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

Ayusha Abbas; Saram Abbas; Kabita Adhikari

cs.LGAug 18, 2026

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

Authors: Ayusha Abbas, Saram Abbas, Kabita Adhikari

Organizations: School of Engineering, Newcastle University, Newcastle upon Tyne, UK

Abstract

Anomaly detectors are hardest to deploy exactly where training data is scarcest: a newly commissioned production line has a handful of verified "golden" samples and no machine-learning engineer on the factory floor. We present a training-free human-in-the-loop framework in which a domain expert corrects a PatchCore detector by direct memory bank editing: no retraining, no gradients, no original training data. A false-positive correction inserts the reviewed image's normal patches through a self-calibrating novelty gate admitting only those beyond the median pool-normal nearest-neighbour distance. From a bank built on only ten golden samples, operator corrections close a median 66% of the gap to an uncorrected fully trained bank (mean 80%, raised by three categories that overshoot parity), significantly improving 12 of 15 MVTec AD categories and harming none: ten samples plus corrections outperform hundreds of samples without them. On already-trained banks the headroom is smaller and concentrated where the bank undersamples normal appearance (gated: toothbrush +0.10, metal nut +0.09, zipper +0.05, screw +0.05), and no category except grid is significantly harmed. Evaluation uses a held-out protocol (20 splits per category, Holm-corrected Wilcoxon), because corrected images entering the bank inflate naive evaluation toward AUROC 1.0 by memorisation. Passive and active querying are statistically indistinguishable; a matched-label-budget control attributes gains to deployment-time label production at 43% of exhaustive-review cost; a defect-memory extension fails decisively. Feedback is simulated from ground truth; live expert trials, where mislabelling is costliest on small banks, remain future work.

Figures & tables

Fig. 1: What training-free correction looks like on images the loop never touches. For toothbrush and metal nut, a held-out normal image is falsely flagged before correction (anomaly heatmap hot everywhere) and cleared after a single simulated review session (16–25 single-decision reviews for the seeds shown; the category means are $\sim$ 13 and $\sim$ 20), while a held-out defect remains confidently flagged, with the anomaly map concentrated on the defect itself (contaminated bristles; flipped part). Heatmaps share a colour scale within each row; scores are the image-level anomaly scores.

Method	No re-	Human	Deploy-	Specific
	training	feedback	ment-time	correction
Autoencoder	✗	✗	✗	✗
AnoGAN	✗	✗	✗	✗
PatchCore	✓	✗	✗	✗
Few-shot AD	Partial	✗	Partial	✗
Continual learning	✗	✗	✓	✗

TABLE I: Positioning of our work relative to prior methods.

Panel of Figure 3 — Fig. 2: Framework overview. A deployed PatchCore model flags errors; a quality engineer (not an ML engineer) resolves each with a single binary decision, which is applied as a direct edit to the memory bank and takes effect immediately. The conventional remedy (bottom, dashed) requires an ML team and a retraining pipeline.

Fig. 3: The novelty gate on a real false positive (toothbrush, seed 10). The threshold

\tau_{\text{nov}}

is the median 1-NN distance of all pool-normal patches to the baseline bank (black outline); of the flagged image’s 784 patches, those within existing coverage (left of

\tau

) are rejected and only the genuinely novel ones (right of

\tau

) are inserted.

Category	Type	$\boldsymbol{\sigma}$	Base.	Train	Test (g/a)
Bottle	Object	50.5	1.0000	209	20/63
Leather	Texture	23.1	1.0000	245	32/92
Carpet	Texture	9.5	0.9932	280	28/89
Wood	Texture	7.6	0.9912	247	19/60
Hazelnut	Object	7.0	0.9964	391	40/70
Tile	Texture	6.0	0.9776	230	33/84

TABLE II: All 15 MVTec AD categories, ordered by separation index

\sigma

(descending).

Fig. 4: Anomaly score distributions for four representative categories spanning the full difficulty range (blue = normal, red = anomalous). Overlap decreases as

\sigma

increases, from heavily overlapping (toothbrush) to fully separated (leather).

Category	Type	Sep.( $\sigma$ )	Full-test Base.	HO Base.	Pass. HO Final	Pass. $\Delta$	Act. $\Delta$	Gated $\Delta$
Toothbrush	Object	2.6	0.8139	0.8264	0.9722	$\mathbf{+0.146}^{*}$	$\mathbf{+0.146}$	$\mathbf{+0.101}^{*}$
Grid	Texture	2.6	0.8839	0.8662	0.8190	$-0.047^{*}$	$-0.056$	$-0.031^{*}$
Screw	Object	3.0	0.8703	0.8703	0.9105	$\mathbf{+0.040}^{*}$	$\mathbf{+0.043}$	$\mathbf{+0.048}^{*}$
Transistor	Object	3.7	0.9558	0.9623	0.9379	$-0.024^{*}$	$-0.031$	$+0.005$
Metal Nut	Object	4.3	0.8803	0.9000	0.9980	$\mathbf{+0.098}^{*}$	$\mathbf{+0.098}$	$\mathbf{+0.089}^{*}$
Pill	Object	4.3	0.9242	0.9283	0.9051	$-0.023^{*}$	$-0.026$	$+0.007$

TABLE III: Complete 15-category HITL results under the held-out protocol (means over 20 seeds). HO = held-out set, never corrected. Bold improvements are consistent across seeds (

\geq

15/20 seeds improved); negative improvements are shown as-is; the ungated correction loop is not beneficial on every category. “Gated

\Delta

” is the passive loop with novelty-gated insertion (Section III-B2 ), the recommended deployment configuration. ∗ marks a

\Delta

significantly different from zero (two-sided Wilcoxon signed-rank on per-seed deltas, Holm-corrected across 15 categories,

\alpha=0.05

); it is reported for the passive and gated configurations. Active deltas are not separately tested against zero, because no category shows a significant passive-vs-active difference under the same test.

