Novel Anomaly Detection Scenarios and Evaluation Metrics to Address the Ambiguity in the Definition of Normal Samples

Reiji Saito; Satoshi Kamiya; Kazuhiro Hotta

cs.CVApr 8, 2026

Novel Anomaly Detection Scenarios and Evaluation Metrics to Address the Ambiguity in the Definition of Normal Samples

Authors: Reiji Saito, Satoshi Kamiya, Kazuhiro Hotta

Organizations: Meijo University, 1-501 Shiogamaguchi, Tempaku-ku, Nagoya 468-8502, Japan

Abstract

In conventional anomaly detection, training data consist of only normal samples. However, in real-world scenarios, the definition of a normal sample is often ambiguous. For example, there are cases where a sample has small scratches or stains but is still acceptable for practical usage. On the other hand, higher precision is required when manufacturing equipment is upgraded. In such cases, normal samples may include small scratches, tiny dust particles, or a foreign object that we would prefer to classify as an anomaly. Such cases frequently occur in industrial settings, yet they have not been discussed until now. Thus, we propose novel scenarios and an evaluation metric to accommodate specification changes in real-world applications. Furthermore, to address the ambiguity of normal samples, we propose the RePaste, which enhances learning by re-pasting regions with high anomaly scores from the previous step into the input for the next step. On our scenarios using the MVTec AD benchmark, RePaste achieved the state-of-the-art performance with respect to the proposed evaluation metric, while maintaining high AUROC and PRO scores. Code: https://github.com/ReijiSoftmaxSaito/Scenario

Figures & tables

Figure 1: Examples that the definition of a normal sample is ambiguous. In the MVTec AD dataset, blue boxes indicate anomalous samples, while orange boxes indicate normal samples.

Figure 2: A2N includes “Anomaly-to-Normal sub-scenario” and “Standard sub-scenario”. In the “Anomaly-to-Normal sub-scenario”, if you want to treat the “Broken” as normal in the “Grid” category, we use half of all “Broken” images as training data and evaluate the remaining half as normal samples. “Standard sub-scenario” in A2N is the same as a standard anomaly detection task trained from normal samples.

Figure 3: N2A includes “Normal-to-Anomaly sub-scenario” and “Standard sub-scenario”. In the “Normal-to-Anomaly sub-scenario”, pseudo-anomalies are added to the test data as anomalous samples. In “Standard sub-scenario”, pseudo-anomalies are added to both training and test data as normal samples.

Figure 4: Examples where the anomaly score map shows a high area in the orange regions, even though they are used as normal samples during training. Note that we use GLASS.

Figure 5: Overview of Repaste. To detect small scratches and dust that are often subject to specification changes, RePaste feeds the input image $x_{\alpha}$ into the model to obtain an anomaly map $A_{\alpha}$ . Then, to extract only the regions with high anomaly scores, $A_{\alpha}$ is binarized using a threshold $\tau$ . Finally, the extracted regions containing small scratches or dust are re-pasted onto the next image $x_{\alpha+1}$ , increasing their occurrence frequency and making them easier to be treated as normal samples.

Model	FastFlow [ 39 ]	PatchCore [ 29 ]	RD4AD [ 9 ]	RD++ [ 32 ]	SimpleNet [ 23 ]	DiAD [ 15 ]	mambaAD [ 11 ]	INP-Former [ 25 ]	UniNet [ 34 ]	Dinomaly [ 13 ]	GLASS [ 6 ]	RePaste
Conf	Arxiv2021	CVPR2022	CVPR2022	CVPR2023	CVPR2023	AAAI2024	NeurIPS2024	CVPR2025	CVPR2025	CVPR2025	ECCV2024	-
A2N	82.11	50.75	65.26	67.68	84.25	54.33	62.39	68.69	61.34	84.70	86.29	86.88
N2A	79.68	50.23	72.70	77.20	75.70	52.48	58.82	60.47	72.87	81.88	83.25	83.75

Table 1: Comparison results for the proposed scenarios on MVTec AD. The evaluation metric is the S-AUROC scores.

