PSI - Issue 84

Available online at www.sciencedirect.com

ScienceDirect

Procedia Structural Integrity 84 (2026) 1347–1352

© 2026 The Authors. Published by ELSEVIER B.V. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0) Peer-review under responsibility of the scientific committee of the Conference Abstract The Structural Local Condition of tunnel linings is primarily influenced by both surface and subsurface features, and their complex interaction. These aspects are typically assessed using laserscanner and GPR surveys. Among the various surface defects detected by laserscanner, cracks represent one of the most critical indicators of structural degradation, and their automated and detailed detection is a major challenge in civil engineering diagnostics. This study focuses on a comparative analysis of two major deep learning tasks for computer vision, object detection and semantic segmentation, with the objective of assessing their suitability for large-scale tunnel lining crack detection and aims to establish the foundations for a scalable crack recognition framework. The system components include pre-trained models for semantic segmentation and object detection, as well as clustering algorithms. These are used to automatically recognize portions of cracks at the pixel level and progressively at higher levels, where the cracks are identified and localized using supervised grayscale images. The analysis is complicated by the inherent heterogeneity of tunnel environments, characterized by different construction methods, concrete decay patterns, and maintenance conditions. Standard pre trained CNN-based models, such as DeepLabV3 for semantic segmentation and YOLOv5 for object detection, were tested; however, their performance was found to be strongly influenced by image quality, dataset variability, and the need for complex pre-processing. The outcomes support the adoption of segmentation-based approaches integrated with hybrid solution methods such as clustering and object detection, that merge different techniques which can be parameterized and scaled at various levels. The proposed method advances the state of the art in automated tunnel inspection by demonstrating how deep learning and machine learning techniques can be effectively combined to reduce errors in crack detection, supporting more reliable assessments of tunnel conditions. III Fabre Conference: Existing Bridges, Viaducts, and Tunnels: Research, Innovation, and Applications Object Detection versus Semantic Segmentation for Large-Scale Tunnel Crack Recognition from Laser-Scanner Data Manuel Mancuso a, *, Francesco Carlicchi b , Diana Salciarini a , Filippo Ubertini a , Alfredo Milani c , Valentina Poggioni b a Department of Civil and Environmental Engineering, University of Perugia, Perugia, Italy b Department of Math & Computer Science, University of Perugia, Perugia, Italy c Dipartimento di Scienze e Tecnologie Applicate, Unilink, Roma, Italy

* Corresponding author. E-mail address: manuel.mancuso@dottorandi.unipg.it

2452-3216 © 2026 The Authors. Published by ELSEVIER B.V. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0) Peer-review under responsibility of the scientific committee of the Conference 10.1016/j.prostr.2026.06.172