基于改进YOLOv8的铁路异物检测算法  

Algorithm of Railway Foreign Object Intrusion Detection Based on Improved YOLOv8

作　　者：赵子艺 张德育[1] 徐双成 ZHAO Ziyi;ZHANG Deyu;XU Shuangcheng(Shenyang Ligong University,Shenyang 110159,China;Shenyang Railway Petrochemical Group Co.,Ltd.,Shenyang 110005,China)

机构地区：[1]沈阳理工大学信息科学与工程学院,沈阳110159 [2]沈阳铁道石油化工集团有限公司,沈阳110005

出　　处：《沈阳理工大学学报》2025年第3期32-38,46,共8页Journal of Shenyang Ligong University

基　　金：辽宁省科技厅揭榜挂帅项目(2022JH1/10800050)。

摘　　要：由于监控场景覆盖面积大、侵限异物相对尺寸小、背景干扰较多等,现有铁路侵限监控系统检测精度较低,且达不到实时检测要求,为此提出一种基于改进YOLOv8的铁路异物检测算法。首先,引用FasterNet中部分卷积的设计思想,在颈部网络采用FasterC2f模块替代原C2f模块,以降低模型参数量和运算量,同时提升模型的特征融合能力;其次,对快速空间金字塔池化(SPPF)模块进行改进,引入大型可分离核注意力(LSKA)机制,提出一种SPPF_LSKA模块结构,以在特征提取时减少复杂背景的干扰,加强对小目标的检测能力;最后,采用WIoU损失函数替代原CIoU损失函数,以降低低质量、小目标样本带来的不良梯度影响。实验结果表明:与原YOLOv8模型相比,改进YOLOv8模型检测的平均精度均值(mAP@0.5)达到92.9%,提升了2.09%,模型权重降低了3.28%,浮点运算量降低了6.17%,在检测精度和效率之间达到了较好的平衡,能够满足实际检测需求。Due to the large coverage area of the monitoring scene,relatively small size of intruding objects and frequent background interference,the present railway intrusion monitoring system has low detection accuracy and cannot meet real-time detection requirements.Therefore,a railway foreign object detection algorithm based on improved YOLOv8 is proposed.Firstly,referencing the design concept of partial convolution in FasterNet,the neck network adopts FasterC2f module instead of the original C2f module to reduce the number of model parameters and computation,while improving the feature fusion ability of the model.Secondly,the fast spatial pyramid pooling(SPPF)module is improved by introducing a large separable kernel attention(LSKA)mechanism and proposing an SPPF_LSKA module structure to reduce the interference of complex backgrounds during feature extraction and enhance the detection ability of small targets.Finally,the WloU loss function is used instead of the original CIoU loss function to reduce the adverse gradient effects caused by low-quality and small target samples.The experimental results show that the improved YOLOv8 model has high mean average precision(mAP@0.5)to 92.9%,an improvement of 2.09%,and compared to the original YOLOv8 model,a reduction of 3.28%in model weight size,and a decrease of 6.17%in floating-point computation.It has achieved a good balance between detection accuracy and efficiency, which can meet practical detection needs.

关 键 词：深度学习 目标检测 YOLOv8 铁路异物 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]