C
CRS AI
DI
Live Experimental Analysis

Model card

Technical transparency for the crack-like region detection model.

Model
crs-crack-v0.3-exp
Status
Experimental AI POC
Human review
Required

Purpose

Segment crack-like regions in still images of concrete surfaces to accelerate visual inspection triage. Not intended for structural safety conclusions or physical measurement.

Training dataset (summary)

Curated internal dataset (placeholder) of concrete surface images with weak crack annotations. Full dataset documentation is in development.

Evaluation metrics (placeholders)

  • IoU (crack class): 0.42 · illustrative
  • Pixel precision: 0.71 · illustrative
  • Pixel recall: 0.55 · illustrative

Real evaluation on a held-out benchmark is planned for the next iteration.

Supported input range

Well-lit, orthogonal photographs of dry concrete surfaces. Minimum recommended resolution 8 MP. Wet, painted, or heavily textured surfaces are out of scope.

Known limitations

  • No physical width measurement (uncalibrated pixel output only).
  • Confusion with surface staining, joints, and wire marks.
  • Reduced recall on hairline cracks under 1 px.
  • Not evaluated on tunnel or underwater imagery.

Known failure examples

Moisture staining detected as crack-like.
Formwork lines detected as long linear cracks.

Version

crs-crack-v0.3-exp · Updated July 2026.
AI-assisted inspection proof of concept. Outputs require professional human review and do not constitute structural engineering conclusions.