DAPQ: Defect-Aware Perceptual Quality Metric for Training-Free Augmentation in Industrial Anomaly Detection

Industrial anomaly detection requires rich normal-class representations, yet real defect exemplars constitute less than 1% of production volume. Existing synthetic augmentation is deployed blindly: no principled mechanism predicts whether a synthesized image will improve downstream detector performance, a critical barrier for resource-constrained deployments. We show this failure stems from a quantifiable misalignment between standard image quality assessment (IQA) metrics and anomaly-detector feature spaces, with Spearman correlations indistinguishable from zero (|ρ| < 0.05). We introduce DAPQ (Defect-Aware Perceptual Quality), the first task-aligned synthesis quality metric for industrial anomaly detection, integrating backbone-aligned feature-space distance (Dfeat), VGG-16 Gram texture consistency (Dtex), engineer perceptual ratings (Dhuman), and mask-conditioned structural fidelity (Dstruct). A training-free proxy achieves Spearman ρ = 0.891 with the masked-region defect signal, which strongly predicts downstream AUROC gain. DAPQ-guided coreset augmentation of PatchCore improves mean image-level AUROC by +10.49 pp across 35 category evaluations on MVTec-AD, VisA, BTAD, and MVTec AD 2, with peak gains of +32.3 pp (bottle) and +24.7 pp (vial); with only 20% real data, DAPQ surpasses full-data training on bottle (AUROC 0.960 vs. 0.630).