| 394 | 0 | 328 |
| 下载次数 | 被引频次 | 阅读次数 |
为实现小样本条件下钢轨表面伤损的精准检测，本文提出一种基于稳定扩散（Stable Diffusion）模型和重参数化视觉Transformer(RepViT)优化的钢轨表面伤损检测方法。以少量综合巡检列车实测轨面数据作为数据源，使用Stable Diffusion模型扩增表面伤损样本，形成含有丰富病害样本数据的轨面图像数据集。利用轻量级RepViT模块优化YOLOv11n模型主干网络的特征提取模块（C3k2），通过深度可分离卷积将空间特征融合和通道融合分离，避免传统卷积中通道混合噪声对空间特征的干扰，减少模型的参数冗余，提升模型的特征交互效率，实现钢轨表面伤损的高效精准检测。将本文方法与基于YOLOv7、YOLOv8、YOLOv11n、SPD-YOLOv11、DETR模型的钢轨表面伤损检测方法进行对比试验，结果表明：与其他模型相比，采用本文基于Stable Diffusion模型的检测方法时，召回率、mAP@0.5和mAP@0.5:0.95指标均有明显提高，参数量和计算量均有明显降低。本文方法能够高效区分不同类别的钢轨表面伤损，提高表面伤损检测的精准度和识别速度。
Abstract：To achieve accurate detection of rail surface damage under small sample conditions, this paper proposed a rail surface damage detection method based on the Stable Diffusion model and the optimization of the reparameterized visual Transformer(RepViT). Using a small amount of comprehensive inspection train measured track surface data as the data source, the Stable Diffusion model was used to amplify surface damage samples, forming a track surface image dataset containing rich disease sample data. Utilizing the lightweight RepViT module to optimize the feature extraction module(C3k2) of the YOLOv11n model backbone network, spatial feature fusion and channel fusion were separated through depthwise separable convolution, which could avoid the interference of channel mixing noise on spatial features in traditional convolution, reduce model parameter redundancy, improve model feature interaction efficiency, and achieve efficient and accurate detection of rail surface damage. Comparative experiments were conducted between the method proposed in this paper and the rail surface damage detection methods based on YOLOv7, YOLOv8, YOLOv11n, SPD-YOLOv11 models, and DETR models. The results show that compared with other models, the detection method based on the Stable Diffusion model proposed in this paper had a higher recall rate, mAP@0.5 and mAP@0.5:0.95 index, while the number of parameters and calculations have significantly decreased. This method can efficiently distinguish different types of rail surface damage, improving the accuracy and recognition speed of surface damage.
[1]钟艳春，张玉华，张世红，等.钢轨探伤车高速高集成探轮的研制与试验[J].铁道建筑，2025,65(1):54-58.ZHONG Yanchun,ZHANG Yuhua,ZHANG Shihong,et al.Development and Testing of High Speed and High Integration Wheel Detector for Rail Flaw Detection Vehicle[J].Railway Engineering,2025,65(1):54-58.
[2]李闯，张银花，田常海，等.高速铁路钢轨服役状态及病害整治研究[J].铁道建筑，2020,60(8):126-129,142.LI Chuang,ZHANG Yinhua,TIAN Changhai,et al.Study on Rail Service Status and Disease Treatment of Rail for High Speed Railway[J].Railway Engineering,2020,60(8):126-129,142.
[3]周韶博，尹利钧，杨光，等.高速铁路钢轨硌伤疲劳扩展模拟试验研究[J].铁道建筑，2022,62(9):25-29.ZHOU Shaobo,YIN Lijun,YANG Guang,et al. Simulation Tests for Fatigue Extension of Rail Indentations in High Speed Railway[J].Railway Engineering,2022,62(9):25-29.
[4]吕晶，杨其全，许鑫，等.重载铁路直线段钢轨断裂原因分析[J].铁道建筑，2021,61(11):124-127.LÜJing,YANG Qiquan,XU Xin,et al.Cause Analysis of Rail Fracture in Straight Section of Heavy Haul Railway[J].Railway Engineering,2021,61(11):124-127.
[5]ZHENG S B, ZHONG Q W, CHEN X Q, et al. The Rail Surface Defects Recognition via Operating Service Rail Vehicle Vibrations[J].Machines,2022,10(9):796.
[6]SWAIN M. J,BALLARD D H.Indexing via Color Histograms[C]//IEEE. Proceedings Third International Conference on Computer Vision. Osaka:IEEE,1990:390-393.
[7]OJALA T,PIETIKAINEN M, HARWOOD D. Performance Evaluation of Texture Measures with Classification Based on Kullback Discrimination of Distributions[C]//IEEE.Proceedings of 12th International Conference on Pattern Recognition.Jerusalem:IEEE,1994:582-585.
[8]朱寒，林丽，陈德全，等.基于多方向改进Sobel算子的PCB图像定位校正方法[J].电子测量与仪器学报，2019,33(9):121-128.ZHU Han,LIN Li,CHEN Dequan,et al. PCB Image Orientation Rectification Method Based on Multi-directional Improved Sobel Operator[J]. Journal of Electronic Measurement and Instrumentation,2019,33(9):121-128.
