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

Joongwon Chae; Runming Wang; Peiwu Qin

cs.CVAug 24, 2026

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

Authors: Joongwon Chae, Runming Wang, Peiwu Qin

Organizations: 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

Abstract

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.

Figures & tables

Figure 1: Clean images miss useful normal patches near defects. Left: two real-defect examples under the clean-only memory and the fixed-budget GT-normal oracle. Right: MVTec macro P-AP at a fixed reference count. Ground-truth-normal candidates within two patch cells recover 94.70% of the oracle gain. The oracle uses ground-truth masks only as a diagnostic.

Figure 2: BoundarySupport exposes normal features under altered context while learning only from pixel-preserved positions. (a) A procedural intervention produces an altered-context view $\tilde{x}$ and nominal insertion mask $A$ . Tokens intersecting the nominal insertion or detected RGB changes form the excluded set $C$ , while the surrounding pixel-preserved positions form the ring $R$ . (b) At a pixel-preserved position $p\in R$ , the clean view and altered-context view yield the clean-view feature $F(x)_{p}$ and altered-context feature $F(\tilde{x})_{p}$ , respectively. (c) The memory branch adds selected altered-context features to the clean candidate pool and selects exactly $B$ references, preserving the original memory budget and test-time readout. (d) The reconstruction branch keeps the original clean objective and adds altered-context encoder targets only at positions in $R$ . Test-time scoring uses the observed image without synthesis or masks.

Branch	Dataset	$\Delta$ I-AUROC	$\Delta$ P-AUROC	$\Delta$ P-AP	$\Delta$ P-F1	$\Delta$ AUPRO
Memory	MVTec	$-0.03\pm 0.02$	$+0.07\pm 0.01$	$\mathbf{+2.21\pm 0.04}$	$+1.31\pm 0.01$	$+0.22\pm 0.01$
Memory	VisA	$+0.02\pm 0.01$	$-0.02\pm 0.01$	$\mathbf{+1.82\pm 0.03}$	$+1.03\pm 0.02$	$-0.03\pm 0.01$
Memory	Real-IAD	$-0.11\pm 0.02$	$-0.01\pm 0.01$	$\mathbf{+1.76\pm 0.04}$	$+1.31\pm 0.02$	$-0.05\pm 0.01$
Recon.	MVTec	$-0.08\pm 0.03$	$+0.16\pm 0.02$	$\mathbf{+5.01\pm 0.07}$	$+2.97\pm 0.14$	$+0.55\pm 0.04$
Recon.	VisA	$+0.05\pm 0.01$	$-0.16\pm 0.03$	$\mathbf{+2.27\pm 0.38}$	$+1.16\pm 0.28$	$+0.21\pm 0.18$
Recon.	Real-IAD	$-0.52\pm 0.04$	$-0.03\pm 0.01$	$\mathbf{+4.20\pm 0.12}$	$+2.56\pm 0.09$	$+0.16\pm 0.03$

Table 1: Paired changes after adding BoundarySupport (mean

\pm

sample standard deviation over three seed-level category macros).

Branch	Dataset	Clean-view target/reference	Altered-context target/reference	Altered $-$ clean	Wins
Reconstruction	MVTec	61.82	73.89	+12.07	15/15
Reconstruction	VisA	44.34	52.79	+8.45	11/12
Memory	MVTec	73.36	75.60	+2.23	15/15
Memory	VisA	52.28	54.11	+1.84	11/12

Table 2: Matched feature-identity controls at seed 0. Reconstruction fixes the synthetic input and ring while changing only the reconstruction target; memory fixes the selected positions and bank construction while changing only the stored feature.

Spatial cohort	Mean $100\Delta s$
Defect patches	$-0.61$
Near-defect normal ( $0<d\leq 2$ )	$-0.37$
Mid-distance normal ( $2<d\leq 4$ )	$+0.11$
Far normal ( $d>4$ )	$+0.19$
Clean-image normal	$+0.21$

Table 3: Distance-resolved MVTec memory score changes at seed 0.

Figure 3: Complementary spatial and score diagnostics. Left: in leather and tile, the fixed-budget oracle gain decreases from near to mid to far candidate bands. Right: reconstruction at seed 0 reduces high anomaly scores on normal regions and also reduces mean defect scores while P-AP rises. These summaries diagnose where scores move; P-AP is determined by the full ranking.

Appendix figures & tables25 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Categories	Memory ratio	Memory seeds	Reconstruction seeds
MVTec AD	15	5%	0,1,2	0,1,2
VisA	12	5%	0,1,2	0,1,2
Real-IAD	30	1%	0,1,2	0,1,2

Appendix

Table 4: Evaluation settings.

Candidate pool	I-AUROC	P-AUROC	P-AP	AUPRO	Rel. bank size
Clean candidates (matched eval.)	99.76	98.80	73.34	96.18	1.00
Clean + GT-normal, expanded	99.71	98.94	76.96	96.64	1.05
Clean + GT-normal, fixed budget	99.71	98.94	76.95	96.64	1.00

Appendix

Table 5: MVTec fixed-budget diagnostic on matched held-out evaluation sets.

Category	Clean (matched eval.)	Expanded GT-normal	Fixed-budget GT-normal	Near-defect GT-normal
bottle	89.02	90.29	90.30	90.21
cable	76.70	77.36	77.32	77.34
capsule	64.40	63.87	63.88	63.47
carpet	80.07	84.04	84.05	83.99
grid	63.74	66.38	66.45	66.26
hazelnut	81.64	84.70	84.62	84.67

Appendix

Table 6: MVTec category P-AP for the clean-only memory and GT-normal oracle variants.

Category	Clean	Fixed-budget GT-normal	Near	Mid	Far
leather	61.97	77.55	78.26	68.22	62.52
tile	76.69	89.76	90.02	78.83	77.36

Appendix

Table 7: Distance-stratified fixed-budget P-AP in the two largest-gain categories.

Category	Clean	Fixed-budget GT-normal	Inpainted-defect GT-normal
leather	61.97	77.55	73.69
tile	76.69	89.76	82.80

Appendix

Table 8: P-AP after removing the visible annotated defect by inpainting.

Source $\rightarrow$ evaluation defects	Clean	Cross-defect GT-normal	$\Delta$ P-AP
leather: cut $\rightarrow$ color/fold/glue/poke	64.65	72.35	+7.69
tile: crack $\rightarrow$ other types	90.65	91.44	+0.79
tile: glue-strip $\rightarrow$ other types	73.98	75.44	+1.46

Appendix

Table 9: Cross-defect transfer P-AP.

Category	Candidate near	Selected near	Enrichment	Selected anchors
leather	0.23%	2.68%	11.73 $\times$	258 / 9,604
tile	0.70%	3.44%	4.95 $\times$	310 / 9,016

Appendix

Table 10: Near-defect candidate and selected-reference shares.

Category	Memory	Good normal q95	Defect-image normal q95	q99	Defect mean
leather	Clean	0.40	0.48	0.61	0.65
leather	Fixed-budget GT-normal	0.40	0.44	0.53	0.60
tile	Clean	0.45	0.55	0.66	0.62
tile	Fixed-budget GT-normal	0.45	0.51	0.60	0.61

Appendix

Table 11: Raw oracle score summaries in leather and tile.

Branch	Dataset	Variant	I-AUROC	I-AP	I-F1	P-AUROC	P-AP	P-F1	AUPRO
Memory	MVTec	Baseline	99.76	99.91	99.38	98.80	73.34	71.04	96.18
Memory	MVTec	BoundarySupport	99.73	99.90	99.31	98.86	75.60	72.33	96.39
Memory	VisA	Baseline	99.17	99.28	97.50	99.08	52.23	54.71	97.06
Memory	VisA	BoundarySupport	99.19	99.30	97.54	99.07	54.05	55.75	97.03
Memory	Real-IAD	Baseline	94.89	94.33	89.00	99.17	52.02	53.11	97.60
Memory	Real-IAD	BoundarySupport	94.79	94.11	88.76	99.16	53.78	54.42	97.54

Appendix

Table 12: Seed-0 category-macro metrics. Three-seed paired changes are reported in Table 1 and Appendix H .

Branch	Dataset	Baseline	BoundarySupport	Paired $\Delta$	Positive pairs
Memory	MVTec	$73.35\pm 0.01$	$75.56\pm 0.03$	$+2.21\pm 0.04$	45/45
Memory	VisA	$52.23\pm 0.02$	$54.05\pm 0.03$	$+1.82\pm 0.03$	36/36
Memory	Real-IAD	$52.02\pm 0.02$	$53.78\pm 0.03$	$+1.76\pm 0.04$	90/90
Recon.	MVTec	$68.96\pm 0.11$	$73.98\pm 0.11$	$+5.01\pm 0.07$	45/45
Recon.	VisA	$50.93\pm 0.26$	$53.20\pm 0.39$	$+2.27\pm 0.38$	29/36
Recon.	Real-IAD	$49.46\pm 0.15$	$53.66\pm 0.18$	$+4.20\pm 0.12$	84/90

Appendix

Table 13: P-AP over three paired seeds (0, 1, 2).

Dataset	Clean-view target	Altered-context target	Altered $-$ clean	Category wins
MVTec	61.82	73.89	+12.07	15/15
VisA	44.34	52.79	+8.45	11/12

Appendix

Table 14: Control 3 P-AP: target identity with synthetic input and preserved positions fixed.

Dataset	Clean-view feature	Altered-context feature	$\Delta$ P-AP	Category wins
MVTec	73.36	75.60	+2.23	15/15
VisA	52.28	54.11	+1.84	11/12

Appendix

Table 15: Control 4 P-AP: stored-feature identity with selected positions and bank construction fixed. Deltas are computed from unrounded values.

Spatial cohort	Mean $100\Delta s$
Defect patches	-0.61
Near-defect normal	-0.37
Mid-distance normal	+0.11
Far normal	+0.19
Clean-image normal	+0.21

Appendix

Table 16: Control 5 mean MVTec memory score change. Values are

100\Delta s

Dataset	Supervised object	$\Delta$ I-AUROC	$\Delta$ I-AP	$\Delta$ I-F1	$\Delta$ P-AUROC	$\Delta$ P-AP	$\Delta$ P-F1	$\Delta$ AUPRO	P-AP wins
MVTec	Pixel-preserved surrounding patch	-0.07	-0.05	-0.18	+0.18	+4.94	+3.10	+0.51	15/15
MVTec	Corresponding clean-view patch	-0.02	0.00	+0.04	-0.01	-0.26	-0.15	-0.01	7/15
MVTec	Patch outside nominal mask	-0.23	-0.20	-0.71	-0.03	+0.70	+0.99	-0.09	8/15
MVTec	Synthetic core	-2.11	-1.44	-0.75	-20.86	-52.28	-45.18	-49.13	0/15
VisA	Pixel-preserved surrounding patch	+0.06	+0.08	-0.11	-0.17	+2.09	+0.92	+0.05	9/12
VisA	Corresponding clean-view patch	-0.27	-0.32	-0.65	-0.07	-0.30	-0.16	-0.28	4/12

Appendix

Table 17: Seed-0 full-dataset reconstruction-object controls.

Setting	$\Delta$ I-AUROC	$\Delta$ I-AP	$\Delta$ I-F1	$\Delta$ P-AUROC	$\Delta$ P-AP	$\Delta$ P-F1	$\Delta$ AUPRO
Three interventions, $r=2$ , $\lambda=0.05$	+0.01	+0.01	-0.21	+0.23	+6.99	+4.18	+0.80
Scar only	+0.01	+0.01	-0.04	+0.26	+8.29	+4.58	+0.80
Poisson paste only	+0.12	+0.07	+0.40	+0.18	+4.45	+2.93	+1.09
Local warp + paste only	+0.04	+0.02	+0.05	+0.15	+4.60	+2.85	+0.70
$r=1$	+0.05	+0.03	-0.04	+0.25	+7.55	+4.31	+0.99
$r=3$	0.00	-0.01	-0.04	+0.23	+6.91	+4.11	+1.21

Appendix

Table 18: Paired changes on the ten-category development subset.

Category	I-AUROC	I-AP	I-F1	P-AUROC	P-AP	P-F1	AUPRO
bottle	100.00 $\rightarrow$ 100.00	100.00 $\rightarrow$ 100.00	100.00 $\rightarrow$ 100.00	99.20 $\rightarrow$ 99.26	89.02 $\rightarrow$ 90.79	82.61 $\rightarrow$ 83.12	97.33 $\rightarrow$ 97.54
cable	99.81 $\rightarrow$ 99.78	99.86 $\rightarrow$ 99.85	98.14 $\rightarrow$ 98.14	98.66 $\rightarrow$ 98.66	76.70 $\rightarrow$ 77.68	72.22 $\rightarrow$ 72.54	95.66 $\rightarrow$ 95.71
capsule	98.75 $\rightarrow$ 98.75	99.67 $\rightarrow$ 99.67	98.48 $\rightarrow$ 98.48	98.98 $\rightarrow$ 98.99	64.40 $\rightarrow$ 65.03	61.17 $\rightarrow$ 61.53	96.66 $\rightarrow$ 96.70
carpet	100.00 $\rightarrow$ 100.00	100.00 $\rightarrow$ 100.00	100.00 $\rightarrow$ 100.00	99.58 $\rightarrow$ 99.61	80.07 $\rightarrow$ 81.90	76.78 $\rightarrow$ 77.68	98.61 $\rightarrow$ 98.70
grid	100.00 $\rightarrow$ 100.00	100.00 $\rightarrow$ 100.00	100.00 $\rightarrow$ 100.00	99.63 $\rightarrow$ 99.67	63.74 $\rightarrow$ 67.48	61.84 $\rightarrow$ 64.04	98.78 $\rightarrow$ 98.91
hazelnut	100.00 $\rightarrow$ 100.00	100.00 $\rightarrow$ 100.00	100.00 $\rightarrow$ 100.00	99.57 $\rightarrow$ 99.59	81.64 $\rightarrow$ 82.75	77.43 $\rightarrow$ 78.43	98.56 $\rightarrow$ 98.65

Appendix

Table 19: Category-level seed-0 results for MVTec Memory. Cells with arrows report baseline

\rightarrow

BoundarySupport.

Category	Baseline P-AP	Uniform $\Delta$ P-AP
candle	46.04	+0.33
capsules	68.93	+1.14
cashew	69.10	+2.75
chewinggum	73.90	+6.19
fryum	46.32	+0.42
macaroni1	27.64	+2.47

Appendix

Table 20: Category-level seed-0 results for VisA Memory, uniform q50. Cells with arrows report baseline

\rightarrow

BoundarySupport.

Category	I-AUROC	I-AP	I-F1	P-AUROC	P-AP	P-F1	AUPRO
audiojack	96.51 $\rightarrow$ 96.62	94.91 $\rightarrow$ 95.00	87.75 $\rightarrow$ 87.44	99.83 $\rightarrow$ 99.84	60.97 $\rightarrow$ 68.74	61.49 $\rightarrow$ 64.60	99.44 $\rightarrow$ 99.46
bottle_cap	96.27 $\rightarrow$ 96.29	96.54 $\rightarrow$ 96.61	90.62 $\rightarrow$ 90.91	99.64 $\rightarrow$ 99.60	35.04 $\rightarrow$ 36.86	38.59 $\rightarrow$ 40.24	98.80 $\rightarrow$ 98.68
button_battery	91.70 $\rightarrow$ 91.29	93.99 $\rightarrow$ 93.77	87.81 $\rightarrow$ 87.84	99.62 $\rightarrow$ 99.62	70.06 $\rightarrow$ 70.51	66.73 $\rightarrow$ 67.07	98.73 $\rightarrow$ 98.75
end_cap	91.12 $\rightarrow$ 90.72	90.85 $\rightarrow$ 90.58	88.39 $\rightarrow$ 88.01	99.38 $\rightarrow$ 99.34	14.10 $\rightarrow$ 14.77	26.00 $\rightarrow$ 26.64	98.01 $\rightarrow$ 97.86
eraser	95.25 $\rightarrow$ 95.14	94.39 $\rightarrow$ 94.32	86.46 $\rightarrow$ 86.05	99.79 $\rightarrow$ 99.79	49.00 $\rightarrow$ 50.48	51.47 $\rightarrow$ 52.09	99.30 $\rightarrow$ 99.31
fire_hood	96.09 $\rightarrow$ 95.72	92.27 $\rightarrow$ 91.51	87.28 $\rightarrow$ 86.55	99.66 $\rightarrow$ 99.63	51.23 $\rightarrow$ 52.77	51.29 $\rightarrow$ 51.81	98.88 $\rightarrow$ 98.81

Appendix

Table 21: Category-level seed-0 results for Real-IAD Memory. Cells with arrows report baseline

\rightarrow

BoundarySupport.

Category	I-AUROC	I-AP	I-F1	P-AUROC	P-AP	P-F1	AUPRO
bottle	100.00 $\rightarrow$ 100.00	100.00 $\rightarrow$ 100.00	100.00 $\rightarrow$ 100.00	99.14 $\rightarrow$ 99.19	88.88 $\rightarrow$ 91.02	83.43 $\rightarrow$ 84.20	97.16 $\rightarrow$ 97.59
cable	99.94 $\rightarrow$ 100.00	99.97 $\rightarrow$ 100.00	99.46 $\rightarrow$ 100.00	98.25 $\rightarrow$ 98.28	76.09 $\rightarrow$ 77.57	73.70 $\rightarrow$ 74.75	93.67 $\rightarrow$ 93.30
capsule	98.96 $\rightarrow$ 98.44	99.77 $\rightarrow$ 99.65	99.08 $\rightarrow$ 99.08	98.68 $\rightarrow$ 98.59	60.32 $\rightarrow$ 61.60	58.54 $\rightarrow$ 59.16	97.74 $\rightarrow$ 97.30
carpet	99.96 $\rightarrow$ 99.84	99.99 $\rightarrow$ 99.95	99.44 $\rightarrow$ 99.44	99.34 $\rightarrow$ 99.42	68.56 $\rightarrow$ 74.44	71.60 $\rightarrow$ 75.10	97.36 $\rightarrow$ 97.96
grid	99.92 $\rightarrow$ 99.92	99.97 $\rightarrow$ 99.97	99.13 $\rightarrow$ 99.13	99.44 $\rightarrow$ 99.55	57.34 $\rightarrow$ 65.98	59.84 $\rightarrow$ 64.83	97.52 $\rightarrow$ 97.89
hazelnut	100.00 $\rightarrow$ 100.00	100.00 $\rightarrow$ 100.00	100.00 $\rightarrow$ 100.00	99.46 $\rightarrow$ 99.57	82.44 $\rightarrow$ 85.67	77.52 $\rightarrow$ 80.50	96.63 $\rightarrow$ 97.50

Appendix

Table 22: Category-level seed-0 results for MVTec Reconstruction. Cells with arrows report baseline

\rightarrow

BoundarySupport.

