系统工程与电子技术 ›› 2026, Vol. 48 ›› Issue (11): 3770-3778.doi: 10.12305/j.issn.1001-506X.2026.11.09

• 传感器与信号处理 • 上一篇    下一篇

基于改进YOLOv8的SAR图像目标检测算法

罗益炜1,2(), 徐正1,2(), 刘向宏3(), 曹越1,2(), 陈庆锌1,2(), 唐云杉1,2()   

  1. 1. 中国科学院空天信息创新研究院,北京 100094
    2. 中国科学院大学电子电气与通信工程学院,北京 100049
    3. 中国人民解放军 93160部队,北京 100041
  • 收稿日期:2025-09-09 修回日期:2026-01-28 接受日期:2026-06-03 出版日期:2026-10-25 发布日期:2026-05-29
  • 通讯作者: 徐正 E-mail:624860529@qq.com;xuzheng292@163.com;liuxh232@163.com;yuecao0927@hotmail.com;chenqx01@aircas.ac.cn;tangyunshan22@mails.ucas.ac.cn
  • 作者简介:罗益炜(2003—),男,硕士研究生,主要研究方向为SAR目标检测、深度学习
    刘向宏(1968—),男,正高级工程师,硕士,主要研究方向为空基预警探测与信息传输
    曹 越(1986—),男,副研究员,博士,主要研究方向为实时信号处理与高性能计算
    陈庆锌(1994—),男,助理研究员,硕士,主要研究方向为雷达系统设计、信号处理
    唐云杉(1994—),男,博士研究生,主要研究方向为SAR目标检测,SAR目标图像生成
  • 基金资助:
    国家自然科学基金(62271363)资助课题

Improved YOLOv8-based target detection algorithm for SAR images

Yiwei Luo1,2(), Zheng Xu1,2(), Xianghong Liu3(), Yue Cao1,2(), Qingxin Chen1,2(), Yunshan Tang1,2()   

  1. 1. Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing 100094,China
    2. School of Electronic,Electrical and Communication Engineering,University of Chinese Academy of Sciences,Beijing 100049,China
    3. Unit 93160 of the PLA,Beijing 100041
  • Received:2025-09-09 Revised:2026-01-28 Accepted:2026-06-03 Online:2026-10-25 Published:2026-05-29
  • Contact: Zheng Xu E-mail:624860529@qq.com;xuzheng292@163.com;liuxh232@163.com;yuecao0927@hotmail.com;chenqx01@aircas.ac.cn;tangyunshan22@mails.ucas.ac.cn

摘要:

针对合成孔径雷达(synthetic aperture radar,SAR)图像目标检测任务,提出一种基于YOLO(you only look once)框架的目标检测算法。针对传统光学检测模型在SAR图像小目标检测中存在的性能局限,通过3个关键技术创新实现性能提升。首先,在损失函数中引入尺度无关的归一化Wasserstein距离,改善小目标检测性能。其次,引入高效多尺度注意力机制,增强模型对SAR图像特征的提取能力。最后,设计一种基于Transformer架构的轻量化检测头,在提高检测精度的同时降低计算复杂度。在高分辨率SAR图像数据集(High-Resolution SAR Images Dataset,HRSID)上的实验结果表明,所提算法较YOLOv8n基准模型在小目标mAP50上提升3.9%,整体mAP50提升1.2%,计算量降低6.1%,实现精度与效率的协同优化。通过另外3个数据集进行泛化实验,进一步验证所提算法具有优良的泛化能力。

关键词: YOLOv8, 合成孔径雷达, 目标检测

Abstract:

This paper proposes an improved object detection algorithm based on the YOLO (you only look once) framework for synthetic aperture radar (SAR) image target detection. To address the performance limitations of traditional optical detection models in detecting small targets in SAR images, this study introduces three key technical innovations. First, a scale-independent normalized Wasserstein distance (NWD) metric is incorporated into the loss function to enhance small target detection. Second, an efficient multi-scale attention (EMA) mechanism is employed to improve the model’s ability to extract features from SAR images. Finally, a lightweight detection head named ATTDetect, based on the Transformer architecture, is designed to increase detection accuracy while reducing computational complexity. Experimental results on the High-Resolution SAR Images Dataset (HRSID) show that compared to the baseline YOLOv8n model, the proposed algorithm improves mAP50 for small targets by 3.9%, increases overall mAP50 by 1.2%, and reduces computational load by 6.1%, achieving a synergistic optimization of accuracy and efficiency. Furthermore, generalization experiments conducted on other three datasets further validate the algorithm’s strong generalization capability.

Key words: YOLOv8, synthetic aperture radar, object detection

中图分类号: