1.State Key Laboratory of Advanced Rail Autonomous Operation, Beijing Jiaotong University, Beijing 100044, China
2.School of Transportation, Beijing Jiaotong University, Beijing 100044, China
3.Beijing Subway Operation Co. , Ltd. , Beijing 100044, China
4.Technology Innovation Research Institute Branch, Beijing Subway Operation Co. , Ltd. , Beijing 100044, China
Show less
文章历史+
Received
Published
2025-08-14
2026-07-01
Issue Date
2026-10-28
PDF (7066K)
摘要
针对轨道扣件缺陷检测中单一可见光图像易受环境干扰、缺乏空间结构信息,导致模型对小尺寸缺陷及扣件空间形变敏感性不足,进而易产生误检和漏检的问题,提出一种基于注意力机制与双模态融合YOLO(Attention and Dual-Modal Fusion YOLO,ADF-YOLO)模型的检测方法。首先,构建ADF-YOLO模型,通过双流骨干网络结构,并行提取轨道扣件可见光图像的纹理特征和深度图像的空间结构特征;其次,为提高对小尺寸缺陷特征的捕获能力,在ADF-YOLO模型骨干网络中引入注意力与多分支卷积协同模块,扩大感受野,强化细节特征表达;最后,为实现对异构特征的高效融合,在ADF-YOLO模型颈部网络引入注意力引导的双模态特征融合模块,充分挖掘双模态特征融合范式,生成鲁棒性更高且对缺陷形态更敏感的融合特征。结果表明:在扣件缺陷多模态数据集上,基于ADF-YOLO模型的检测方法,在交并比阈值为0.5和0.5~0.95时平均精度均值分别达90.9%和70.5%;与多种单模态和多模态目标检测模型相比,基于ADF-YOLO模型的检测方法在检测精度上具有显著优势。提出的ADF-YOLO模型可为轨道扣件缺陷巡检任务提供可靠的技术支撑。
Abstract
In rail fastener defect detection, single-modal visible-light images are susceptible to environmental interference and lack spatial structural information. This limitation reduces the model’s sensitivity to small-sized defects and spatial deformations of fasteners. Consequently, it increases the rate of false and missed detections. To address these issues, this paper proposes a detection method based on the Attention and Dual-Modal Fusion YOLO (ADF-YOLO) model. First, the ADF-YOLO model is constructed, which employs a dual-stream backbone architecture to concurrently extract texture features from visible-light images and spatial structural features from depth images of railway fasteners. Second, an Attention and Multi-Branch Convolution Synergy Module is introduced into the backbone network to improve the capture of small-sized defect features. This module expands the receptive field and enhances the representation of detailed features. Finally, an Attention-Guided Dual-Modal Feature Fusion Module is integrated into the neck network for the efficient fusion of heterogeneous features. This module fully exploits the dual-modal feature fusion paradigm. Consequently, it generates more robust fused features that are highly sensitive to defect morphology. The results demonstrate that on a multimodal fastener defect dataset, the ADF-YOLO model detection method achieves a mean average precision (mAP) of 90.9% and 70.5% at IoU thresholds of 0.5 and 0.5-0.95, respectively. The ADF-YOLO method significantly outperforms various single-modal and multi-modal object detection models in detection accuracy. The proposed ADF-YOLO model provides reliable technical support for railway fastener defect inspection tasks.
采用精确率P(Precision),召回率R(Recall),平均精度均值mAP(Mean Average Precision),小目标类别的平均精度APS(Average Precision to Small)评价模型性能。P为检测结果中正确检测的比例,R为所有目标中被成功检测的比例,mAP为平均精度AP(Average Precision)的平均值,是评价模型综合性能的关键指标,其中AP为模型识别某一类别目标的平均精度。一般采用mAP@0.5和mAP@0.5:0.95展示mAP的大小,二者分别表示当预测框与真实框的交并比(Intersection over Union,IoU)为0.5时mAP的大小,以及当IoU从0.5到0.95以0.05为步长变化时,各IoU值下的mAP的平均值。APS指小目标类别的平均精度,计算方式为提取螺帽丢失、螺帽松动和平垫圈丢失3类小目标缺陷在IoU=0.5:0.95阈值下的单类平均精度,计算其算术平均值。各评价指标的计算式分别为
QinYong, CaoZhiwei, SunYongfu, et al. Research on active safety methodologies for intelligent railway systems[J]. Engineering, 2023, 27: 266-279.
[2]
RenShaoqing, HeKaiming, GirshickR, et al. Faster R-CNN: towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6): 1137-1149.