Category	I-AUROC	I-AP	I-F1	P-AUROC	P-AP	P-F1	AUPRO
candle	98.14 $\rightarrow$ 98.11	98.20 $\rightarrow$ 98.13	93.53 $\rightarrow$ 93.60	99.43 $\rightarrow$ 99.41	43.30 $\rightarrow$ 43.60	48.85 $\rightarrow$ 50.79	95.50 $\rightarrow$ 95.00
capsules	98.28 $\rightarrow$ 98.83	98.83 $\rightarrow$ 99.19	97.03 $\rightarrow$ 97.51	99.60 $\rightarrow$ 99.61	64.05 $\rightarrow$ 67.34	63.54 $\rightarrow$ 64.32	97.60 $\rightarrow$ 98.29
cashew	98.48 $\rightarrow$ 98.48	99.36 $\rightarrow$ 99.35	97.49 $\rightarrow$ 96.48	97.91 $\rightarrow$ 96.36	64.18 $\rightarrow$ 61.39	62.17 $\rightarrow$ 61.15	94.81 $\rightarrow$ 93.57
chewinggum	99.86 $\rightarrow$ 99.86	99.93 $\rightarrow$ 99.93	99.00 $\rightarrow$ 98.52	99.28 $\rightarrow$ 99.20	64.59 $\rightarrow$ 75.91	66.86 $\rightarrow$ 70.49	90.79 $\rightarrow$ 89.38
fryum	99.42 $\rightarrow$ 99.36	99.71 $\rightarrow$ 99.69	97.54 $\rightarrow$ 97.51	96.57 $\rightarrow$ 96.15	48.57 $\rightarrow$ 47.31	52.05 $\rightarrow$ 50.61	93.26 $\rightarrow$ 93.58
macaroni1	98.45 $\rightarrow$ 98.62	98.38 $\rightarrow$ 98.54	94.74 $\rightarrow$ 95.57	99.69 $\rightarrow$ 99.71	33.49 $\rightarrow$ 37.50	39.81 $\rightarrow$ 41.72	97.09 $\rightarrow$ 97.37

Appendix

Table 23: Category-level seed-0 results for VisA Reconstruction. Cells with arrows report baseline

\rightarrow

BoundarySupport.

Category	I-AUROC	I-AP	I-F1	P-AUROC	P-AP	P-F1	AUPRO
audiojack	94.22 $\rightarrow$ 93.78	91.73 $\rightarrow$ 91.35	82.38 $\rightarrow$ 80.92	99.81 $\rightarrow$ 99.82	62.21 $\rightarrow$ 66.55	61.46 $\rightarrow$ 64.26	98.16 $\rightarrow$ 98.31
bottle_cap	95.78 $\rightarrow$ 95.49	95.75 $\rightarrow$ 95.34	88.93 $\rightarrow$ 87.78	99.56 $\rightarrow$ 99.57	34.92 $\rightarrow$ 38.60	38.69 $\rightarrow$ 41.58	97.14 $\rightarrow$ 96.95
button_battery	90.96 $\rightarrow$ 89.80	92.74 $\rightarrow$ 92.19	87.23 $\rightarrow$ 86.29	99.50 $\rightarrow$ 99.54	54.89 $\rightarrow$ 62.53	65.91 $\rightarrow$ 67.10	96.58 $\rightarrow$ 96.36
end_cap	87.06 $\rightarrow$ 84.72	87.12 $\rightarrow$ 84.88	85.65 $\rightarrow$ 84.36	99.29 $\rightarrow$ 99.28	17.66 $\rightarrow$ 17.46	26.22 $\rightarrow$ 26.72	96.86 $\rightarrow$ 97.17
eraser	93.78 $\rightarrow$ 93.72	93.00 $\rightarrow$ 92.74	85.28 $\rightarrow$ 84.72	99.78 $\rightarrow$ 99.82	51.88 $\rightarrow$ 58.67	53.61 $\rightarrow$ 58.13	98.05 $\rightarrow$ 98.23
fire_hood	95.96 $\rightarrow$ 94.89	92.61 $\rightarrow$ 89.27	87.32 $\rightarrow$ 84.59	99.77 $\rightarrow$ 99.78	57.42 $\rightarrow$ 63.05	58.13 $\rightarrow$ 59.93	98.09 $\rightarrow$ 97.88

Appendix

Table 24: Category-level seed-0 results for Real-IAD Reconstruction. Cells with arrows report baseline

\rightarrow

BoundarySupport.

