基于SAR图像的小目标检测算法  

Small target detection algorithm based on SAR images

作　　者：许峻滔 王卓薇[1,2] 赵艮平 XU Jun-tao;WANG Zhuo-wei;+;ZHAO Gen-ping(School of Computers,Guangdong University of Technology,Guangzhou 510006,China;Key Laboratory of Cyber-Physical System,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学计算机学院,广东广州510006 [2]广东工业大学信息物理融合系统重点实验室,广东广州510006

出　　处：《计算机工程与设计》2025年第2期570-577,共8页Computer Engineering and Design

基　　金：以农业领域为单元的广东省现代农业产业共性关键技术研发创新团队建设基金项目(2023KJ147);广州市基础与应用研究基金项目(202201010273);教育部与北京航天宏图信息技术有限公司合作教育基金项目(220802313160749);广东省农业科技创新推广项目的部分支持基金项目(2019KJ147);广东省信息物理融合系统重点实验室基金项目(2020B1212060069)。

摘　　要：针对SAR图像上的小目标难以被检测的问题,提出一种STSAR-YOLOv5目标检测算法。在YOLOv5基础上增加一层检测头捕捉小目标;在骨干网络部分加入BoTNet模块,其具有全局信息获取能力,能够在高度密集场景下定位小目标。在自主构建的SSAD数据集上进行对比实验,实验结果表明,STSAR-YOLOv5的检测效果较YOLOv5有明显提升,该方法平均检测精度(mAP_0.5:0.95)提高了7.8%。相比YOLOv5算法,该算法对SAR图像中的小目标表现出较好的检测性能和鲁棒性。To address the challenge of detecting small targets in SAR images,an object detection algorithm,named STSAR-YOLOv5,was introduced.An extra detection head was added to the YOLOv5 base for capturing small targets.The BoTNet module was incorporated into the backbone network,which possessed the capability to acquire global information for precise localization of small targets,particularly in highly dense scenarios.Comparative experiments were conducted on a self-constructed SSAD dataset.The results of these experiments demonstrate a significant improvement in the detection performance of STSAR-YOLOv5 compared to YOLOv5,with an average detection accuracy(mAP_0.5:0.95)increase of 7.8%.In comparison to the YOLOv5 algorithm,this method exhibits superior detection performance and robustness for small targets in SAR images.

关 键 词：合成孔径雷达 目标检测 小目标 遥感图像 深度学习 YOLO模型 计算机视觉 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]