VerdictAD: Few-Shot Anomaly Detection via Decision-Guided Refinement

Shangquan Sun; Wenjun Luo; Yuanzhi Wang; Ruiqi Li; Ruoyu Chen; Jiachen Liu; Shijian Lu

cs.CVNov 2, 2026

VerdictAD: Few-Shot Anomaly Detection via Decision-Guided Refinement

Authors: Shangquan Sun, Wenjun Luo, Yuanzhi Wang, Ruiqi Li, Ruoyu Chen, Jiachen Liu, Shijian Lu

Organizations: Nanyang Technological University · Michigan State University · University of Michigan

Abstract

Detecting defects in real-world industrial inspection often requires learning from only a few normal images per category. Recent few-shot anomaly detection methods have achieved strong performance by measuring feature differences between query images and normal references. However, visual differences do not always indicate genuine defects, as normal appearance variations can also produce high anomaly scores. Conversely, structural or logical anomalies may not produce strong local feature differences, making them difficult to detect. To address this limitation, we introduce VerdictAD, a training-free framework that refines existing anomaly detectors through visual decisions. Rather than replacing existing detectors, VerdictAD adopts a typed decision model to assess the query image and its local regions against a few normal references. By directly producing structured decision probabilities without text generation, the model provides complementary evidence for refining anomaly responses at different spatial scales. Building on this capability, we introduce two modules: Global Verdict Refinement (GVR) uses image-level decisions to refine the overall anomaly score, while Local Verdict Refinement (LVR) aggregates regional decisions to refine the anomaly map. Combined with a simple DINOv2-based detector, VerdictAD achieves competitive performance in one-shot and few-shot anomaly detection. More importantly, as a plug-and-play framework, it consistently improves existing anomaly detectors across MVTec AD, VisA, and MVTec LOCO AD, achieving strong results across one-shot and few-shot settings. Our code and experimental logs are publicly available at https://github.com/sgstq0811/VerdictAD-Jev-AD.

Figures & tables

Figure 1: Performance improvements with VerdictAD. (a) Anomaly detection and localization performance under the 1-, 2-, and 4-shot settings on MVTec AD and VisA. (b) Image-level and pixel-level AUROC before and after applying VerdictAD to different anomaly detectors.

Panel of Figure 2 — Figure 2: Examples of visual differences that produce misleading anomaly responses. (a) Juice bottle: Different liquid colors lead to false anomaly responses on the bottle logo. (b) Screw: Variations in screw orientation cause widespread false responses. (c) Cashew: Textured backgrounds are incorrectly identified as anomalous regions. (d) Breakfast box: Missing dried fruits cause a logical anomaly, but the baseline incorrectly highlights the entire tray. VerdictAD uses visual decisions to suppress misleading responses while preserving useful anomaly evidence.

Figure 3: Overview of VerdictAD. A base detector $\mathcal{D}$ produces an anomaly map $\mathbf{M}$ and a score $s$ from the query $\mathbf{X}$ and the references $\mathcal{R}$ . The frozen decision model $\mathcal{J}$ answers a reference-free and a reference-conditioned question on the whole image, giving the image logit $z$ that gates $s$ into $S$ . It also answers one question per window $b_{i}$ , either on a crop pair or on the full image with a red box, giving window logits $z_{i}$ that form the verification map $\mathbf{W}$ and reweight $\mathbf{M}$ into $\mathbf{M}^{\prime}$ .

Panel of Figure 4 — Figure 4: Qualitative comparison of anomaly localization results on VisA, MVTec-AD, and MVTec LOCO AD. Each example shows the input image, ground truth (GT), and anomaly-map overlays from AnomalyDINO, SubspaceAD, and VerdictAD. Compared with the baseline detectors, VerdictAD generally produces more focused responses and suppresses distracting activations on irrelevant regions, while preserving the responses around true defects. The LOCO examples further illustrate that VerdictAD can improve localization quality for both structural and logical anomalies when the original anomaly map already contains useful evidence.

		MVTec AD				VisA
		Image-level		Pixel-level		Image-level		Pixel-level
Setup	Method	AUROC	AUPR	AUROC	AUPRO	AUROC	AUPR	AUROC	AUPRO
1-shot	SPADE [ 15 ]	$81.0$	$90.6$	$91.2$	$83.9$	$79.5$	$82.0$	$95.6$	$84.1$
	PatchCore [ 52 ]	$83.4$	$92.2$	$92.0$	$79.7$	$79.9$	$82.8$	$95.4$	$80.5$
	WinCLIP [ 33 ]	$93.1$	$96.5$	$95.2$	$87.1$	$83.8$	$85.1$	$96.4$	$85.1$
	PromptAD [ 40 ]	$94.6$	$97.1$	$95.9$	$87.9$	$86.9$	$88.4$	$96.7$	$85.1$

Table 1: Few-shot anomaly detection and localization on MVTec AD and VisA (%). Values with subscripts report mean and standard deviation. Bold and underlined means indicate the best and second-best results within each shot and metric; the final four rows in each shot are excluded from ranking.

\Delta

denotes the difference from the corresponding baseline in percentage points.

	Image-level		Pixel-level
Method	AUROC	AUPR	AUROC	AUPRO
PatchCore [ 52 ]	$67.4_{\scriptscriptstyle\pm 3.3}$	$77.6_{\scriptscriptstyle\pm 2.5}$	$74.9_{\scriptscriptstyle\pm 0.0}$	$66.7_{\scriptscriptstyle\pm 0.5}$
PatchCore + Ours	${78.0}_{\scriptscriptstyle\pm 2.4}$	${86.5}_{\scriptscriptstyle\pm 1.5}$	$79.9_{\scriptscriptstyle\pm 0.3}$	$68.7_{\scriptscriptstyle\pm 0.4}$
DINOv2-L [ 48 ]	$68.7_{\scriptscriptstyle\pm 0.8}$	$80.3_{\scriptscriptstyle\pm 0.8}$	$77.1_{\scriptscriptstyle\pm 0.1}$	$62.8_{\scriptscriptstyle\pm 0.3}$
DINOv2-L + Ours	${77.5}_{\scriptscriptstyle\pm 0.1}$	${86.3}_{\scriptscriptstyle\pm 0.1}$	$80.0_{\scriptscriptstyle\pm 0.1}$	$65.4_{\scriptscriptstyle\pm 0.2}$
AnomalyDINO [ 16 ]	$67.4_{\scriptscriptstyle\pm 2.0}$	$77.7_{\scriptscriptstyle\pm 1.2}$	${81.5}_{\scriptscriptstyle\pm 0.2}$	$59.5_{\scriptscriptstyle\pm 0.8}$

Table 2: Four-shot anomaly detection and localization on MVTec LOCO AD (%). Results are reported as mean and standard deviation over three seeds. Best and second-best results are bold and underlined.

	MVTec AD		VisA		LOCO AD
Method	AUROC	AUPR	AUROC	AUPR	AUROC	AUPR
AnomalyCLIP	91.55	96.36	81.97	85.34	61.02	74.75
AnomalyCLIP + Ours	92.97	97.19	90.70	92.82	65.10	79.14
AdaCLIP	90.08	95.74	86.31	88.68	60.38	73.39
AdaCLIP + Ours	92.39	97.09	92.08	93.71	64.80	78.51

Table 3: Zero-shot image-level AUROC and AUPR (%) on three datasets. Ours uses reference-free global decisions with a gating exponent of

0.25

. Best and second-best results are bold and underlined.

Figure 5: Image-level anomaly-score distributions of SubspaceAD before (top) and after (bottom) GVR under the 4-shot setting. Scores are normalized within each class for visualization.

Figure 6: Image-level ROC curves for SubspaceAD before and after GVR on three industrial anomaly detection benchmarks (4-shot, single-seed analysis). The ranking-correction statistics are discussed in the text; no separate ranking plot is included here.

Figure 7: Sensitivity analyses of VerdictAD. (a) Effect of the global gating exponent

\alpha

, with separate vertical scales for MVTec AD/VisA and LOCO AD. (b) Effect of the local reweighting exponent

\gamma

on Pixel AUROC and AUPRO, evaluated on an independent synthetic validation set. (c) Effect of the DINOv2-L feature layer on pixel-level detection performance.

