率提高了5.09%, 精度提高了1.4%。

关键词: 合成孔径雷达, 目标检测, 近岸舰船目标, 多特征融合, 可变形卷积神经网络

Abstract:

In order to solve the problems of low detection rate, high false alarm rate and high missed detection rate of near-shore ship target detection from synthetic aperture radar (SAR) images caused by the vulnerability of SAR images to background clutter, a deformable feature fusion you only look once 5 (DFF-Yolov5) algorithm is proposed for the detection of near-shore ship targets in SAR images with complex background. The algorithm is based on the Yolov5 target detection algorithm, with two improvements in the feature extraction network: feature refinement and multi-feature fusion. A special data set for near-shore ship target ditection in complex background of SAR images is costructed. In the feature extraction network, a deformable convolutional neural network is used to change the position of the target sampling points to enhance the feature extraction capability of the target and improve the detection rate of SAR images ship targets in complex background. In the multi-feature fusion network structure, cascade and parallel pyramids are used to perform feature fusion at different levels. At the same time, cavity convolution is used to expand the visual field of feature extraction, enhance the adaptability of the network to near-shore multi-scale ship targets in complex background, and reduce the false alarm rate of SAR image ship target detection in complex backgrounds. Through the test experiments on the constructed complex background near-shore ship, the results show that the average accuracy of DFF-Yolov5 is 85.99%, compared with the original Yolov5, the average accuracy of the proposed method is improved by 5.09% and the precision is improved by 1.4%.

Key words: synthetic aperture radar (SAR), target detection, near-shore ship target, multi-feature fusion, deformable convolutional neural network

中图分类号: