Unsupervised Industrial Anomaly Detection and Localization via Anomaly Score Fusion of Dual-Layer DINOv3 Features

Unsupervised industrial anomaly detection uses normal training images to identify and localize defects. The challenge addressed here is localizing small defects amid normal appearance variation without requiring different feature layers to share one normal reference. We propose dual-layer DINOv3 anomaly-score fusion (DT-FS), which allows the two layers to retrieve different normal references before their anomaly maps are combined. Separate memory banks are constructed from the fourth-to-last and last transformer blocks. The two distance maps are normalized and combined with fixed equal weights, without fine-tuning the backbone or training a fusion module. Under a matched five-seed protocol, DT-FS improves pixel-level average precision over feature concatenation by 2.678 and 3.920 percentage points on MVTec AD and VisA, and by 6.470 points on supplementary MPDD. An eight-configuration ablation supports the selected layer pair, while a DINOv2 comparison shows positive fusion gains on all three datasets. A PaDiM baseline, normalization analysis and runtime measurements further assess the method. The results support independent retrieval followed by score fusion for pixel-level defect localization, while showing that image-wise normalization can weaken image-level detection and that separate retrieval increases inference time.