A vision-language based image measurement framework for quantitative assessment of railway surface defects

Accurate quantification of railway surface defects is essential for infrastructure condition assessment, maintenance prioritisation, and traceable engineering decision-making. This study presents an image-based measurement framework in which binary segmentation serves as one component of a broader defect-extent measurement chain. The primary measurand is the Defect Extent Index (DEI), defined from defect-pixel counts at the inference resolution, with calibration-dependent pathways for approximate physical-area estimation. The segmentation component, FMSegNet, integrates a domain-adapted FLAVA encoder with synthetic multi-scale aggregation and a lightweight decoder for binary defect delineation. On the ProRail validation split (100 images), FMSegNet achieved IoU = 76.42% and Dice = 84.79%, the highest single-run performance among the evaluated trainable baselines; a repeated-seed comparison with SETR indicated competitive performance without statistically consistent superiority. Measurement-oriented evaluation demonstrated strong agreement between predicted and annotation-derived DEI values: Bland–Altman bias = +0.00581, Pearson r = 0.9909, and expanded uncertainty U(DEI) = 0.0543 (k = 2). Repeated inference under identical digital input conditions yielded CV = 0.000%, indicating deterministic inference repeatability; full measurement-chain repeatability, requiring repeated physical acquisition under varying imaging conditions, was outside the scope of this study. Pipeline-level transfer without fine-tuning on RSDD Type-I (n = 16) yielded IoU = 0.517 ± 0.040 and Dice = 0.681 ± 0.035, demonstrating framework portability under domain shift rather than stable cross-domain generalisation.