StructCore: Structure-Aware Image-Level Scoring for Training-Free Unsupervised Anomaly Detection

Joongwon Chae; Lihui Luo; Yang Liu; Runming Wang; Dongmei Yu; Zeming Liang; Xi Yuan; Dayan Zhang

cs.CVFeb 19, 2026

StructCore: Structure-Aware Image-Level Scoring for Training-Free Unsupervised Anomaly Detection

Authors: Joongwon Chae, Lihui Luo, Yang Liu, Runming Wang, Dongmei Yu, Zeming Liang, Xi Yuan, Dayan Zhang, +3 more

Organizations: Tsinghua University Shenzhen International Graduate School, Shenzhen, China · Ratel Soft · Affiliated Fifth Hospital, Wenzhou Medical University, Wenzhou, Zhejiang, China · Chinese Medicine Guangdong Laboratory

Abstract

Max pooling is the de facto standard for converting anomaly score maps into image-level decisions in memory-bank-based unsupervised anomaly detection (UAD). However, because it relies on a single extreme response, it discards most information about how anomaly evidence is distributed and structured across the image, often causing normal and anomalous scores to overlap. We propose StructCore, a training-free, structure-aware image-level scoring method that goes beyond max pooling. Given an anomaly score map, StructCore computes a low-dimensional structural descriptor phi(S) that captures distributional and spatial characteristics, and refines image-level scoring via a diagonal Mahalanobis calibration estimated from train-good samples, without modifying pixel-level localization. StructCore achieves image-level AUROC scores of 99.6% on MVTec AD and 98.4% on VisA, demonstrating robust image-level anomaly detection by exploiting structural signatures missed by max pooling.

Figures & tables

Figure 1: Overview of the StructCore framework. Normal images are encoded by a frozen DINOv2 ViT-B/14 backbone using multi-layer skip feature extraction. Patch features from multiple layers are concatenated and compressed via a fixed random projection, followed by greedy coreset selection to construct a category-specific memory bank. At inference time, an optional routing bank selects a relevant memory subset for efficient nearest-neighbor matching, producing an anomaly score map. While the base image-level score is obtained by conventional max pooling, StructCore augments it with a structure-aware score computed from a low-dimensional descriptor of the anomaly map, calibrated using statistics from train-good samples. StructCore refines image-level decisions without altering pixel-level localization.

Component	Setting
Backbone	DINOv2 ViT-B/14 (frozen)
Input	resize $448\times 448$ , center-crop $392\times 392$
Layers	skip5: $[-1,-3,-6,-9,-12]$ (concat fusion)
Projection	fixed random projection to $D{=}512$ (RP-512)
Memory bank	coreset ratio = 1% (default), 10% (robustness)
Patch scoring	kNN $k{=}5$ , mean over neighbors (squared $\ell_{2}$ )

Table 1: Paper-default configuration. We use a 1% coreset by default and report 10% as a robustness check.

Panel of Figure 2 — Figure 2: Base vs. hybrid image scores illustrating where StructCore changes the image-level decision. The x-axis is the base score $S_{\mathrm{base}}=\max(S)$ and the y-axis is the final score $S_{\mathrm{hyb}}=S_{\mathrm{base}}+\lambda_{\mathrm{auto}}\,D_{\mathrm{struct}}$ . Points are colored by ground-truth label. Thresholds $\tau_{\mathrm{base}}$ and $\tau_{\mathrm{hyb}}$ are set to the 99.5% quantile of train-good scores. Shaded cross-over regions mark samples whose decisions differ between $S_{\mathrm{base}}$ and $S_{\mathrm{hyb}}$ (counts are shown in each panel).

Dataset	Method	Backbone	I-AUROC	P-AUROC
MVTec-AD [ 1 ]	CFA [ 40 ]	WideResNet-50	99.2	98.2
	PaDiM [ 7 ]	WideResNet-50	95.5	97.3
	PatchCore (max pooling) [ 9 ]	WideResNet-50	98.9	98.1
	PatchCore + StructCore (struct only)	WideResNet-50	99.1	98.1
	PatchCore + StructCore (hybrid)	WideResNet-50	99.2	98.1
	Ours (DINOv2, max pooling)	DINOv2 ViT-B/14 (skip-layer)	98.7	98.2

Table 2: Average performance (Mean) on MVTec-AD. Each entry reports I-AUROC / P-AUROC . Results are only comparable within each backbone group. CFA/PaDiM/PatchCore follow their standard WideResNet-50 settings, while our variants use a frozen DINOv2 ViT-B/14 backbone with skip-layer fusion. Within the DINOv2 group, Base denotes max pooling on the same anomaly maps, and Struct/Hybrid modify only the image-level scoring (pixel-level metrics remain unchanged). For PatchCore-based runs in this table, the coreset ratio is set to 1%. Best results are highlighted in bold within each backbone group.

Dataset	Method	Image-level			Pixel-level			AUPRO
		I-AUROC	I-AP	I-F1-max	P-AUROC	P-AP	P-F1-max
MVTec-AD [ 1 ]	RD4AD [ 3 ]	94.6	96.5	95.2	96.1	48.6	53.8	91.1
	SimpleNet [ 41 ]	95.3	98.4	95.8	96.9	45.9	49.7	86.5
	DeSTSeg [ 42 ]	89.2	95.5	91.6	93.1	54.3	50.9	64.8
	UniAD [ 43 ]	96.5	98.8	96.2	96.8	43.4	49.5	90.7
	ReContrast [ 44 ]	98.3	99.4	97.6	97.1	60.2	61.5	93.2

Table 3: Performance (%) on MVTec-AD and VisA. Our StructCore results are reported with a 1% coreset. Best results are in bold.

Coreset	Extract (ms)	kNN (ms)	Post (ms)	Total (ms)	FPS	GPU peak (MB)
1%	9.60	0.20	0.32	10.12	98.85	2333
5%	9.68	0.68	0.31	10.67	93.71	2459
10%	9.73	1.29	0.27	11.30	88.53	2619

Table 4: Scalability with memory-bank size on MVTec AD (RTX 3090, batch=16). “Extract” includes DINOv2 feature extraction and random projection; “kNN” is FAISS search; “Post” includes Gaussian blur and StructCore scoring. Coreset ratio controls the memory-bank size.

$\phi(S)$	Base	Hybrid	$\Delta$
– (Base only)	98.68	–	–
$\{\sigma_{S}\}$	98.68	99.57	+0.89
$\{\mathrm{topk\_mean}_{r}\}$	98.68	99.57	+0.88
$\{\mathrm{TV}\}$	98.68	99.07	+0.38
$\{\sigma_{S},\mathrm{topk\_mean}_{r}\}$	98.68	99.66	+0.97
$\{\sigma_{S},\mathrm{TV}\}$	98.68	99.55	+0.87

Table 5: Ablation of

\phi(S)

components on MVTec AD (mean I-AUROC, %). Base uses max pooling; Hybrid uses

S_{\mathrm{hyb}}=S_{\mathrm{base}}+\lambda_{\mathrm{auto}}D_{\mathrm{struct}}

\Delta

denotes the improvement over Base.

Distance	MVTec AD (15)	VisA (12)
$\ell_{1}$ (std)	0.9965 $\pm$ 0.0057	0.9743 $\pm$ 0.0347
Diagonal Mahalanobis	0.9962 $\pm$ 0.0059	0.9837 $\pm$ 0.0319
Chebyshev (std)	0.9960 $\pm$ 0.0060	0.9842 $\pm$ 0.0293
Manhattan	0.9934 $\pm$ 0.0108	0.9789 $\pm$ 0.0339
Euclidean	0.9927 $\pm$ 0.0117	0.9767 $\pm$ 0.0349
Chebyshev	0.9922 $\pm$ 0.0123	0.9761 $\pm$ 0.0351

Table 6: Ablation of structural distance for computing

D_{\mathrm{struct}}

(mean image-level AUROC, mean

\pm

std over categories). Standardized variants (suffix _std ) normalize each dimension of

\phi(S)

by the train-good standard deviation.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Category	Base	$\{\sigma_{S}\}$	$\{\mathrm{topk\_mean}_{r}\}$	$\{\mathrm{TV}\}$	$\{\sigma_{S},\mathrm{topk\_mean}_{r}\}$	$\{\sigma_{S},\mathrm{TV}\}$	$\{\mathrm{topk\_mean}_{r},\mathrm{TV}\}$	$\{\sigma_{S},\mathrm{topk\_mean}_{r},\mathrm{TV}\}$
Bottle	100.00	100.00	100.00	100.00	100.00	100.00	100.00	100.00
Cable	92.71	99.04	98.41	94.77	99.23	98.97	98.35	99.10
Capsule	95.81	98.52	98.21	98.88	98.28	98.60	98.32	98.40
Hazelnut	100.00	100.00	100.00	100.00	100.00	100.00	100.00	100.00
Metal Nut	100.00	100.00	100.00	99.76	100.00	99.90	100.00	100.00
Pill	98.06	99.15	99.59	98.28	99.32	99.35	99.59	99.56

Appendix

Table 7: Per-category structural feature ablation on MVTec AD (image-level AUROC, %). Columns show image-level scoring variants: Base is max pooling; others are Hybrid scores with the indicated

\phi(\cdot)

Category	Base (Max)			Struct (Ours)			Hybrid (Ours)			Pixel-level Metrics
Category	AUROC	AP	F1-max	AUROC	AP	F1-max	AUROC	AP	F1-max	AUROC	AP	F1-max	AUPRO
Bottle	100.0	100.0	100.0	100.0	100.0	100.0	100.0	100.0	100.0	98.8	86.1	81.5	95.9
Cable	97.7	98.7	93.9	99.4	99.7	97.3	99.5	99.7	97.2	94.5	52.6	62.5	88.0
Capsule	95.6	99.0	96.8	97.8	99.4	98.6	97.8	99.5	98.2	98.3	59.7	56.8	94.5
Carpet	100.0	100.0	100.0	100.0	100.0	100.0	100.0	100.0	100.0	99.3	73.4	70.4	97.8
Grid	99.9	100.0	99.1	99.9	100.0	99.1	100.0	100.0	100.0	99.6	61.9	61.5	98.6

Appendix

Table 8: Detailed per-category performance on MVTec AD (1% coreset) . Best Image-level AUROC is highlighted in bold .

Category	Base (Max)			Struct (Ours)			Hybrid (Ours)			Pixel-level Metrics
Category	AUROC	AP	F1-max	AUROC	AP	F1-max	AUROC	AP	F1-max	AUROC	AP	F1-max	AUPRO
Bottle	100.0	100.0	100.0	100.0	100.0	100.0	100.0	100.0	100.0	99.0	87.1	81.0	96.5
Cable	97.4	98.5	93.1	99.7	99.8	98.4	99.8	99.9	98.4	95.4	53.5	63.3	89.4
Capsule	97.5	99.4	98.6	98.3	99.6	98.2	98.4	99.6	98.6	98.6	61.0	58.1	95.5
Carpet	100.0	100.0	100.0	100.0	100.0	100.0	100.0	100.0	100.0	99.4	74.4	70.9	97.9
Grid	99.2	99.7	99.1	99.9	100.0	99.1	99.8	99.9	99.1	99.6	63.7	63.5	98.7

Appendix

Table 9: Detailed per-category performance on MVTec AD (10% coreset) .

Category	Base (Max)			Struct (Ours)			Hybrid (Ours)			Pixel-level Metrics
Category	AUROC	AP	F1-max	AUROC	AP	F1-max	AUROC	AP	F1-max	AUROC	AP	F1-max	AUPRO
Candle	96.7	97.0	89.7	98.4	98.4	93.8	98.2	98.3	92.9	99.4	39.4	45.6	98.2
Capsules	98.9	99.4	96.5	96.8	97.8	93.8	98.6	99.1	97.5	99.3	62.9	64.0	97.7
Cashew	94.2	97.3	90.8	98.1	99.0	96.6	98.2	99.1	95.2	96.9	58.4	59.3	91.1
Chewinggum	99.4	99.7	98.0	99.3	99.7	99.0	99.6	99.8	99.0	99.4	75.6	69.8	98.1
Fryum	98.6	99.3	95.3	98.1	99.1	94.9	98.6	99.4	95.8	95.2	43.2	49.5	85.2

Appendix

Table 10: Detailed per-category performance on VisA (1% coreset) .

Category	Base (Max)			Struct (Ours)			Hybrid (Ours)			Pixel-level Metrics
Category	AUROC	AP	F1-max	AUROC	AP	F1-max	AUROC	AP	F1-max	AUROC	AP	F1-max	AUPRO
Candle	96.3	96.7	90.9	97.9	98.0	93.5	97.9	98.0	92.8	99.6	40.6	47.4	98.6
Capsules	99.3	99.6	97.5	97.0	98.0	95.0	98.5	98.8	98.0	99.5	63.8	63.9	98.3
Cashew	96.6	98.3	94.0	98.2	99.0	97.6	98.6	99.2	97.5	93.7	52.2	55.6	83.7
Chewinggum	99.4	99.7	98.5	99.1	99.6	98.5	99.4	99.8	99.0	99.3	77.7	71.7	98.0
Fryum	99.1	99.6	96.6	97.8	98.9	96.0	99.1	99.6	96.5	95.7	43.3	48.8	86.1

Appendix

Table 11: Detailed per-category performance on VisA (10% coreset) .

Layer idx	MVTec (15)	VisA (12)
-1	98.58	80.75
-2	98.61	92.97
-3	98.78	95.91
-4	98.94	97.11
-5	98.94	96.65
-6	98.86	97.00

Appendix

Table 12: Single-layer depth sweep on MVTec AD and VisA. All settings are fixed: RP dim=256,

k

=5, with the indicated

\phi(\cdot)

Images are resized to 448 and center-cropped to 392. We report image-level AUROC (%).

