Data & Methodology

Crack detection uses a deep-learning segmentation model trained on concrete surface datasets. Reported metrics (recall 91%, draft time 3 days to 4 hours) were measured in our pilot inspection of one bridge under documented conditions. Width estimates convert pixel measurements using the zone GSD; without GSD, values remain in pixels. Limitations: performance varies with lighting, surface condition and image quality. These results are an inspection aid; final judgement is the responsibility of a qualified engineer.