quotation:[Copy]
[Copy]
【Print page】 【Download 【PDF Full text】 【View/Add Comment】【Download reader】【 Close 】

←Previous page|Page Next →

Back Issue    Advanced search

This Paper:Browse 248   Download 224 本文二维码信息
码上扫一扫!
面向卫星遥感数据车辆识别的改进YOLOv9算法
王钦艺,郝庆涛,张鹏飞,孙雷刚,左璐,鲁军景,范小勇,常方哲
0
(1.河北省科学院地理科学研究所,石家庄 050011;2.河北省地理信息开发应用技术创新中心,石家庄 050011;3.河北省交通规划设计研究院有限公司,石家庄 050011)
摘要:
针对基于遥感的车辆检测场景中普遍存在的纹理表征不足、误检漏检多、感受野尺度受限等问题,提出了一种基于深度学习的目标检测算法Y0L0v9-Car。该算法主要改进是:首先,在目标检测头融入多尺度特征,增加对于浅层特征的捕捉能力;其次,引入多尺度空洞注意力(Multi-scale Dilated Attention,MSDA )机制,通过扩大感受野,增加信息聚合能力,减少背景中冗余信息的影响;最后,采用EIoU(Enhanced Intersection over Union Loss)损失函数,减少边界框回归的几何差异,提高模型泛化性与鲁棒性。此外,构建了一个车辆目标检测的遥感数据集。实验结果表明,相比于YOLOv9模型,YOLOv9-Car算法在平均精度、准确率和召回率方面分别提升了7.1%、10.3%和8.1%。
关键词:  卫星遥感  车辆目标检测  YOLOv9  多尺度空洞注意力(MSDA)  多尺度特征层
DOI:10.20079/j.issn.1001-893x.250702004
基金项目:河北省科学院基本科研业务费制度试点项目(2025PF10)
Vehicle Recognition Algorithm Based on Improved YOLOv9 for Satellite Remote Sensing Data
WANG Qinyi,HAO Qingtao,ZHANG Pengfei,SUN Leigang,ZUO Lu,LU Junjing,FAN Xiaoyong,CHANG Fangzhe
(1.Institute of Geographical Sciences,Hebei Academy of Sciences,Shijiazhuang 050011,China;2.Hebei Technology Innovation Center for Geographic Information Application,Shijiazhuang 050011,China; 3.Hebei Provincial Communications Planning,Design and Research Institute,Shijiazhuang 050011,China)
Abstract:
A target detection algorithm named Y0L0v9-Car based on deep learning is proposed to address the common problems in remote sensing-based vehicle detection scenarios,such as insufficient texture representation,numerous false detections and missed detections,and limited receptive field size.The main improvements of the algorithm are as follows:Firstly,multi-scale features are integrated into the detection head to enhance the ability to capture shallow features.Secondly,the multi-scale dilated attention(MSDA) mechanism is introduced to expand the receptive field,improve information aggregation,and reduce the impact of redundant background information.Finally,the enhanced intersection over union loss(EIoU) is adopted to minimize geometric discrepancies in bounding box regression,thereby enhancing the model’s generalization capability and robustness.Additionally,a remote sensing dataset is constructed specifically for vehicle target detection.Experimental results demonstrate that,compared with YOLOv9,the proposed YOLOv9-Car algorithm achieves improvements of 7.1%,10.3%,and 8.1% in average precision,accuracy,and recall rate,respectively.
Key words:  satellite remote sensing  vehicle target detection  YOLOv9  multi-scale dilated attention(MSDA)  multi-scale feature layer