基于多尺度特征聚合的遥感图像目标检测算法  

Object Detection Algorithm of Remote Sensing Image Based on Multi-scale Feature Aggregation

作　　者：刘伟东 周华平 LIU Weidong;ZHOU Huaping(School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan Anhui 232001,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001

出　　处：《兰州工业学院学报》2025年第2期46-53,共8页Journal of Lanzhou Institute of Technology

基　　金：安徽省重点研发计划国际科技合作专项(202004b11020029)。

摘　　要：由于遥感图像目标检测中尺度变换以及小目标特征信息丢失等问题的存在,导致现有模型检测精度不佳。为解决上述问题,以YOLOv8为基线网络进行改进,提出了一种基于多尺度特征聚合的遥感图像目标检测算法。首先,引入LSK(Large Selective Kernel)构建LSKC2f模块,通过LSKC2f自适应地调整感受野大小,帮助模型充分提取多尺度目标的特征信息;然后,设计一个全局特征聚合(Global Feature Aggregation)模块,将浅层特征图与多个不同尺度的深层特征图进行融合并输出,提升小目标检测精度。实验结果表明:相较于基线网络,改进的新算法在具有挑战性的公共数据集DIOR上的检测精度提升了1.4%达到了73%,证明了本算法的有效性。Due to issues such as scale variations and the loss of small object feature information in remote sensing image object detection,the detection accuracy of existing models is not good.To address these problems,this study proposes an object detection algorithm of remote sensing image based on multi-scale feature aggregation by using YOLOv8 as the baseline network.Firstly,the LSK(Large Selective Kernel)is introduced to construct the LSKC2f module,which adaptively adjusts the receptive field size to help the model fully extract feature information of multi-scale objects.Then,a Global Feature Aggregation module is designed to fuse shallow feature maps with multiple deep feature maps of different scales,thereby enhancing the detection accuracy of small objects.Experimental results show that compared to the baseline network,the proposed algorithm improves the detection accuracy on the challenging public dataset DIOR by 1.4%,reaching 73%,demonstrating the effectiveness of the algorithm.

关 键 词：计算机视觉 遥感图像目标检测 YOLOv8 多尺度 小目标 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]