摘要
针对遥感影像交通设施检测精度不高的问题,提出一种改进YOLOv5的交通设施检测算法YOLOv5-Traffic。首先将非对称卷积核组引入骨干网络,在不增加额外算力的前提下提高对样本特征的提取能力;其次在特征融合网络中引入了GsOP注意力,跨层特征拼接通道与更大尺度的特征图输出窗口,以提高对小目标的检测精度;最后使用Alpha IoU计算目标框回归损失,提高预测框回归的精准度。在两组不同数据集上的实验结果表明,所提出的改进算法较YOLOv5在精确率、准确率、平均精度均值方面均有明显提升,在满足实时检测要求的前提下,具备更高的检测精度与更强的泛化能力。
Aiming at the problem of low detection accuracy of traffic facilities in remote sensing images,an improved YOLOv5 traffic facility detection algorithm YOLOv5-Traffic is proposed.First,the asymmetric convolution kernel group is introduced into the backbone network to improve the ability to extract sample features without adding additional computing power;then the GsOP attention,the cross-layer feature splicing channel,and larger scale output window of the feature map are introduced into the feature fusion network to improve the detection accuracy of small targets;finally,Alpha IoU is used to calculate the regression loss of the target frame to improve the accuracy of the prediction frame regression.The experimental results on two sets of different data sets show that the proposed improved algorithm has significantly improved accuracy,correctness,and average accuracy compared with YOLOv5,and has higher detection accuracy and stronger generalization ability under the premise of meeting the real-time detection requirements.
作者
韩禹
HAN Yu(Liaoning Provincial Natural Resources Affairs Service Center,Shenyang 110034,China)
出处
《测绘与空间地理信息》
2025年第10期112-115,125,共5页
Geomatics & Spatial Information Technology
