PSI - Issue 84

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

1348

Keywords: landslides; highway bridges; InSAR; FEM simulation; infrastructure monitoring.

1. Introduction In recent decades, a large portion of the existing tunnel stock in Italy has approached the end of its service life. Most of these structures were built before the 1980s using construction techniques that often-lacked waterproofing systems. Age-related deterioration, evolving operational loads, and environmental exposure make structural condition assessment a central aspect of tunnel asset management and risk mitigation. In this framework, the integration of AI based diagnostic tools into traditional inspection workflows represents a valuable opportunity to support systematic and repeatable damage recognition. According to the Italian Ministry of Infrastructure and Transport guidelines (MIMS, 2022), both Structural Local (SLO) and Structural Global and Geotechnical (SGG) vulnerability are related to surface conditions and to deep-lining defects, such as cracks, cavities, and reductions in lining thickness. Among surface defects, cracks are a primary indicator of structural vulnerability, as they may reflect durability issues, local structural damage, or unfavourable ground–structure interaction. Large-scale crack identification is commonly based on manual interpretation of images obtained from photogrammetric surveys or laser-scanner–derived orthophotos. Although these techniques allow full lining documentation, interpretation remains time-consuming, labour-intensive, and strongly dependent on operator expertise, introducing subjectivity and limiting scalability at a national level. Recent studies (Zhou et Al, 2023; Zhenqing et Al, 2019) have explored the use of Convolutional Neural Networks (CNNs) for automatic crack detection; however, most rely on ad-hoc, high-resolution datasets with limited variability, reducing model generalisation and applicability to real inspection scenarios. This study introduces a CNN-based methodology for crack recognition on tunnel linings using greyscale orthophotos derived from laser-scanner acquisition systems, at resolutions compatible with nationwide inspection programmes. The work focuses on comparing two major computer vision tasks—object detection (OD) and semantic segmentation (SS)—to assess their suitability for large-scale tunnel crack detection and to investigate how tunnel conditions influence model performance, with the aim of laying the foundations for a scalable prototype system. 2. Dataset and methods The dataset consists of greyscale orthophotos representing complete tunnel linings derived from laser-scanner surveys. Visible cracks were manually annotated by technical specialists during national Tunnel Assessment Campaigns (MIMS Guideline, 2020) and provided as vectorial polylines with tunnel x–y references, forming the supervised learning ground truth. In addition to the image–label pairs, metadata were introduced to support result interpretation and to characterise dataset quality. These include: (i) IQOA static and hydraulic indices describing structural degradation and water infiltration conditions, derived from consolidated engineering assessment procedures; (ii) a label quality and noise condition index, defined to quantify annotation reliability and the level of visual disturbance affecting crack recognisability. Noise sources include illumination artefacts, tunnel equipment, surface deterioration, and staining related to water infiltration. Both indices are defined on a four-level qualitative scale. 3. Object detection vs semantic segmentation Two CNN-based approaches were investigated: object detection and semantic segmentation. Object detection was implemented using a fine-tuned YOLOv5 network, which provides bounding-box representations of cracks with high computational efficiency and global contextual awareness. Semantic segmentation was performed using a U-Net trained from scratch and a pre-trained DeepLabV3 network, classify each pixel as crack or background and enabling detailed geometric reconstruction at higher computational cost.