Figure 3: Qualitative results on MVTec AD (all categories).

Figure 4: Qualitative results on VisA (all categories).

Explore similar work

Sep 4, 2026

Training-Free Logical and Structural Anomaly Detection via Calibrated Fusion

Industrial anomaly detection must handle two distinct defect families: structural anomalies, which manifest as local texture corruptions, and logical anomalies, which violate global rules on object count, composition, or arrangement. Existing detectors typically favor one family at the expense of the other. In particular, training-free methods effectively exploit frozen representations but lack an explicit notion of object count, while methods that reason about counts usually rely on category-specific component modeling. We show that counting ability can be introduced into training-free anomaly detection without additional training or part-level supervision. Our key idea is a normal-set calibration that aligns heterogeneous anomaly cues using statistics from normal images, enabling their direct fusion within a unified training-free framework. Built upon this calibration, our detector combines complementary frozen cues to address both logical and structural anomalies. On MVTec-LOCO, our method achieves image-level AUROCs of 89.0 and 95.9 on logical and structural anomalies, respectively, yielding a 92.5 average---the best among training-free detectors in our comparison. It remains competitive with methods requiring network training or part annotations, while its structural variant matches PatchCore on MVTec-AD (99.1 image-AUROC), suggesting that the proposed calibration generalizes beyond logical anomaly detection.

Changyi Li, Miao Yu, Kai Dong +1

Aalto University, 02150 Espoo, Finland · Harbin Engineering University, 150001 Harbin, China

Anomaly Detection Industrial Anomaly Detection

Ask PDF

Sep 18, 2026

A Principled Approach to Unsupervised Anomaly Detection

Traditional unsupervised anomaly detection (UAD) methods are designed to flag or localise deviations from a normative distribution, ignoring the underlying generative mechanisms of the anomalies. Yet the nature of an anomaly is often as important as its presence. We reformulate UAD as a Bayesian inverse problem, in which the objective is to infer the most probable corruption responsible for each observation. Our framework yields a probabilistic anomaly score as the energy of the inferred corruption parameters, and serves as a principled recipe for developing new UAD algorithms. We derive several existing methods as instances of the general framework, each corresponding to the same energy score under different modelling choices. Experimentally, we study the framework's components in a controlled setting, and improve object-class AUROC on the MVTec AD dataset by 2.3% by adapting the underlying corruption model. Finally, we validate the framework on a brain MRI benchmark, achieving strong detection performance while producing estimates of pathology intensity, bias, and geometry. Code is available at https://github.com/jgmyles/inverse-uad.

James Myles, Matthew Baugh, Johanna P. Müller +2

Imperial College London, UK · Friedrich-Alexander University Erlangen-Nürnberg, Germany · Nanyang Technological University, Singapore

Unsupervised Anomaly Detection Anomaly Detection Industrial Anomaly Detection

Ask PDF

Jul 6, 2026

ProCon: Projection-Consistency Memory for Training-Free Anomaly Detection

Memory-based anomaly detection is attractive because it localizes defects from normal images without training a decoder or synthesizing pseudo anomalies. However, most memory methods still use the memory bank as a nearest-neighbor lookup table: a test patch is treated as normal if it has one nearby normal anchor. This hard retrieval view is vulnerable to false-normal matches and does not test whether the patch is consistently supported by a local normal neighborhood. We propose ProCon, a training-free framework that turns memory retrieval into decoder-free reconstruction. ProCon softly projects each test patch onto nearby normal memory vectors and uses the projection residual as anomaly evidence. To stabilize this residual, it constructs seed-perturbed layer-wise memories, aggregates bank residuals by a median, and fuses depth-specific residual maps by layer consensus. ProCon requires no decoder training, backbone fine-tuning, learned fusion weights, or pseudo-anomaly supervision. Across MVTec-AD, VisA, and Real-IAD under the single-category evaluation protocol, ProCon achieves strong image- and pixel-level performance under seven standard metrics, including image AUROC scores of 99.8%, 99.2%, and 93.2%, respectively. Ablations show that the gains come from replacing hard retrieval with soft normal projection and stabilizing the residuals through memory and depth consensus. The code is available at https://github.com/jw-chae/Procon

Joongwon Chae, Lihui Luo, Yang Liu +4

Tsinghua University Shenzhen International Graduate School, Shenzhen, China · Affiliated Fifth Hospital, Wenzhou Medical University, Wenzhou, Zhejiang, China · Chinese Medicine Guangdong Laboratory +1

Industrial Anomaly Detection Anomaly Localization Reconstruction-Based Anomaly Detection Anomaly Detection

Ask PDF