PSI - Issue 84

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

317

3. Case study: Computer-Vision-based damage detection on eight PRC bridges The study involved eight girder bridges with I-PRC section beams, characterized by span numbers ranging from 3 to 23 and cross-section configurations from 3 to 9 girders (Santarsiero et al. 2025). All the bridges were located in the Basilicata region (southern Italy). For these bridges, a total of 1,993 images were available from inspection campaigns. These images were not captured for AI purposes, meaning they often represented panoramic scenes with oblique views and overlapping frames. Moreover, variable lighting and focus were observed, resulting low-quality imagery. This dataset therefore provides a realistic test environment for AI-assisted inspection.

Table 1. Image annotation. Class Id Class name

Number of annotations

0 1 2 3 4 5 6 7 8 9

Rebar oxidation / corrosion (PRC) Concrete delamination / spalling (PRC)

2,259

545

Concrete delamination – top surface (PRC) 135

Crack (generic)

47

Anchorage failure of safety barrier Corrosion-related cracking (PRC)

7

260 137

Active moisture stains Passive moisture stains

1,023

Exposed and corroded stirrups (RC) Exposed and corroded stirrups (PRC) Runoff / water streaks (generic defect)

1

897 550 215

10 11

Honeycombing (PRC)

Total

6,086

As mentioned, the LG2020 inspection forms include 46 different defect classes for structural components. However, only 12 defect categories were actually observed in the analyzed dataset. These were extracted and used for model annotation. Manual annotation involved the use of 436 training images, producing 6,086 labelled instances followed by data augmentation (blur and exposure variation), increasing robustness against poor image quality. Training was performed on the augmented dataset, while 258 images (belonging to 4 spans of the longest bridge in the dataset) were used for testing. The goal was not only to detect defects, but also to verify whether the model correctly identified the most severe defect affecting each structural element, since this is the key driver for Class of Attention assignment. In panoramic views showing several structural elements, it is almost impossible for the model to predict the precise location of every single defect. Therefore, the task should not be interpreted as a conventional object detection problem aimed at the accurate prediction of both class and spatial location of defects. Instead, it can be more appropriately framed as an image-level classification task focused on identifying the damage mechanisms actually present in the inspected structural elements, since these directly govern the condition assessment of the bridge. Consequently, traditional performance metrics such as Precision, Recall, and F1-score cannot be evaluated on the basis of Intersection over Union (IoU), which would require an exact spatial correspondence between predicted and ground truth defect locations. In fact, the model inference applied to a test dataset of about 50 images not used in the training process revealed Recall and Precision values around 0.7 when the IoU was set the minimum value of 0.1. This is in order to attribute a lower importance to the defect localization. This demonstrates that, if used to assess the defect presence, the model behaves rather well, apart from the possible improvements that can be made through the dataset enhancement. 4. Results and discussion The case study confirms that deep learning-based object detection represents a promising tool for supporting bridge inspection activities, particularly when a code-conforming defect taxonomy is embedded in the classification logic. On the four spans analyzed, the pictures were organized in separate folders and inferred through the trained YOLO model that proved capable of: