TOLO-based intelligent detection and classification of high-speed railway tunnel portal types

Abstract

The portal type of high-speed railway (HSR) tunnels has a direct impact on aerodynamic phenomena and operational safety when trains enter and exit. In much of the existing inspection and field measurement data, portal type information is predominantly stored in image form, requiring manual interpretation and experience-based classification to annotate portal configurations. This process is labor-intensive, inefficient, and often inconsistent, limiting the large-scale reuse of available datasets and the parametric modeling of portal configurations. To address the need for automatic portal recognition under complex backgrounds and with significant scale variations, this study proposes an intelligent vision approach based on a pipeline that integrates portal detection with configuration classification, enabling rapid portal localization and automated type identification to provide structured inputs for subsequent aerodynamic response analysis. A lightweight improved detector, termed TOLO, is developed by incorporating the parameter-free spatial attention module (SimAM) and the efficient channel attention (ECA) module into the YOLO11 framework, enabling high-accuracy real-time detection of tunnel portals. Additionally, a YOLO11-based classification model is employed to identify portal types from the detected portal regions, forming an integrated detection–classification framework. A dedicated dataset covering multiple representative portal types is established for both object detection and image classification. Systematic comparative experiments are conducted to evaluate the baseline YOLO11 and attention-enhanced variants in terms of convergence behavior, computational and resource costs, and performance metrics including precision, recall, and mean average precision (mAP). The results demonstrate that, with only marginal increases in parameters and computational complexity, TOLO consistently outperforms the baseline model in detection accuracy and robustness, especially for distant small-scale portals and scenarios involving strong background interference. The adopted classification model reliably discriminates typical portal types —such as end-wall, oblique-cut, and extended end-wall configurations —providing dependable structural-type inputs for subsequent aerodynamic response prediction. Finally, a joint framework encompassing detection, classification, and aerodynamic response prediction is proposed to support intelligent portal recognition and aerodynamic assessment for HSR tunnels.

Keywords

High-speed railway / Tunnel portal type / Object detection / Attention mechanism / Deep learning

References

[1] W. Ma, G. Wen, H. Zhu, J. Han, C. Wang, J. Tian, A. Cheng, Sonic boom and alleviated measurements in Wan’an Tunnel of Beijing—Hong Kong high—speed railway, Tunnel Constr. 44 (2024) 1431-1439, doi: 10.3973/j.issn.2096—4498.2024.07.010.

[2] B. Guo, G. Geng, L. Zhu, H. Shi, Z. Yu, High—speed railway intruding object image generation with generative adversarial networks, Sensors 19 (2019) 3075, doi: 10.3390/s19143075.

[3] H. Niu, T. Hou, Fast detection study of foreign object intrusion on railway track, Arch. Transp. 47 (2018) 79-89, doi: 10.5604/01.3001.0012.6510.

[4] H.R. Wang, P. Dai, J.B. Liu, J. Shi, H.R. Song, Z.C. Gu, Development of a video detection system for foreign object intrusion in high—speed railway environment, Comput. Meas. Control 32 (2024) 86-91, doi: 10.16526/j.cnki.11—4762/tp.2024.10.012.

[5] H.C. Fu, T.B. Bai, G.Y. Xu, H. Zong, J.M. Duan, Crack detection in track slabs based on an improved YOLOv8 algorithm, J. Beijing Jiaotong Univ. 48 (2024) 133-143, doi: 10.11860/j.issn.1673—0291.20240024.

[6] S. Li, T. Shi, F. Jingke, Improved road damage detection algorithm of YOLOv8, Comput. Eng. Appl. 59 (2023) 165-174, doi: 10.3778/j.issn.1002—8331.2306—0205.

[7] B.Q. Xia, L.W. Wang, Z.Y. Shi, Y.Q. Huang, J. Lian, Development and application of automatic monitoring system for structural deformation of tunnel surrounding rock, Railw. Technol. Innov. (2022) 48-53, doi: 10.19550/j.issn.1672—061x.2021.11.30.002.

