Crack inspection moves past pixels: IAC turns masks into millimeter widths and a continuous severity score

Image-based crack analysis moves from marking crack pixels to reporting widths in millimeters and a continuous severity score that can be aligned with inspection standards. Most prior image methods stop after locating crack pixels, leaving width to be measured on site with calipers or comparator cards, and the few severity methods sort images into discrete classes without calibrated physical geometry. Across 41 digital-caliper reference widths, estimated widths differed from caliper values by 7.33% on average, with 6.29% for the 29 linear-crack locations and 9.84% for the 12 web-crack locations, all below the 10% target set for the recommended resolution range.

The framework first segments cracks with a vision foundation model, then skeletonizes the mask and applies a distance transform for local width, and finally converts pixels to millimeters using ground sampling distance. The segmentation backbone is the Tiny variant of SAM2, and training updates only the affine scale and shift parameters of LayerNorm modules while all other weights stay frozen, so far fewer parameters are trained than in full fine-tuning. On OmniCrack30k, SACv2-Tiny reaches 62.49% F1 and 45.45% IoU with 38.9M total parameters and 9.94 FPS, improving F1 by 1.27 percentage points over the SAC baseline while cutting parameters by 56.8% and raising throughput by 34.9%.

The severity score combines maximum crack width and affected-area ratio into a bounded continuous score from 0 to 100, with the width term anchored to half the critical width specified by the applicable standard. Width alone does not capture the added risk of branching cracks, while the segmented crack-area ratio carried negligible discriminatory information over the observed range and was therefore removed. Across the ten measured cracks, five wide linear cracks fell in the Critical band and the two narrowest in Medium, while all three web cracks were Critical with scores from 53.4 to 58.7 because of their larger affected-area ratios.

The pipeline targets concrete surface crack inspection, takes images from handheld cameras, mobile robots or drones, and outputs calibrated crack width, affected area and a continuous severity score from 0 to 100. Width measurement is reliable at the recommended ground sampling distance or finer, which can be reached by adjusting stand-off distance and focal length. Score thresholds and exponents can be recalibrated to align with other inspection regulations, including pavement-specific manuals.

The severity score currently considers only maximum crack width and spatial extent, leaving crack orientation and pattern type outside the assessment. Score thresholds, exponents and reference width are configurable, so the current Low, Medium and Critical bands apply to this concrete inspection scenario. Width error rises markedly as ground sampling distance coarsens and masks become more fragmented, with only part of the crack recovered as four disconnected components at 0.303 mm/px. Data, models and code are available from the corresponding author upon reasonable request.