基于双注意力和GSSN轻量化的钢轨扣件缺陷检测

doi:10.19678/j.issn.1000-3428.0068491

摘要/Abstract

摘要：

铁路钢轨扣件的缺陷检测是铁道安全巡检中极为重要的工作之一。为提高钢轨扣件维护工作的效率, 基于深度学习的方式进行巡检。而使用当前主流的目标检测模型进行钢轨扣件缺陷的检测时, 模型体积大、参数量多等因素导致无法同时平衡检测准确度和速度。采用压缩与激活(SE)注意力机制与坐标注意力(CA)机制组成的双注意力机制对YOLOv5模型进行改进; 重新设计网络, 选用MobileNetv3作为主干网络, 同时引入含有GSConv模块的Slim-Neck结构和轻量级上采样算子, 以降低计算成本; 将YOLOv5的坐标损失函数修改为SIoU, 以提升训练时的收敛速度, 使得模型更加轻量化。使用改进后的模型在钢轨扣件数据集上进行测试, 结果显示, 单张扣件图片的检测时间为53.8 ms, 检测速度为17.9帧/s, 并且模型大小仅有8.3 MB, 符合模型体积小、检测效果佳的要求。

关键词: 钢轨扣件, 目标检测, YOLOv5, 注意力机制, 卷积网络, 轻量化

Abstract:

The defect detection of railway rail fasteners is a critical task in railway safety inspections. To improve the efficiency of rail fastener maintenance, a deep learning-based approach is implemented for inspections. When using current major object detection models to detect defects in rail fasteners, the large size and numerous parameters of the models make it challenging to balance detection accuracy and speed. The YOLOv5 model is improved by adopting the Squeeze-and-Excitation (SE) attention mechanism and the Coordinate Attention (CA) mechanism. The network is redesigned, with MobileNetv3 selected as the backbone network. Additionally, a Slim-Neck structure incorporating the GSConv module and a lightweight upsampling operator are introduced to reduce computational cost. Finally, the coordinate loss function of YOLOv5 is modified to SIoU to enhance convergence speed during training and make the model more lightweight. The improved model is tested on the rail fastener dataset. Results show that the detection time for a single fastener image is 53.8 ms, with a detection speed of 17.9 frames per second and a model size of only 8.3 MB. These results satisfy the requirements for small model size and effective detection.

Key words: rail fastener, target detection, YOLOv5, attention mechanism, convolutional network, lightweight

张元, 吕德芳, 孟建军, 祁文哲. 基于双注意力和GSSN轻量化的钢轨扣件缺陷检测[J]. 计算机工程, 2025, 51(2): 289-299.

ZHANG Yuan, LÜ Defang, MENG Jianjun, QI Wenzhe. Defect Detection of Rail Fasteners Based on Double Attention and GSSN Lightweight[J]. Computer Engineering, 2025, 51(2): 289-299.

https://www.ecice06.com/CN/Y2025/V51/I2/289

图/表 20

图1 SE注意力机制工作原理

Fig.1 Working principle of SE attention mechanism

图2 CA机制工作原理

Fig.2 Working principle of CA mechanism

图3 双注意力机制结构

Fig.3 Structure of dual-attention mechanism

图4 上采样算子CARAFE的结构

Fig.4 Structure of upsampling operator CARAFE

图5 经GSConv方法后的DSC过程

Fig.5 DSC process after GSConv

图6 跨级网络VoV-GSCSP及其瓶颈单元结构

Fig.6 Structures of VoV-GSCSP and bottleneck

图7 改进后的YOLOv5模型结构

Fig.7 Model structure of modified YOLOv5

图8 距离损失和角度损失的计算示意图

Fig.8 Schematic diagrams of calculations of distance loss and angle loss

图9 模型训练过程中各项指标变化趋势

Fig.9 Change trend of each index in process of model training

图10 改进前后模型的P-R曲线

Fig.10 P-R curves of model before and after improvement

图11 使用改进前后的模型的误检情况对比

Fig.11 Comparisons of false detections of pre- and post-improvement models

图12 使用改进前后的模型进行扣件缺陷检测的部分效果对比

Fig.12 Comparisons of partial effects of using pre- and post-improvement models for fastener defect detections

