A Calibrated Post-Processing Layer for Tolerance-Compliant Crack-Width Measurement

Crack width governs how inspectors rate and prioritize concrete deterioration, yet automated pipelines are still judged by pixel-overlap scores (IoU, Dice) that do not explicitly target width measurement. We show geometrically that high pixel overlap can coexist with substantial width error and experimentally demonstrate a systematic over-segmentation bias across modern crack-segmentation algorithms. We present Parallel Consensus Fusion (PCF), a deliberately lightweight measurement-reliability layer, applicable to any algorithm whose operating point over-segments, that corrects this bias by intersecting the learned mask with an independently derived contrast-enhanced threshold mask; its few parameters are fixed once per algorithm and domain through a coverage-constrained calibration rather than assumed universal. Evaluated across ten segmentation algorithms on an uncontrolled field set (75 caliper-measured physical regions of interest, ROIs, on 15 cracks, imaged at three distances for 225 image-ROI observations) and an external dataset of 19,098 width-labeled ROIs, PCF behaves as a calibrated over-segmentation corrector whose benefit is predictable from each algorithm's over-segmentation bias. It raises the within-tolerance ROI rate-the share of regions whose measured width falls within an engineering tolerance-for nine of ten algorithms and collapses architecturally diverse algorithms into a single width-accuracy band. We argue that the within-tolerance ROI rate, not overlap, is the metric that matters for inspector-grade crack-width measurement; we evaluate it primarily at a strict ±0.10 mm band, with the looser AASHTO allowable-error bands as robustness checks.