Eval	Scenario	FastFlow	PatchCore	RD4AD	RD++	SimpleNet	DiAD	mambaAD	INP-Former	UniNet	Dinomaly	GLASS	RePaste
I-AUROC	$A2N_{S}$	96.02	98.33	96.97	97.09	98.60	84.40	98.63	96.65	96.07	98.49	99.54	99.45
	$A2N_{A2N}$	91.05	91.31	89.07	89.16	93.47	77.06	90.13	92.37	87.68	93.74	94.65	94.77
	$N2A_{S}$	93.46	93.67	90.47	89.88	94.24	85.42	94.92	95.20	88.80	96.28	95.68	95.68
	$N2A_{N2A}$	96.16	95.06	95.11	95.36	97.29	82.83	95.32	97.79	94.92	98.24	97.45	97.61
P-AUROC	$A2N_{S}$	98.60	99.10	97.12	97.01	98.85	93.87	99.13	98.75	98.81	99.08	99.37	99.33
	$A2N_{A2N}$	98.58	99.12	95.70	96.46	98.86	94.45	99.06	98.86	98.75	99.11	99.31	99.31

Table 2: Comparison results of

A2N_{S}

A2N_{A2N}

N2A_{S}

and

N2A_{N2A}

on MVTec AD. The evaluation metric is the mean I-AUROC, P-AUROC, and PRO.

Figure 6: Qualitative visualization for pixel-level anomaly segmentation.

S-AUROC
	GLASS	RePaste	RePaste
Mixup			✓
$A2N$	86.29	87.48	86.88
$N2A$	83.25	78.26	83.75
I-AUROC
$A2N_{S}$	99.54	99.53	99.45

Table 3: Comparison of GLASS, RePaste w/ and w/o Mixup.

Figure 7: The difference between RePaste with and without Mixup. The blue box regions show the RePaste areas. RePaste Without Mixup creates boundary discontinuities.

Anomaly Detection

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

See all 652 papers

What Remains Normal? Clean Images Miss Useful Near-Defect Normal Patches for Anomaly Detection
Aug 24, 2026Joongwon Chae, Runming Wang, Peiwu QinIndustrial Anomaly Detection Reconstruction-Based Anomaly Detection
Institute of Biopharmaceutical and Health Engineering, Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · RatelSoft · Guangdong Provincial Laboratory of Traditional Chinese Medicine, Hengqin, Guangdong, China
Beyond Normal References: Discriminative Few-Shot Anomaly Detection
May 22, 2026Huan Wang, Jun Shen, Jun Yan +1Few-Shot Anomaly Detection Few-Shot Learning
Singapore Management University, Singapore · University of Wollongong, Australia
BoRAD: Bootstrap your Own Representations for Multi-class Anomaly Detection
Jun 12, 2026Duy Hoang Khuong, Tri Nguyen Minh, Ngu Huynh Cong VietRepresentation Learning Industrial Anomaly Detection
Department of Artificial Intelligence FPT University Ho Chi Minh City, Vietnam · Department of IT FPT University Ho Chi Minh City, Vietnam · Department of Computing Fundamental FPT University Ho Chi Minh City, Vietnam

Aug 24, 2026J/K move · Enter open · S save

What Remains Normal? Clean Images Miss Useful Near-Defect Normal Patches for Anomaly Detection

Joongwon Chae, Runming Wang, Peiwu Qin

Institute of Biopharmaceutical and Health Engineering, Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · RatelSoft · Guangdong Provincial Laboratory of Traditional Chinese Medicine, Hengqin, Guangdong, China

Industrial Anomaly Detection Reconstruction-Based Anomaly Detection

Open paper AskPDF arXiv

Normal-only industrial anomaly detectors use patches from clean training images as normal references or reconstruction targets. This assumes that clean patches are sufficient for the normal regions encountered at test time. We test that assumption directly. On MVTec AD, admitting ground-truth-normal patches from real defect images to a DINOv2 memory candidate pool raises pixel average precision (P-AP) from 73.34 to 76.95 while keeping the encoder, test-time score, and number of stored references fixed. Patches within two patch cells of the annotated defect recover 94.70% of this gain. We then ask whether useful patches of this kind can be exposed using clean training images alone. BoundarySupport inserts a procedural synthetic defect to alter surrounding context, excludes every token intersecting the nominal insertion or a detected RGB change, and learns only from pixel-preserved neighboring patches. Across three paired seeds, the same principle improves P-AP in all six memory and reconstruction settings across MVTec, VisA, and Real-IAD. Matched controls identify the altered-context feature itself as the useful normal evidence: with synthetic input or selected positions fixed, altered-context features outperform their clean-view counterparts as both reconstruction targets and memory references. On MVTec memory, the final score change is also spatially selective, with larger reductions on normal patches next to defects than on mid-distance or far-normal patches in all 15 categories. Code is publicly available at https://github.com/jw-chae/boundary_support.

Abstract

Anomaly Detection

Explore similar work

What Remains Normal? Clean Images Miss Useful Near-Defect Normal Patches for Anomaly Detection

Beyond Normal References: Discriminative Few-Shot Anomaly Detection

BoRAD: Bootstrap your Own Representations for Multi-class Anomaly Detection