Configuration	MVTec AD	VisA	LOCO AD
Reference-free decision	91.8	89.8	63.3
Reference-conditioned decision	93.9	86.3	69.5
Combined global decisions	96.3	91.7	73.8
DINOv2-L detector	95.5	93.7	67.1
DINOv2-L + GVR	98.3	96.9	76.2

Table 4: Ablation of global decisions. Image AUROC (%).

Windows	Logical	Structural	All
No LVR	75.98 / 60.77	94.22 / 89.20	77.05 / 69.34
Crop only	77.77 / 63.51	95.06 / 90.76	78.31 / 70.60
Box only	79.90 / 64.43	94.98 / 90.23	80.48 / 71.46
Crop + Box	79.31 / 64.70	95.21 / 90.67	79.87 / 71.60

Table 5: Window-view ablation on MVTec LOCO AD, with the 4-shot SubspaceAD detector (single seed). Metrics are pixel AUROC / AUPRO (%).

Anomaly Detection

Anomaly detection finds rare or unexpected patterns in data, from defective products and system faults to fraud and security threats.

See all 690 papers

SubspaceAD: Training-Free Few-Shot Anomaly Detection via Subspace Modeling
Feb 26, 2026Camile Lendering, Erkut Akdag, Egor BondarevFew-Shot Anomaly Detection Anomaly Detection
AIMS Group, Department of Electrical Engineering, Eindhoven University of Technology
Hypergraph-Enhanced Training-Free and Language-Free Few-Shot Anomaly Detection
May 11, 2026Guohuan Xie, Xin He, Dingying Fan +2Few-Shot Anomaly Detection Medical Image Anomaly Detection
Nankai University · Tianjin University of Technology · Tsinghua University
ConceptADapt: Concept-guided Adaptive Feature Reconstruction with Dynamic Attention for Few-Shot Industrial Anomaly Detection
Aug 6, 2026Yufei Li, Yicheng Ruan, Long Tian +2Few-Shot Anomaly Detection Industrial Anomaly Detection
School of Computer Science and Technology, Xidian University, Xi’an, China

Feb 26, 2026J/K move · Enter open · S save

SubspaceAD: Training-Free Few-Shot Anomaly Detection via Subspace Modeling

Camile Lendering, Erkut Akdag, Egor Bondarev

AIMS Group, Department of Electrical Engineering, Eindhoven University of Technology

Few-Shot Anomaly Detection Anomaly Detection Industrial Anomaly Detection Reconstruction-Based Anomaly Detection

Open paper AskPDF arXiv

Detecting visual anomalies in industrial inspection often requires training with only a few normal images per category. Recent few-shot methods achieve strong results employing foundation-model features, but typically rely on memory banks, auxiliary datasets, or multi-modal tuning of vision-language models. We therefore question whether such complexity is necessary given the feature representations of vision foundation models. To answer this question, we introduce SubspaceAD, a training-free method, that operates in two simple stages. First, patch-level features are extracted from a small set of normal images by a frozen DINOv2 backbone. Second, a Principal Component Analysis (PCA) model is fit to these features to estimate the low-dimensional subspace of normal variations. At inference, anomalies are detected via the reconstruction residual with respect to this subspace, producing interpretable and statistically grounded anomaly scores. Despite its simplicity, SubspaceAD achieves state-of-the-art performance across one-shot and few-shot settings without training, prompt tuning, or memory banks. In the one-shot anomaly detection setting, SubspaceAD achieves image-level and pixel-level AUROC of 97.1% and 97.5% on the MVTec-AD dataset, and 93.2% and 98.2% on the VisA dataset, respectively, surpassing prior state-of-the-art results. Code and demo are available at https://github.com/CLendering/SubspaceAD.

Abstract

Anomaly Detection

Explore similar work

SubspaceAD: Training-Free Few-Shot Anomaly Detection via Subspace Modeling

Hypergraph-Enhanced Training-Free and Language-Free Few-Shot Anomaly Detection

ConceptADapt: Concept-guided Adaptive Feature Reconstruction with Dynamic Attention for Few-Shot Industrial Anomaly Detection