PSI - Issue 84

Giuseppe Santarsiero et al. / Procedia Structural Integrity 84 (2026) 313–320

315

• improve repeatability, scalability, and speed of inspection workflows. Such AI assessment techniques can help in the application of the latest regulatory framework, which foresees specific inspection forms for each bridge structural component (Santarsiero et al. 2023). The present paper focuses primarily on the damage detection module for RC and PRC bridges, illustrating the methodology and results obtained from a dataset of eight girder bridges inspected through conventional procedures. The work investigates the feasibility of applying YOLO-based object detection to classify degradation phenomena according to the Italian code taxonomy, assessing both the capability of the model and the limitations imposed by real inspection imagery. This goes beyond the identification of just a few defects like corrosion and cracks as in other studies (Liu et al. 2025). At the same time, the paper briefly frames this module within the broader inspection platform, where a second AI-assisted tool supports bearing identification and code-conforming defect interpretation. This highlights the potential of integrated AI-assisted workflows to improve the overall consistency of bridge inspection activities. 2. Methodology: a platform for AI-assisted bridge visual inspection The AI-assisted platform under development at the University of Basilicata is conceived as a modular system, where different AI components act on the inspection images to extract meaningful, code-compliant information. The platform currently includes two fully-operational pillars: (i) Automatic detection and classification of defects in RC/PRC elements using YOLO-based DL models (main subject of this paper); (ii) Intelligent inspection of bridge bearings (APPOGGI framework), based on a DL classifier for bearing type recognition combined with a Large Language Model constrained by engineering rules, which automatically translates visual evidence into LG2020 compliant inspection records. Although technically distinct, both modules share the same logic with an input made by photographs taken during real inspections, processing through AI-based classification aligned with normative taxonomies, providing as output structured data suitable for inspection forms. Thus, the platform is not meant to replace human inspectors, but rather to act as a decision support tool capable of highlighting potential defects, standardizing terminology, accelerating data processing, reducing subjectivity as well as effective storage and repeatable analysis from multiple inspectors at different times. The core component analyzed in this paper is the YOLOv8n object detection model, trained to identify the main damage classes defined in the Italian inspection forms for: RC beams, PRC beams, slabs, piers, columns (for multi column piers), abutments. The model simultaneously performs localization and classification, returning, for each detected object defect type, bounding box, and confidence score. This output is then mapped to the LG2020 inspection sheet fields, enabling direct comparison with human inspection data. In fact, Fig. 1 illustrates the conceptual framework adopted for Computer Vision–based damage detection in reinforced concrete (RC) and prestressed reinforced concrete (PRC) bridge elements, with a specific focus on the harmonization of defect classification across different structural components. On the left-hand side, the original inspection forms prescribed by the Italian Guidelines are shown separately for each element type, including abutments, piers, columns, RC beams, PRC beams, and slabs. As can be noted, dapped end beams are excluded from this analysis since they do not have a dedicated inspection form (Santarsiero and Picciano 2024). Each form contains its own list of defect codes and descriptions, reflecting the traditional element-by-element organization of visual inspections. Moreover, each defect is accompanied by its severity index (G) that can vary from 1 (minimum) to 5 (maximum). High severity defects push the structure towards high structural risk and, therefore, the detection of these defects is of fundamental importance. Although technically correct, this structure leads to fragmentation, redundancies, and difficulties when attempting automated analysis across heterogeneous image datasets.