摘要
针对现有的重载铁路扣件检测算法在复杂背景下难以满足检测要求、漏检率高的问题,提出一种基于YOLOv8n改进的扣件状态检测算法YOLO-CSM,该模型在骨干层引入了压缩-激励结构(SENetV2),增强了模型在不同通道的特征提取能力,提升了对细节特征的表达能力。在颈部中引入跨尺度特征融合结构(CCFF),融合不同尺度的特征信息,使上下文信息整合能力得到提升。为了突出扣件目标特征并减少复杂背景的影响,在网络中添加多尺度扩张注意力机制(MSDA),使模型关注点聚焦于扣件处,提高扣件定位精度。实验结果显示,YOLO-CSM网络在自制的包含四种重载铁路扣件状态的数据集上m AP@50-95和精度分别达到了94.0%和98.9%,相比于YOLOv8n提升了4.8%和3.8%,证明所提方法有效提升了在复杂背景下的扣件检测算法的性能。
The existing detection algorithms for heavy-haul railway fasteners are difficult to meet the detection requirements in complex backgrounds and have a high missed detection rate.An improved fastener state detection algorithm YOLO-CSM based on YOLOv8n is proposed.In this model,a compression-excitation structure(SENetV2)is introduced in the backbone layer,which enhances the feature extraction ability of the model in different channels and improves the expression of detailed features.A cross-scale feature fusion structure(CCFF)is introduced in the neck to fuse the feature information of different scales,thereby improving the integration ability of contextual information.To highlight the target features of the fasteners and reduce the influence of complex backgrounds,a multi-scale dilated attention mechanism(MSDA)is added to the network to focus the model's attention on the fasteners and improve the positioning accuracy of the fasteners.Experimental results show that,on the self-made dataset which contains four states of heavy haul railway fasteners,the mAP@50-95 and accuracy of the YOLO-CSM network reach 94.0%and 98.9%respectively,which are 4.8%and 3.8%higher than those of YOLOv8n.The proposed method effectively improves the performance of the fastener detection algorithm in complex backgrounds.
作者
张宇
丁建明
ZHANG Yu;DING Jianming(State Key Laboratory of Rail Transit Vehicle System,Southwest Jiaotong University,Chengdu 610031,China)
出处
《机械》
2025年第5期68-74,80,共8页
Machinery
关键词
扣件检测
压缩-激励模块
多尺度特征融合
多尺度扩张注意力机制
fastener detection
squeeze-and-excitation networks
cross-scale feature fusion
multi-scale dilated attention operation