[8] L.L. Yang, Y.L. Xie, Y.L. Lu, X.L. He, S.Y. Zhao, Application of machine vision in railway tunnel construction, Sci. Technol. Innov. (2025) 40-43, doi: 10.15913/j.cnki.kjycx.2025.07.011.

[9] P. Zhao, W.B. Ma, J. Wang, J.J. Zhang, Multimodal vision—guided method for identifying apparent defects in railway tunnels, Railw. Stand. Des. 69 (2025) 337-346, doi: 10.13238/j.issn.1004—2954.202506060005.

[10] J.Y. Fu, H.Y. Yang, X.R. Liang, H.Y. Wang, S.W. Yang, P. Zhao, Quantitative characterization and identification method of spatial geometric features of tunnel cracks based on panoramic images, J. Railw. Sci. Eng. 22 (2025) 5630-5643, doi: 10.19713/j.cnki.43—1423/u.T20250397.

[11] X. Wu, Y. Jiang, J. Wang, M. Kusaba, T. Taniguchi, T. Yamato, A new health assessment index of tunnel lining based on the digital inspection of surface cracks, Appl. Sci. 7 (2017) 507, doi: 10.3390/app7050507.

[12] Z. Yin, Z. Lei, A. Zheng, J. Zhu, X.Z. Liu, Automatic detection and association analysis of multiple surface defects on shield subway tunnels, Sensors 23 (2023) 7106, doi: 10.3390/s23167106.

[13] T.Y. Zhang, C.Y. Suen, A fast parallel algorithm for thinning digital patterns, Commun. ACM 27 (1984) 236-239, doi: 10.1145/357994.358023.

[14] Z. Tian, C.H. Shen, H. Chen, T. He, FCOS: a simple and strong anchor—free object detector, IEEE Trans. Pattern Anal. Mach. Intell. 44 (2020) 1922-1933, doi: 10.1109/TPAMI.2020.3032166.

[15] F.C. Akyon, S.O. Altinuc, A. Temizel, Slicing aided hyper inference and fine—tuning for small object detection, in: Proceedings of the 2022 IEEE International Conference on Image Processing (ICIP), IEEE, Bordeaux, 2022, pp. 966-970, doi: 10.1109/ICIP46576.2022.9897990.

[16] A.A. Murat, M.S. Kiran, A comprehensive review on YOLO versions for object detection, Eng. Sci. Technol. Int. J. 70 (2025) 102161, doi: 10.1016/j.jestch.2025.102161.

[17] A.F. Rasheed, M. Zarkoosh, YOLOv11 optimization for efficient resource utilization, The J. Supercomput. 81 (2025) 1085, doi: 10.1007/s11227—025—07520—3.

[18] A. Bewley, Z. Ge, L. Ott, F. Ramos, B. Upcroft, Simple online and realtime tracking, in: Proceedings of the 2016 IEEE International Conference on Image Processing (ICIP), IEEE, Phoenix, 2016, pp. 3464-3468, doi: 10.1109/ICIP.2016.7533003.

[19] T.H. Ling, H.N. Chen, K. Zhang, B. Yang, L. Zhang, Lightweight network identification method of rock lithology and its application in tunnel engineering, J. Railw. Sci. Eng. 20 (2023) 3604-3615, doi: 10.19713/j.cnki.43—1423/u.t20221939.

[20] J. Redmon, A. Farhadi, YOLOv3: an incremental improvement, arXiv preprint arXiv:1804.02767, 2018, doi: 10.48550/arXiv.1804.02767.

[21] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.Y. Fu, A.C. Berg, SSD: single shot MultiBox detector, in: Proceedings of the European Conference on Computer Vision (ECCV), Springer, Cham, 2016, pp. 21-37, doi: 10.1007/978—3—319—46448—0_2.

[22] S. Ren, K. He, R. Girshick, J. Sun, Faster R—CNN: towards real—time object detection with region proposal networks, IEEE Trans. Pattern Anal. Mach. Intell. 39 (2017) 1137-1149, doi: 10.1109/TPAMI.2016.2577031.