PSI - Issue 84

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

316

Fig. 1. Framework for defect mapping and image annotation.

The right-hand side of the figure shows the key step of the proposed Computer Vision (CV) framework: the unification of all damage types into a single joint defect table, obtained by merging the defect lists extracted from multiple inspection forms. This process results in a common taxonomy of 46 damage classes, where each defect is uniquely defined, consistently coded, and linked to the structural elements in which it may occur. In this unified representation, damage phenomena such as corrosion-related cracking, spalling, moisture stains, reinforcement exposure, and deformation are no longer tied to a specific element form, but become part of a shared semantic space. This unified defect taxonomy is a fundamental enabler for AI-based inspection. It allows the deep learning model to be trained on images belonging to different RC and PRC elements while relying on a single, coherent classification scheme, rather than multiple element-specific label sets. As a result, the Computer Vision system can recognize similar degradation mechanisms across beams, slabs, piers, and abutments in a consistent manner, improving robustness, scalability, and interpretability of the automated damage detection process. This unified taxonomy was paramount also in the training process in order to reduce fragmentation and duplication of defects across several structural elements in order to obtain a more concise class definition table. As can be seen in Fig. 1, the presence of unified defect table enables the image annotation across the selected image dataset.