PSI - Issue 84

Manuel Mancuso et al. / Procedia Structural Integrity 84 (2026) 1347–1352

1352

offers computational efficiency and global contextual interpretation, it proved less suitable for crack detection due to the thin, elongated nature of cracks and the strong influence of object-driven false positives. Semantic segmentation demonstrated greater robustness and flexibility, allowing pixel-level reconstruction, multi-scale noise filtering, and integration of post-processing strategies tailored to tunnel imagery. Both approaches were strongly affected by tunnel specific conditions, including construction age, waterproofing systems, surface degradation, illumination, and crack morphology. Overall, semantic segmentation emerged as the most appropriate framework for scalable tunnel crack recognition. Its ability to support geometric reconstruction and spatial mapping along the tunnel alignment makes it particularly suitable for integration into inspection workflows. From an engineering perspective, a conservative system design that prioritises robustness and tolerates a limited number of false positives remains essential for practical deployment in real-world tunnel assessment scenarios. Acknowledgements This study was supported by ALPINA spa. We acknowledge the contributions of the project “Assessment, Risk Classification and Interventions on TUNnel LINings (ARCITUNLIN)” funded by FABRE – “Research consortium for the evaluation and monitoring of bridges, viaducts and other structures” (www.consorziofabre.it/en). Any opinion expressed in the paper does not necessarily reflect the view of the funders. References Cao Vu Dung, Le Duc Anh – Autonomous concrete crack detection using deep fully convolutional neural network. Automation in Construction, 99 (2019) 52-58. F.Foria , M. Calicchio, A. Tarquini, G. Miceli & D. Chiaino, D. Cuccato, S. Rinaldo, G. Bomben, G. Rossetti & A. Allegro - Artificial intelligence and image processing in the MIRET approach for the water detection and integrated geotechnical management of existing mechanized tunnels: Methodology, algorithm and case study (2022) The Evolution of Geotech – 25 years of Innovation MIMS – Ministero delle Infrastrutture e della mobilità Sostenibile - Linee guida per la redazione del progetto di fattibilità tecnica ed economica da porre a base dell’affidamento di contratti pubblici di lavori del PNRR e del PNC (Art. 48, comma 7, del decreto-legge 31 maggio 2021, n. 77, convertito nella legge 29 luglio 2021, n. 108)

PyTorch Contributors. Deeplabv3 with resnet-101 backbone. Accessed: 2025-10-02, 2023. PyTorch Documentation, torchvision package.

Ultralytics, YOLOv5, GitHub repository, https://github.com/ultralytics/yolov5, accessed november 2025.

Zhenqing Liu, Yiwen Cao, Yize Wang, Wei Wang – Computer vision-based concrete crack detection using U-net fully convolutional neural networks. Atuomation in Construction 104 (2019) 129-139. Z. Zhou, J. Zhang, and C. Gong. Hybrid semantic segmentation for tunnel lining cracks based on swin transformer and convolutional neural network. Computer-Aided Civil and Infrastructure Engineering, 38(10):2491–2510, 2023.