Category	Cold Base	Post-HITL	Full Ref.	$\boldsymbol{\Delta}$	Recovery	Reviews	Labels to Parity
Bottle	0.957	0.996	1.000	$+0.039^{*}$	0.92	14	5 (7/20)
Cable	0.756	0.866	0.926	$+0.110^{*}$	0.65	50	35 (1/20)
Capsule	0.560	0.699	0.940	$+0.139^{*}$	0.37	80	— (0/20)
Carpet	0.987	0.996	0.993	$+0.008^{*}$	—	18	8 (2/20)
Grid	0.681	0.719	0.866	$+0.038$	0.20	37	24 (6/20)
Hazelnut	0.921	0.998	0.996	$+0.077^{*}$	1.03	16	4 (14/20)

TABLE IV: Cold-start results: initial bank built from only

k=10

training normals (78 patches), passive querying with novelty-gated insertion, 90-round budget, means over 20 seeds. “Full Ref.” = full-training-bank held-out baseline on the same seed (paired). Recovery = fraction of the cold-to-full gap closed (shown where the gap exceeds 0.02). “Labels to parity” = median reviews to reach the full-training reference, with the number of seeds reaching it. ∗ = Holm-significant

\Delta

(Wilcoxon,

\alpha=0.05

); no category is significantly negative.

Fig. 5: Held-out trajectories (mean

\pm 1

std over 20 seeds). (a–c) Cold-start recovery from 10 golden samples with the novelty gate: metal nut crosses its full-training baseline around round 50 and ends above it; zipper converges to just below parity after a transient early dip (corrections on a 78-patch bank briefly disturb scores before coverage accumulates); screw is still climbing when the 90-round budget ends, far from parity (the budget-limited case). (d) Pill on a fully trained bank: the ungated loop drifts below baseline while the novelty-gated loop stays above it, showing the gate’s effect in a single panel.

Fig. 6: Held-out ROC curves before (grey) and after correction under passive (blue) and active (red dashed) querying, for the three largest-improvement categories at their median-improvement seed. The corrected ROC dominates the baseline everywhere; passive and active endpoints are nearly identical.

Fig. 7: Held-out AUROC over correction rounds for passive (blue) and active (red) querying, on six categories spanning the difficulty range (mean over 20 seeds; bands are the 95% confidence interval of the mean,

t

-based with

n=20

, not the seed-to-seed spread). The bands overlap at every round except on metal nut, where active querying rises faster over the first few rounds before both strategies converge to the same endpoint: a difference in speed, not destination.

Fig. 8: Per-seed distributions of held-out AUROC improvement (20 seeds per category; blue = passive, red = active). Toothbrush, metal nut, zipper and screw improve consistently; grid, transistor, pill, capsule, cable and tile centre at or below zero. Passive and active distributions overlap heavily on every category.

Mode	HO Final	$\boldsymbol{\Delta}$	HO AUCC	FP/FN corr.
FP-only	0.9722	$+0.146$	0.709	7.5/0
Both	0.9722	$+0.146$	0.563	7.6/5.0
FN-only	0.8139	$-0.013$	0.004	0/8.8

TABLE V: FP vs. FN correction ablation, toothbrush passive, held-out protocol (means over 20 seeds). Restricted modes exclude the skipped error type from review;

\Delta

is held-out final minus baseline (0.826).

Fig. 9: FP vs. FN correction mix across all 15 categories (mean counts over 20 seeds). Blue = FP, red = FN. Bottle and leather are ceiling controls.

Explore similar work

Training-Free Logical and Structural Anomaly Detection via Calibrated Fusion
Sep 4, 2026Changyi Li, Miao Yu, Kai Dong +1Anomaly Detection Industrial Anomaly Detection
Aalto University, 02150 Espoo, Finland · Harbin Engineering University, 150001 Harbin, China
ProCon: Projection-Consistency Memory for Training-Free Anomaly Detection
Jul 6, 2026Joongwon Chae, Lihui Luo, Yang Liu +4Industrial Anomaly Detection Anomaly Localization
Tsinghua University Shenzhen International Graduate School, Shenzhen, China · Affiliated Fifth Hospital, Wenzhou Medical University, Wenzhou, Zhejiang, China · Chinese Medicine Guangdong Laboratory +1
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

Sep 4, 2026J/K move · Enter open · S save

Training-Free Logical and Structural Anomaly Detection via Calibrated Fusion

Changyi Li, Miao Yu, Kai Dong, Yu Xiao

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

Anomaly Detection Industrial Anomaly Detection

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

Training-Free Logical and Structural Anomaly Detection via Calibrated Fusion

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

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