Dataset	$\Delta$ normal q95	$\Delta$ normal q99	$\Delta$ defect mean	$\Delta$ P-AP
MVTec	-1.91	-4.91	-4.01	+4.94
VisA	-0.28	-0.88	-1.29	+2.09
Real-IAD	-0.32	-1.13	-2.49	+4.20

Appendix

Table 25: Seed-0 reconstruction score changes.

Dataset	Abs. q95	Abs. q99	Excess q95	Excess q99	Defect mean	$\Delta$ P-AP
MVTec	-0.35	-1.02	-0.51	-1.14	-0.51	+2.26
VisA	+0.30	-0.43	-0.33	-0.92	-0.52	+1.82
Real-IAD	+0.55	+0.09	-0.11	-0.48	-0.21	+1.76

Appendix

Table 26: Seed-0 memory score changes under the uniform headline policies.

Case	Observed change
Reconstruction / VisA / cashew	P-AP -2.79; P-AUROC -1.56; q99 -0.74; defect mean -1.79.
Reconstruction / VisA / pcb3	P-AP -2.42; AUPRO -0.78.
Reconstruction / VisA / fryum	P-AP -1.26; AUPRO +0.32.
Reconstruction / Real-IAD / rolled strip base	P-AP -2.92 while P-AUROC and AUPRO increase slightly.
Reconstruction / Real-IAD / end cap	P-AP -0.20; I-AUROC -2.33.
Memory / Real-IAD / u block	P-AP +3.32 while absolute q95 and q99 increase.

Appendix

Table 27: Selected seed-0 negative and metric-trade-off cases.

Dataset	Baseline hours	BoundarySupport hours	Ratio
MVTec	5.81	15.17	2.61 $\times$
VisA	4.99	12.71	2.55 $\times$
Real-IAD	12.60	31.21	2.48 $\times$

Appendix

Table 28: Seed-0 reconstruction category-GPU hours. Totals are sums over categories, not parallel wall-clock time.

Explore similar work

Aug 22, 2026

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

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.

Hannaneh Kalantary, Javad Khoramdel

University of Padua · University of Vienna

Industrial Anomaly Detection Anomaly Detection Few-Shot Anomaly Detection

Ask PDF

Aug 27, 2026

ShiftSplit-AD: Separating Domain Shift from Defects in Foundation-Feature Visual Anomaly Detection

Visual anomaly detectors based on frozen foundation-model features commonly score distances from test patches to a memory of normal features. Benign acquisition changes can also enlarge these distances, confounding domain variation with defects. We investigate whether structured decomposition of nearest-normal DINOv2 residuals can suppress shift-induced evidence while retaining unseen defects. ShiftSplit-AD decomposes the patch residual matrix into low-rank and row-sparse components and scores the sparse component, with an optional low-rank/sparse fusion. The experiments expose a central trade-off rather than a universal separation: genuine defects can contain correlated, low-dimensional structure, so filtering broad residual activity may also remove defect information. On AeBAD-S, using settings fixed after Bottle development, sparse-only scoring improves image AUROC from 0.6780 to 0.7294 and AUPRC from 0.8052 to 0.8465. Paired bootstrap 95% intervals for the improvements are [0.0238, 0.0808] and [0.0170, 0.0650], respectively. However, sparse-only scoring reduces mean clean AUROC from 0.9890 to 0.9133 on four held-out MVTec categories and degrades Bottle localization. These findings show that residual decomposition can help when domain shift strongly contaminates anomaly evidence, but preserving defect structure remains the limiting problem.

Muhamathu Ameer Ali Aacaas Muhamath

Department of Electrical Engineering University of Moratuwa Moratuwa, Sri Lanka

Anomaly Detection Anomaly Localization

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