铁道科学与工程学报2025，Vol.22Issue(4)：1840-1852,13.DOI:10.19713/j.cnki.43-1423/u.T20241087
DETR-MCA:基于探地雷达图像的隧道衬砌内部缺陷的智能检测算法
DETR-MCA:intelligent detection algorithm for tunnel lining defects based on GPR images
摘要
Abstract
The detection of internal defects of tunnel lining is very important for the safe operation of tunnels.To address the problems of traditional ground penetrating radar(GPR)image recognition methods,such as high requirements for technicians,difficulty in complex image recognition,and inability to realize end-to-end recognition,this paper proposed an automatic detection algorithm for hidden defects in radar images based on transformer framework,named DETR-MCA.A novel and efficient multi-scale attention module,Multi-scale convolutional block attention module(MCA),was embedded in the original end-to-end object detection with transformers(DETR)architecture.It allows the model to dynamically focus on the most useful features according to task requirements and content context,so as to improve the convergence speed of DETR model and the detection accuracy of small targets.Finally,the parallel computing capability of the encoder decoder structure based on global attention computing was used to realize the end-to-end problem of cavity and uncompacting defect detection and identification.In order to solve the problem of scarcity of the measured data set,the GPR-measured tunnel defect data set was constructed.Through the traditional data enhancement method,a total of 1 427 cavity data and 669 uncompacting defect data were obtained.Second,to bolster the robustness of the predictive model,transfer learning was employed using measured GPR datasets from two tunnel components:rebars and steel arches,encompassing a total of 5,333 images.Under the same experimental conditions,the proposed model was compared with six deep learning network models:Faster R-CNN,YOLOv3,YOLOv8,Mask R-CNN,RMTDet,and DINO.The proposed model achieved an average precision of 97.1%for different hidden defects in the tunnels,outperforming other models.The results show that DETR-GPR model has high recognition accuracy and strong robustness,which can provide reference for the detection of different defects in complex tunnel environments.关键词
探地雷达/隧道衬砌缺陷/深度学习/目标检测/TransformerKey words
ground penetrating radar/tunnel lining/deep learning/object detection/Transformer分类
信息技术与安全科学引用本文复制引用
侯斐斐,张智轩,崔广炎,樊欣宇,吕飞..DETR-MCA:基于探地雷达图像的隧道衬砌内部缺陷的智能检测算法[J].铁道科学与工程学报,2025,22(4):1840-1852,13.基金项目
自主式交通系统运行及环境状态全息感知技术(SQ2022YFB4300022) （SQ2022YFB4300022）
中南大学教育教学改革项目(2023CG020) （2023CG020）
