系统工程与电子技术 ›› 2023, Vol. 45 ›› Issue (4): 1032-1039.doi: 10.12305/j.issn.1001-506X.2023.04.12

• 传感器与信号处理 • 上一篇    

基于特征增强网络的SAR图像舰船目标检测

张冬冬, 王春平, 付强   

  1. 陆军工程大学石家庄校区电子与光学工程系, 河北 石家庄 050003
  • 收稿日期:2021-04-19 出版日期:2023-03-29 发布日期:2023-03-28
  • 通讯作者: 张冬冬
  • 作者简介:张冬冬(1993—), 男, 硕士研究生, 主要研究方向为目标识别
    王春平(1965—), 男, 教授, 博士, 主要研究方向为图像工程、火力与指挥控制
    付强(1981—), 男, 讲师, 博士, 主要研究方向为智能视觉与目标检测

Ship target detection in SAR image based on feature-enhanced network

Dongdong ZHANG, Chunping WANG, Qiang FU   

  1. Department of Electronic and Optical Engineering, Army Engineering University Shijiazhuang Campus, Shijiazhuang 050003, China
  • Received:2021-04-19 Online:2023-03-29 Published:2023-03-28
  • Contact: Dongdong ZHANG

摘要:

合成孔径雷达(synthetic aperture radar, SAR)图像场景复杂度高、舰船目标尺度小, 传统方法检测效率低、虚警概率大。针对以上问题, 提出一种特征增强网络用于SAR图像舰船目标检测。首先, 利用I-Darknet-53(improved Darknet-53)提取特征信息, 构建4层特征金字塔丰富低层特征。其次, 将多个特征层进行跨尺度连接, 使低层细节信息更易于向高层语义信息映射, 增强特征的传播和重用。最后, 利用多尺度注意力模型增强特征信息, 为检测器提供高质量的判断依据。试验结果表明, 所提算法在SSDD数据集上的平均检测精度为95%。相较于其他网络模型, 所提算法具有明显优势。

关键词: 合成孔径雷达图像, 目标检测, 特征增强, 多尺度融合, 多尺度注意力

Abstract:

Traditional detection methods are inefficient and have a high probability of false alarm due to the high complexity of synthetic aperture radar (SAR) image scenes and small scale of ship targets. To address these problems, this paper proposes a feature-enhanced network for SAR image ship target detection is proposed. Firstly, feature information is extracted using I-Darknet53 (improved Darknet-53), and a four-layer feature pyramid is constructed to enrich low-level features. Secondly, multiple feature layers are connected across scales to make low-level detail information easier to map to high-level semantic information, thus enhancing the propagation and reuse of features. Finally, the feature information is enhanced using a multi-scale attention model to provide a high-quality judgment basis for the detector. The experimental results show that the average detection accuracy of the proposed algorithm on the SSDD dataset is 95%. The proposed algorithm has high precision compared with other network models.

Key words: synthetic aperture radar (SAR) image, target detection, feature enhancement, multi-scale fusion, multi-scale attention

中图分类号: