A Comparative Study of Local-Global Anomaly-Map Fusion Strategies in EfficientAD

Abstract

Industrial anomaly detection must identify both local structural defects and global logical violations while preserving localization quality and practical inference cost. EfficientAD provides complementary local student-teacher and global autoencoder-student anomaly maps, yet its fixed equal average assumes that both maps are equally reliable at every location. This paper evaluates whether the joint-objective score-preserving spatial gate (A3) improves the fixed equal-averaging baseline (B2) under a Test-isolated protocol containing 14 frozen fusion configurations, three detector seeds, and three fusion seeds where applicable. On the frozen 3,651-image MVTec Logical Constraints Anomaly Detection (MVTec LOCO AD) archive, A3 achieved 83.35% macro image-level area under the receiver operating characteristic curve (AUROC) compared with 83.67% for B2, yielding −0.32 percentage points with a paired 95% confidence interval of [−2.56, +1.74]. Relative to B2, A3 changed structural AUROC by +0.95 points, logical AUROC by −1.58 points, and the area under the saturated per-region overlap curve up to a 0.05 false-positive rate (AU-sPRO@0.05) by −0.88 points. Although A3 added only 769 parameters, its fusion-only median latency was 1.463 ms versus 0.757 ms for B2 at batch size 1 on an RTX 4070 SUPER, a 1.93-fold ratio. A3 therefore did not demonstrate a reliable advantage or meet the pre-specified success criteria. This negative conclusion is limited to EfficientAD, MVTec LOCO AD, and the evaluated structural-proxy supervision.