参考文献 25

1	翁氏根, 赫丹, 王阳, 等. 扣件失效对城市轨道交通列车-浮置板式轨道系统竖向振动响应的影响. 铁道科学与工程学报, 2008, 5 (2): 29- 33.
	WENG S G , HE D , WANG Y , et al. Influence of fastener failure on responses of vertical vibration of urban rail transit train-floating slab track system. Journal of Railway Science and Engineering, 2008, 5 (2): 29- 33.
2	GIBERT X , PATEL V M , CHELLAPPA R , et al. Deep multitask learning for railway track inspection. IEEE Transactions on Intelligent Transportation Systems, 2016, 18 (1): 153- 164.
3	MAIWALDL D , FASS U , LITSCHKE H . Railcheck system: automated optoelectronic inspection of rail systems. Der Eisenbahningenieur, 1998, (7): 33- 37.
4	张未. 德国RAILCHECK光电式自动化钢轨检测系统在轨道检查车中的应用. 哈尔滨铁道科技, 2001, (4): 3- 4.
	ZHANG W . Application of German RAILCHECK photoelectric automatic rail detection system in track inspection vehicles. Harbin Railway Technology, 2001, (4): 3- 4.
5	STELLA E, MAZZEO P, NITTI M, et al. Visual recognition of missing fastening elements for railroad maintenance[C]//Proceedings of the IEEE 5th International Conference on Intelligent Transportation Systems. Washington D. C., USA: IEEE Press, 2002: 94-99.
6	贾利红. 基于DSP的高速轨道扣件检测算法研究及系统设计[D]. 成都: 电子科技大学, 2014.
	JIA L H. Research of the detection algorithm and design the system on high speed railway fastener based on DSP[D]. Chengdu: University of Electronic Science and Technology of China, 2014. (in Chinese)
7	钱广春. 基于计算机视觉的铁路扣件缺失快速探测方法研究[D]. 上海: 上海交通大学, 2011.
	QIAN G C. Research of the rapid detection method on missing railway fastener based on computer vision[D]. Shanghai: Shanghai Jiao Tong University, 2011. (in Chinese)
8	李彧雯. 轨道部件图像检测系统设计[D]. 南昌: 南昌大学, 2012.
	LI Y W. The design of railway components image detection system[D]. Nanchang: Nanchang University, 2012. (in Chinese)
9	WEI X K , YANG Z M , LIU Y X , et al. Railway track fastener defect detection based on image processing and deep learning techniques: a comparative study. Engineering Applications of Artificial Intelligence, 2019, 80, 66- 81. doi: 10.1016/j.engappai.2019.01.008
10	史燕燕, 史殿习, 乔子腾, 等. 小样本目标检测研究综述. 计算机学报, 2023, 46 (8): 1753- 1780.
	SHI Y Y , SHI D X , QIAO Z T , et al. A survey on recent advances in few-shot object detection. Chinese Journal of Computers, 2023, 46 (8): 1753- 1780.
11	CHEN J W , LIU Z G , WANG H R , et al. Automatic defect detection of fasteners on the catenary support device using deep convolutional neural network. IEEE Transactions on Instrumentation and Measurement, 2018, 67 (2): 257- 269. doi: 10.1109/TIM.2017.2775345
12	李嘉新, 侯进, 盛博莹, 等. 基于改进YOLOv5的遥感小目标检测网络. 计算机工程, 2023, 49 (9): 48, 256-264. URL
	LI J X , HOU J , SHENG B Y , et al. Remote sensing small object detection network based on improved YOLOv5. Computer Engineering, 2023, 49 (9): 48, 256-264. URL
13	BOCHKOVSKIY A, WANG C Y, MARK LIAO H Y. YOLOv4: optimal speed and accuracy of object detection[EB/OL]. (2020-04-23)[2023-09-07]. https://arxiv.org/abs/2004.10934.
14	REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: unified, real-time object detection[EB/OL]. (2015-06-08)[2023-09-07]. https://arxiv.org/abs/1506.02640.
15	HEGISTE V, LEGLER T, RUSKOWSKI M. Federated ensemble YOLOv5: a better generalized object detection algorithm[EB/OL]. (2023-06-30)[2023-09-07]. https://arxiv.org/abs/2306.17829.
16	HU J, SHEN L, ALBANIE S, et al. Squeeze-and-excitation networks[EB/OL]. (2017-09-05)[2023-09-07]. https://arxiv.org/abs/1709.01507.
17	HOU Q B, ZHOU D Q, FENG J S. Coordinate attention for efficient mobile network design[EB/OL]. (2021-04-04)[2023-09-07]. http://export.arxiv.org/abs/2103.02907v1.
18	HOWARD A, SANDLER M, CHU G, et al. Searching for MobileNetV3[EB/OL]. (2019-05-06)[2023-09-07]. https://arxiv.org/abs/1905.02244.
19	LI Y D , HAN Z Q , XU H Y , et al. YOLOv3-Lite: a lightweight crack detection network for aircraft structure based on depthwise separable convolutions. Applied Sciences, 2019, 9 (18): 3781. doi: 10.3390/app9183781
20	CHOLLET F. Xception: deep learning with depthwise separable convolutions[EB/OL]. (2016-10-06)[2023-09-07]. https://arxiv.org/abs/1610.02357.
21	陈智超, 焦海宁, 杨杰, 等. 基于改进MobileNet v2的垃圾图像分类算法. 浙江大学学报(工学版), 2021, 55 (8): 1490- 1499.
	CHEN Z C , JIAO H N , YANG J , et al. Garbage image classification algorithm based on improved MobileNet v2. Journal of Zhejiang University (Engineering Science), 2021, 55 (8): 1490- 1499.
22	HU J , WANG Z B , CHANG M J , et al. PSG-Yolov5: a paradigm for traffic sign detection and recognition algorithm based on deep learning. Symmetry, 2022, 14 (11): 2262.
23	LI H L, LI J, WEI H B, et al. Slim-neck by GSConv: a better design paradigm of detector architectures for autonomous vehicles[EB/OL]. (2022-06-06)[2023-09-07]. https://www.researchgate.net/publication/361135191_Slimneck_by_GSConv_A_better_design_paradigm_of_detector_architectures_for_autonomous_vehicles.
24	WANG B N. A parallel implementation of computing mean average precision[EB/OL]. (2022-06-19)[2023-09-07]. https://arxiv.org/abs/2206.09504v1.
25	GEVORGYAN Z. SIoU loss: more powerful learning for bounding box regression[EB/OL]. (2022-05-25)[2023-09-07]. https://arxiv.org/abs/2205.12740.

