摘要
针对现有方法在多源遥感影像目标检测过程中,存在小目标检测精度不高、多场景泛化能力不强等问题,本文提出一种基于YOLOv8的改进方法。在特征提取网络中,引入广义高效层聚合网络及三重注意力机制,提高模型对小目标特征的挖掘能力与学习能力;在特征融合阶段,引入语义-细节特征灌注层,充分融合深、浅层的目标特征;在检测阶段,增加用于小尺寸目标的第四组检测头,并使用滑动函数来计算置信度损失,提高模型对困难样本的关注度。实验结果表明,改进模型对于光学及合成孔径雷达遥感影像中的多类目标,都表现出了优于YOLOv8模型的检测精度与多场景泛化能力,能够在城市规划、应急救援等场景下发挥应用价值。
Aiming at the problems of low detection accuracy and weak generalization ability of small targets in the process of implementing multi-source remote sensing image target detection in the existing model,an improved method based on YOLOv8 is proposed.In the feature extraction network,the generalized efficient aggregation network and triple attention mechanism are introduced to improve the mining ability and learning ability of the model for small target features;In the feature fusion stage,semantic detail feature perfusion layer is introduced to fully integrate the deep and shallow target features;In the detection phase,the fourth group of detection heads for small-size targets is added,and the sliding function is used to calculate the confidence loss,so as to improve the attention of the model to difficult samples.The experimental results show that the improved model has better detection accuracy and multi-scene generalization ability than YOLOv8 for many kinds of targets with different scales in optical and synthetic aperture radar remote sensing images,and can play an application value in urban planning,emergency rescue,and other scenes.
作者
杨德志
YANG Dezhi(Beijing Dongfang Zhiyuan Technology Co.,Ltd.,Beijing 100191,China)
出处
《测绘与空间地理信息》
2026年第4期126-129,133,共5页
Geomatics & Spatial Information Technology
关键词
多源遥感目标检测
YOLOv8
广义高效层聚合网络
三重注意力
语义细节灌注模块
multi-source remote sensing target detection
YOLOv8
generalized efficient layer aggregation network
triple attention
semantic detail perfusion module