[9]CANNY J.A Computational Approach to Edge Detection[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,1986,6:679-698.
[10]GABOR D. Theory of Communication[J]. Journal of the Institution of Electrical Engineers:PartⅢ：Radio and Communication Engineering,1946,93(26):429-457.
[11]LOWE D G. Distinctive Image Features from Scale-invariant Keypoints[J]. International Journal of Computer Vision,2004,60(2):91-110.
[12]DALAL N,TRIGGS B. Histograms of Briented Gradients for Human Detection[C]//IEEE. IEEE Computer Society Conference on ComputerVision andPattern Recognition(CVPR2005), San Diego:IEEE,2005:886-893.
[13]任盛伟，李清勇，许贵阳，等.鲁棒实时钢轨表面擦伤检测算法研究[J].中国铁道科学，2011,32(1):25-29.REN Shengwei,LI Qingyong,XU Guiyang,et al.Research on Robust Fast Algorithm of Rail Surface Defect Detection[J].China Railway Science,2011,32(1):25-29.
[14]赵宏伟，黄雅平，王胜春，等.一种基于空间滤波的钢轨表面擦伤检测改进算法[J].计算机科学，2014,41(1):130-133,137.ZHAO Hongwei,HUANG Yaping,WANG Shengchun, et al.Rail Surface Defect Detection Algorithm Based on Spatial Filtering[J].Computer Science,2014,41(1):130-133,137.
[15]张琳娜，岑翼刚.铁轨图像的低秩矩阵分解缺陷检测[J].信号处理，2019,35(2):667-675.ZHANG Linna,CEN Yigang.Low-rank Matrix Decomposition for Rail Track Image Defect Detection[J]. Journal of Signal Processing,2019,35(2):667-675.
[16]SHANG L,YANG Q,WANG J,et al.Detection of Rail Surface Defects Based on CNN Image Recognition and Classification[C]//IEEE. Proceedings of 2018 20th International Conference on Advanced Communication Technology(ICACT),Chuncheon:IEEE,2018:45-51.
[17]侯博文，杨晓，高亮.基于深度残差网络的轨道结构病害识别[J].铁道学报，2020,42(8):100-106.HOU Bowen,YANG Xiao,GAO Liang,et al.Identificat ion of Track Structure Diseases Based on Deep Residual Network[J]. Journal of the China Railway Society,2020,42(8):100-106.
[18]CHEN X B,ZHANG H M. Rail Surface Defects Detection Based on Faster R-CNN[C]//IEEE. 2020 International Conference on Artificial Intelligence and Electromechanical Automation(AIEA). Tianjin:IEEE,2020:819-822.
[19]罗晖，徐广隆.基于图像增强与深度学习的钢轨表面缺陷检测[J].铁道科学与工程学报，2021,18(3):623-629.LUO Hui,XU Guanglong. Rail Surface Defect Detection Based on Image Enhancement and Deep Learning[J]. Journal of Railway Science and Engineering,2021,18(3):623-629.
[20]杨佳佳，许贵阳，白堂博，等.基于改进YOLOX的轻量级钢轨表面缺陷检测算法[J].铁道建筑，2023,63(7):34-39.YANG Jiajia,XU Guiyang,BAI Tangbo,et al. Lightweight Rail Surface Defect Detection Algorithm Based on Improved YOLOX[J].Railway Engineering,2023,63(7):34-39.
[21]陈仁祥，潘升，杨黎霞，等.基于改进YOLOv7的钢轨表面缺陷检测[J].铁道工程学报，2024,41(7):18-24.CHEN Renxiang,PAN Sheng,YANG Lixia,et al.Rail Surface Defect Detection Based on Improved YOLOv7[J].Journal of Railway Engineering Society,2024,41(7):18-24.
[22]RIPPEL O, MULLER M, MERHOF D. GAN-based Defect Synthesis for Anomaly Detection in Fabrics[C]//IEEE. 202025th IEEE International Conference on Emerging Technologies and FactoryAutomation(ETFA). Vienna:IEEE,2020:534-540.
[23]ZHANG G , CUI K , HUNG T Y ,et al.Defect-GAN:Highfidelity Defect Synthesis for Automated Defect Inspection[C]//IEEE. 2021 IEEE/CVF Winter Conference on Applications of Computer Vision(WACV 2021),IEEE,2021:2524-2534.
[24]ZHANG H, PAN D, LIU J,et al.A Novel MAS-GAN-based Data Synthesis Method for Object Surface Defect Detection[J].Neurocomputing,2022,499:106-114.
基本信息:
中图分类号:U216.3
引用信息:
[1]刘子煜,刘金朝,孙凯仪,等.基于Stable Diffusion和RepViT-YOLOv11模型的钢轨表面伤损检测方法[J].铁道建筑,2026,66(05):1-10.
基金信息:
中国国家铁路集团有限公司科技研究开发计划（K2022T006）; 中国铁道科学研究院集团有限公司基金（2024YJ295）
2025-10-30
2025
2025-11-29
2026-04-13
2026
1
2026-04-03
2026-04-03
2026-04-03