[3]
HeKaiming, GkioxariG, DollárP, et al. Mask R-CNN[C]//2017 IEEE International Conference on Computer Vision (ICCV), October 22-29, 2017, Venice, Italy. New York: IEEE Press, 2017: 2980-2988.
LiangNan, ZhangWei, LiuYanglong, et al. Fastener clips defect detection based on improved faster R-CNN in high-speed railway[J]. Journal of Terahertz Science and Electronic Information Technology, 2024, 22(11): 1221-1227, 1269.
[6]
黄午祥.基于深度学习的铁轨表面缺陷检测技术研究[D].武汉:湖北工业大学,2023.
[7]
HuangWuxiang. Research on rail surface defect detection technology based on deep learning[D].Wuhan: Hubei University of Technology, 2023.
XuGuiyang, LiJinyang, BaiTangbo, et al. Detection method of track fastener state based on improved mask R-CNN[J]. China Railway Science, 2022, 43(1): 44-51.
WuSongying, LiuLinya, JiangJiaming, et al. Detection algorithm of improved YOLOv5s of railway fastener under complex background[J]. China Railway Science, 2023, 44(3): 53-63.
WuFu, JiangPengmin, LiZhongxue, et al. Lightweight detection method for track fastener status based on improved YOLOv8s[J]. Journal of Beijing Jiaotong University, 2024, 48(5): 59-68.
[16]
HeJin, WangWei, LvFengmao, et al. Multi-scale CNN-transformer hybrid network for rail fastener defect detection[J]. IEEE Transactions on Intelligent Transportation Systems, 2025, 26(6): 8894-8906.
[17]
Gómez-ChovaL, TuiaD, MoserG, et al. Multimodal classification of remote sensing images: a review and future directions[J]. Proceedings of the IEEE, 2015, 103(9): 1560-1584.
[18]
GaoYang, CaoZhiwei, QinYong, et al. Railway fastener anomaly detection via multisensor fusion and self-driven loss reweighting[J]. IEEE Sensors Journal, 2024, 24(2): 1812-1825.
[19]
WangJingpeng, SongKechen, ZhangDefu, et al. Collaborative learning attention network based on rgb image and depth image for surface defect inspection of no-service rail[J]. IEEE/ASME Transactions on Mechatronics, 2022, 27(6): 4874-4884.
[20]
GeXuanyu, CaoZhiwei, QinYong, et al. An anomaly detection method for railway track using semisupervised learning and vision-lidar decision fusion[J]. IEEE Transactions on Instrumentation and Measurement, 2024, 73: 5023215.
[21]
陈仕涛.基于多传感器信息融合的轨道异物检测研究与实现[D].南京:南京邮电大学,2023.
[22]
ChenShitao. Research and implementation of track abnormal object detection based on multi-sensor information Fusion[D]. Nanjing: Nanjing University of Posts and Telecommunications, 2023.
ZhaoHongwei, ZhengJiajun, ZhaoXinxin, et al. Rail surface defect method based on bimodal-modal deep learning[J]. Computer Engineering and Applications, 2023, 59(7): 285-293.
SunTieqiang, WeiGuanghui, SongChao, et al. Multi-modal rail surface defect detection method based on YOLO[J]. Electronic Measurement Technology, 2024, 47(21): 72-81.
[27]
CaiXinhao, LaiQiuxia, WangYuwei, et al. Poly Kernel Inception Network for Remote Sensing Detection[C]//2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 17-21, 2024, Seattle, WA, USA. New York: IEEE Press, 2024: 27706-27716.
[28]
ZhangYin, YeMu, ZhuGuiyi, et al. FFCA-YOLO for small object detection in remote sensing images[J]. IEEE Transactions on Geoscience and Remote Sensing, 2024, 62: 5611215.
[29]
HuJie, ShenLi, SunGang. Squeeze-and-excitation networks[C]//2018 IEEE/CVF IEEE Conference on Computer Vision and Pattern Recognition, June 18-22, 2018, Salt Lake City, UT, USA. New York: IEEE Press, 2018: 7132-7141.
WanDahang, LuRongsheng, FangYang, et al. YOLOv11-RGBT: towards a comprehensive single-stage multispectral object detection framework[EB/OL]. (2025-06-18).
[32]
ShenJifeng, ChenYifei, LiuYue, et al. ICAFusion: iterative cross-attention guided feature fusion for multispectral object detection[J]. Pattern Recognition, 2024, 145: 109913.
[33]
ChengXiaolong, GengKeke, WangZiwei, et al. SLBAF-Net: super-lightweight bimodal adaptive fusion network for UAV detection in low recognition environment[J]. Multimedia Tools and Applications, 2023, 82(30): 47773-47792.