When More References Hurt: Contamination-Aware DINOv2 Memory Banks for Few-Shot Steel Defect Detection

Hannaneh Kalantary; Javad Khoramdel

cs.CVAug 22, 2026

When More References Hurt: Contamination-Aware DINOv2 Memory Banks for Few-Shot Steel Defect Detection

Authors: Hannaneh Kalantary, Javad Khoramdel

Organizations: University of Padua · University of Vienna

Abstract

Patch-memory anomaly detectors assume that their reference bank is normal, an assumption that is difficult to guarantee when additional industrial images are unverified. We study whether a few trusted normal images can safely recover useful normal patches from such references without defect masks. Starting from the DINOv2 patch-memory formulation used by AnomalyDINO, we score candidate patches by distance to a clean seed bank, discard the most suspicious 20%, merge the retained patches with the seed, and enforce a fixed budget by greedy coreset selection. On Severstal, naive additional references contain 9.46% anomalous patches; the proposed trim rejects 78.1% of them and reduces residual contamination to 2.59%. At an equal 51,200-patch development budget, the proposed bank reaches 0.1084 AUPRC versus 0.0950 for naive expansion, 0.0952 for random removal, and 0.1030 for eight clean images. Injecting only 0.5% anomalous patches into a clean bank reduces AUPRC from 0.1030 to 0.0759. On all five completed held-out pairs, the proposed bank improves over naive expansion, with a mean gain of 0.0142 AUPRC. Reference purity is therefore a first-order design variable, and unverified images are useful only when their contribution is filtered explicitly.

Figures & tables

Figure 1: DT20 construction and inference. Two trusted references form the clean seed memory; eight unverified references provide candidates. Each candidate is scored by nearest distance to the seed, the largest 20% are rejected, and the retained patches are merged with the seed and reduced to a 51,200-patch memory by greedy coreset selection. Query patches are scored by nearest-memory distance. No defect masks are used by DT20.

Bank	Reject recall	Final contam.	Normal retain	Reject prec.
Naive	0.000	0.0946	1.000	0.000
Random size-matched	0.195	0.0953	0.799	0.092
DT20	0.781	0.0259	0.861	0.369
Oracle	1.000	0.0000	1.000	1.000

Table 1: Purification quality on fold 0, seed 42.

Setting	AUPRC	AUROC	F1-max
Random size-matched	0.0951	0.7971	0.1935
Fixed trim 5%	0.0987	0.8020	0.1986
Fixed trim 10%	0.1022	0.8062	0.2039
Auto percentile 95	0.1007	0.8045	0.2015
Fixed trim 20%	0.1084	0.8132	0.2151

Table 2: Fold-0 purification controls used to freeze the trim rate.

Panel of Figure 2 — Figure 2: Qualitative DT20 purification on a representative unverified reference image. (a) Input image; (b) ground-truth defect mask, shown only for retrospective analysis and never seen by DT20; (c) per-patch distance $d_{u}$ to the clean seed bank; (d) the 20% of patches rejected by the distance trim; (e) the retained 80% merged into the memory. Rejected high-distance patches in (d) concentrate around the true defect region in (b), consistent with the 78.1% reject recall reported in Table 1 .

Method	Clean	Add.	Final bank	AUPRC	AUROC	F1-max	Fixed F1
Clean 1	1	0	6,400	0.0700	0.703	0.135	0.062
Clean 2	2	0	12,800	0.1020	0.818	0.191	0.181
Clean 4	4	0	25,600	0.1050	0.813	0.200	0.177
Clean 8	8	0	51,200	0.1030	0.795	0.199	0.147
Naive, random budget	2	8	51,200	0.0950	0.797	0.193	0.080
Random20 + greedy	2	8	51,200	0.0952	0.7974	0.1936	0.0274

Table 3: Fold-0 reference composition and matched-budget controls.

Condition	AUPRC	AUROC	F1-max	Fixed F1
Clean 2	$0.102\pm 0.017$	$0.784\pm 0.030$	$0.189\pm 0.027$	$0.098\pm 0.099$
Naive 2+8	$0.101\pm 0.011$	$0.796\pm 0.020$	$0.198\pm 0.017$	$0.110\pm 0.025$
Percentile-95 purified	$0.103\pm 0.012$	$0.797\pm 0.020$	$0.200\pm 0.019$	$0.027\pm 0.037$
Oracle purified	$\mathbf{0.110}\pm 0.011$	$\mathbf{0.804}\pm 0.020$	$\mathbf{0.211}\pm 0.017$	$0.142\pm 0.021$

Table 4: Five-seed fold-0 reference-composition study.

Figure 3: AUPRC under controlled anomaly replacement at fixed 51,200-patch memory size.

Contamination	Injected patches	AUPRC
0	0	0.1030
0.5%	256	0.0759
1%	512	0.0710
2%	1,024	0.0667
5%	2,560	0.0632
10%	5,120	0.0606

Table 5: Controlled replacement contamination.

Condition	$n$	AUPRC mean [95% CI]	AUROC	F1-max	Fixed F1
Clean 2	6	0.0979 [0.085, 0.111]	0.780	0.187	0.063
Clean 8	6	0.1406 [0.115, 0.173]	0.823	0.234	0.112
Naive 2+8	6	0.0919 [0.085, 0.099]	0.781	0.180	0.111
Random20 2+8	6	0.0921 [0.085, 0.099]	0.782	0.181	0.010
DT20 2+8	5	0.1062 [0.100, 0.113]	0.790	0.203	0.038
Oracle 2+8	5	0.0998 [0.090, 0.109]	0.791	0.193	0.140

Table 6: Available held-out results on folds 1–2. DT20/oracle use five completed cells; others use six.

Explore similar work

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
What Memory Composition Does Not Tell Us About Anomaly Detection
Aug 24, 2026Joongwon Chae, Runming Wang, Peiwu QinAnomaly Detection Data Contamination
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
Training-Free Human-in-the-Loop Anomaly Detection via Memory Bank Correction
Aug 18, 2026Ayusha Abbas, Saram Abbas, Kabita AdhikariAnomaly Detection Industrial Anomaly Detection
School of Engineering, Newcastle University, Newcastle upon Tyne, UK

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

Explore similar work

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

What Memory Composition Does Not Tell Us About Anomaly Detection

Training-Free Human-in-the-Loop Anomaly Detection via Memory Bank Correction