[1]	黄舒怡, 谭光. 基于分区的高效视频目标检测[J]. 计算机工程, 2025, 51(2): 65-77.
[2]	张欢, 王晨, 单景东, 仇润鹤. 基于领域自适应与注意力机制的电梯安全风险预测[J]. 计算机工程, 2025, 51(2): 86-93.
[3]	孙浩淼, 李宗民, 肖倩, 孙文洁, 张雯欣. AI-Curling: 一种冰壶现场分析与决策方法[J]. 计算机工程, 2025, 51(2): 102-110.
[4]	刘圣杰, 何宁, 王鑫, 于海港, 韩文静. 基于轻量级高分辨率网络的人体姿态估计算法[J]. 计算机工程, 2025, 51(2): 278-288.
[5]	许明, 屈泰澎, 姜彦吉. 改进YOLOv7在复杂场景下的交通标志检测算法[J]. 计算机工程, 2025, 51(2): 335-343.
[6]	安国成, 王晓龙, 江波, 幸健. 复杂环境下高速服务区禁停检测算法[J]. 计算机工程, 2025, 51(2): 356-364.
[7]	张兴鹏, 何东, 杨模, 叶杭滨. 基于多尺度注意力和数据增强的细胞核分割[J]. 计算机工程, 2025, 51(2): 387-396.
[8]	李海丰, 刘森森, 王怀超, 李南莎, 张艺凡. 融合关联关系推理的机场道面地下病害检测算法[J]. 计算机工程, 2025, 51(2): 397-406.
[9]	罗旭东, 袁笛, 常晓军, 何震宇. 基于不确定性启发图像增强的水下目标跟踪[J]. 计算机工程, 2025, 51(1): 11-19.
[10]	胡升龙, 陈彬, 张开华, 宋慧慧. 场景结构知识增强的协同显著性目标检测[J]. 计算机工程, 2025, 51(1): 31-41.
[11]	周雪阳, 傅启明, 陈建平, 陈延明, 陆悠, 王蕴哲. 基于证据和图推理的文档级关系抽取方法: 以医学关系为例[J]. 计算机工程, 2025, 51(1): 106-117.
[12]	肖超恩, 李子凡, 张磊, 王建新, 钱思源. 基于Transformer模型与注意力机制的差分密码分析[J]. 计算机工程, 2025, 51(1): 156-163.
[13]	张倡倡, 吕卫东, 蔡子杰, 刘炎奎. 基于域泛化的轻量化图像分类算法[J]. 计算机工程, 2025, 51(1): 182-189.
[14]	杨红菊, 吉昌. 学习驱动的图像压缩算法研究[J]. 计算机工程, 2025, 51(1): 190-197.
[15]	阳丽莎, 李茂军, 胡建文, 王鼎湘. 基于改进YOLOv7-tiny的带钢表面缺陷检测算法[J]. 计算机工程, 2025, 51(1): 208-215.

选择文件类型/文献管理软件名称

选择包含的内容