系统工程与电子技术 ›› 2022, Vol. 44 ›› Issue (12): 3703-3709.doi: 10.12305/j.issn.1001-506X.2022.12.14

• 传感器与信号处理 • 上一篇    

基于YOLO框架的无锚框SAR图像舰船目标检测

贾晓雅1,2, 汪洪桥1,*, 杨亚聃3, 崔忠马2, 熊斌2   

  1. 1. 火箭军工程大学作战保障学院, 陕西 西安 710025
    2. 北京遥感设备研究所, 北京 100854
    3. 中国航天科工集团有限公司科研生产部, 北京 100048
  • 收稿日期:2021-07-08 出版日期:2022-11-14 发布日期:2022-11-24
  • 通讯作者: 汪洪桥
  • 作者简介:贾晓雅(1993—), 女, 硕士研究生, 主要研究方向为SAR图像目标检测|汪洪桥(1979—), 男, 副教授, 博士, 主要研究方向为人工智能、机器学习、遥感图像处理|杨亚聃(1983—), 男, 高级工程师, 硕士, 主要研究方向为电子对抗|崔忠马(1978—), 男, 研究员, 硕士, 主要研究方向为雷达系统|熊斌(1982—), 男, 高级工程师, 博士, 主要研究方向为雷达系统
  • 基金资助:
    陕西省自然科学基础研究计划(2020JM-358)

Anchor free SAR image ship target detection method based on the YOLO framework

Xiaoya JIA1,2, Hongqiao WANG1,*, Yadan YANG3, Zhongma CUI2, Bin XIONG2   

  1. 1. Department of Information Engineering, Rocket Force University of Engineering, Xi'an 710025, China
    2. Beijing Institute of Remote Sensing Equipment, Beijing 100854, China
    3. Scientific Research and Production Department, China Aerospace Science and Industry Corporation, Beijing 100048, China
  • Received:2021-07-08 Online:2022-11-14 Published:2022-11-24
  • Contact: Hongqiao WANG

摘要:

面向合成孔径雷达(synthetic aperture radar, SAR)多目标检测应用, 提出了一种基于YOLO (you only look once) 框架的无锚框SAR图像舰船目标检测方法。该方法针对YOLOv3锚框需要预先设定且无法完美契合的弊端, 通过采用无锚框方法更好适应所检测目标的大小, 便于多尺度目标使用。在此基础上, 给CSPDarknet53网络增加了注意力机制作为特征提取网络, 然后经过能够增大感受野的改进特征金字塔网络(feature pyramid network, FPN)后, 把特征图传给无锚框检测头, 有效提升了目标类别和位置的预测精度。实验证明, 所提算法在公开SAR舰船数据集上平均精度比YOLOv3提高3.8%,达到了94.8%, 虚警率降低4.8%。

关键词: 合成孔径雷达图像, YOLO, 无锚框, 舰船目标检测

Abstract:

For synthetic aperture radar (SAR) multi-target detection applications, this paper proposes an anchor free SAR image ship target detection method based on the you only look once (YOLO) framework. This method is aimed at the disadvantage that the YOLOv3 anchor needs to be preset and cannot fit perfectly. By adopting the anchor free method, it can better adapt to the size of the detected target and facilitate the use of multi-scale targets. On this basis, the attention mechanism is added to the CSPDarknet53 network as a feature extraction network, and then after an improved feature pyramid network (FPN) that can increases the receptive field, the feature map is transmitted to the anchor free detection head, which effectively improves the prediction precision of the target category and location. Experiments show that the improved algorithm has an average precision of 3.8% higher than YOLOv3 on the public SAR ship data set, reaching 94.8%, and the false alarm rate is reduced by 4.8%.

Key words: synthetic aperture radar (SAR) image, YOLO, anchor free, ship target detection

中图分类号: