Provenance-aware image phenotyping links ISO 4628 grades to localized defect burden in epoxy coatings aged by salt spray

Visual standards provide practical descriptors of coating degradation, but ordinal grades compress heterogeneous surface morphologies and do not directly connect local defect evidence with overall image condition. Here, we developed a provenance-aware image phenotyping framework for epoxy-coated 921 A steel after neutral salt-spray exposure. The dataset comprised 68 source images from 68 independently prepared coated panels, each imaged once at its designated endpoint. The framework integrates ISO 4628-linked rusting and blistering grades, whole-image descriptors, localized defect annotations, and model-assisted outputs within a traceable panel/source-image structure. Ri, G, and S were associated but not interchangeable, with Spearman coefficients of 0.71, 0.41, and 0.54 for Ri-G, Ri-S, and G-S, respectively. Image phenotypes complemented the ordinal grades; Ri correlated with brightness (rho = 0.54), dark-pixel fraction (rho = −0.51), and luminance contrast (rho = 0.38). Aggregating 5138 unique bounding-box annotations to the panel/source-image level linked Ri with boxes per tile (rho = 0.75), box-derived area-burden index (rho = 0.78), and empty-tile fraction (rho = −0.65). Because the grades and annotations came from the same photographs, these associations reflect internal consistency in visible-surface description rather than independent validation of corrosion rate, adhesion loss, impedance, or under-film damage. Independent-panel grouped evaluation yielded more conservative model estimates, with no advantage of the enhanced detector configurations over the baseline. The workflow therefore provides a traceable quantitative description of visible coating-surface condition while retaining explicit limits on mechanistic interpretation and external generalization.