PSI - Issue 84

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

318

• correctly identifying the most severe damage in many cases; • recognizing corrosion-related damage consistently with human reports; • operating robustly even under heterogeneous visual conditions.

As can be seen from Table 2, related to one out of the four analyzed spans, the model was able to correctly identify the maximum severity defect on all the structural elements matching the results of human inspection. However, defects like deformed longitudinal bars and transverse cracks were not recognized automatically due to many factors.

Table 2. Human and Computer Vision defect detection on a bridge span. Severity

Maximum severity defects

Element

Human inspection YOLO

Human inspection

YOLO

Rebar oxidation, Deformed long. rebar, Transverse crack Rebar oxidation, Deformed long. rebar Rebar oxidation, Transverse crack Rebar oxidation, Transverse crack Rebar oxidation, Transverse crack Rebar oxidation, Transverse crack

Maximum value

5

5

Rebar oxidation

Piers

5 5 5 5 5

5 5 5 5 5

Rebar oxidation Rebar oxidation Rebar oxidation Rebar oxidation Rebar oxidation

Cap beams

Beams

Transverse beams

Slabs

However, some limitations emerged. First, inspection images are frequently not optimized for AI analysis. Panoramic shots containing several elements reduce detectability accuracy, while low resolution, blur, and lighting variations reduce model confidence. This suggests that basic photographic guidance for inspectors could significantly enhance AI performance without modifying standard workflows. Second, human inspection is not error-free. Some discrepancies between YOLO output and inspection forms were found to derive from subjective or incorrect human assessments, reinforcing the potential of AI as a consistent support tool rather than a replacement technology.

Fig. 2. Example of inspection image after the YOLO inference.