Incremental dimension reduction for efficient and accurate visual anomaly detection

Teng-Yok Lee

cs.CVFeb 27, 2026

Incremental dimension reduction for efficient and accurate visual anomaly detection

Authors: Teng-Yok Lee

Organizations: Information R&D Technology Center, Mitsubishi Electric Coporation

Abstract

While nowadays visual anomaly detection algorithms use deep neural networks to extract salient features from images, the high dimensionality of extracted features makes it difficult to apply those algorithms to large data with 1000s of images. To address this issue, we present an incremental dimension reduction algorithm to reduce the extracted features. While our algorithm essentially computes truncated singular value decomposition of these features, other than processing all vectors at once, our algorithm groups the vectors into batches. At each batch, our algorithm updates the truncated singular values and vectors that represent all visited vectors, and reduces each batch by its own singular values and vectors so they can be stored in the memory with low overhead. After processing all batches, we re-transform these batch-wise singular vectors to the space spanned by the singular vectors of all features. We show that our algorithm can accelerate the training of state-of-the-art anomaly detection algorithm with close accuracy.

Figures & tables

Figure 1: Overview of the algorithm that use a batch of vectors

X_{b}

to update the singular values

S^{1,b}

and vectors

U^{1,b}

. The variables in the dashed boxes will be released after the corresponding operation.

Figure 2: Rotation of all projected $V_{b}$ to the final space of $U^{1,B}$ . The variables in the dashed boxes will be released after the corresponding operation.

		WideResNet50 [ 11 ]							ResNet18 [ 5 ]
Update parameter	$k$	N/A	128			256			N/A	128			256
Update parameter	$n_{b}$	N/A	16K	32K	All	16K	32K	All	N/A	16K	32K	All	16K	32K	All
Image AUROC	All	99.1	98.9	98.9	98.9	99.1	99.1	99.0	97.3	96.9	97.1	96.9	97.2	97.2	97.1
	Texture	99.3	98.9	99.1	99.0	99.3	99.2	99.2	97.9	96.7	96.9	96.3	97.5	97.3	97.1
	Object	99.0	98.9	98.9	98.9	99.0	99.0	99.0	97.0	97.1	97.2	97.1	97.0	97.2	97.1
Pixel AUTOC	All	97.9	97.7	97.7	97.7	97.8	97.8	97.8	96.9	96.6	96.6	96.5	96.8	96.8	96.8

Table 1: Average AUROC on MVTec AD [ 1 ] . Different numbers (

k

) of singular values and batch sizes (

n_{b}

) were tested when using dimension reduction.

$k$	N/A	128			256
$n_{b}$	N/A	16K	32K	All	16K	32K	All
carpet	99.3	98.0	98.1	98.1	98.7	98.8	98.9
grid	99.3	98.9	99.5	99.0	99.5	99.1	99.2
leather	100.0	100.0	100.0	100.0	100.0	100.0	100.0
tile	99.3	99.3	99.5	99.4	99.3	99.2	99.3
wood	98.9	98.4	98.3	98.5	98.8	98.8	98.8

Table 2: Category-wise Image AUROC (%) on MVTec AD [ 1 ] with WideResNet50 [ 11 ] . Different numbers (

k

) of singular values and batch sizes (

n_{b}

) were tested when using dimension reduction.

$k$	N/A	128			256
$n_{b}$	N/A	16K	32K	All	16K	32K	All
GPU	130	39	38	37	53	51	51
CPU	23998	3333	2991	2965	5314	5073	4947

Table 3: Training time (seconds) on category hazelnut of MVTec AD [ 1 ] with WideResNet50 [ 11 ] . Please refer to the caption of Table 2 for the meaning of

k

and

n_{b}

$k$	N/A	128			256
$n_{b}$	N/A	16K	32K	All	16K	32K	All
GPU	128	37	36	35	50	49	48
CPU	26473	2979	2975	2795	4952	4792	4841

Table 4: Training time (seconds) on category hazelnut of MVTec AD [ 1 ] with ResNet18 [ 5 ] . Please refer to the caption of Table 2 for the meaning of

k

and

n_{b}

AUROC	Method	L	Can.C.	Cho.C.	Cho.P.	Confet.	Gum.B.	Haz.T.	Lic.S.	Lollip.	Marsh.	Pep.C.	Avg.
Image	Ours	All	53.1	95.0	77.0	87.8	80.9	69.9	87.4	68.6	96.5	88.2	80.4
	Ours	0	51.8	96.3	79.7	89.3	76.1	71.0	83.8	67.7	96.0	85.9	79.8
	PaDiM	0	47.2	93.6	71.2	87.7	86.5	69.9	86.2	67.3	96.2	84.5	79.0
Pixel	Ours	All	91.6	97.3	86.6	98.2	93.5	85.7	94.1	96.3	97.0	97.3	93.8
	Ours	0	89.8	97.0	86.2	98.2	93.6	86.1	93.6	96.4	96.6	96.3	93.4
	PaDiM	0	92.7	96.6	87.1	97.6	93.9	88.1	93.7	96.8	97.5	97.0	94.1

Table 5: AUROC on the dataset Eyecandies [ 2 ] with PaDiM [ 4 ] and ours, which used

k=128

and

n_{b}=32

K. We used WideResNet50 [ 11 ] as the feature extractor for both. The values were measured on the test images captured under lighting condition 0, whereas column L means lighting conditions of training images.

Dimensionality Reduction

Dimensionality reduction maps high-dimensional data to compact representations for visualization, prediction, compression, and scientific modeling.

See all 145 papers

Deep Convolutional Large-Margin $\ell_p$ -SVDD for Visual Anomaly Detection
Jul 24, 2026Alireza Dastmalchi Saei, Shervin Rahimzadeh ArashlooKernel Methods One-Class Classification
Department of Computer Engineering, Bilkent University, 06800 Ankara, Türkiye
SubspaceAD: Training-Free Few-Shot Anomaly Detection via Subspace Modeling
Feb 26, 2026Camile Lendering, Erkut Akdag, Egor BondarevFew-Shot Anomaly Detection Anomaly Detection
AIMS Group, Department of Electrical Engineering, Eindhoven University of Technology
Augmenting Visual Anomaly Detection with Automated Interpretability
Sep 27, 2026Antonio De Santis, Arsenio Leo, Marco BrambillaInterpretable Anomaly Detection
Politecnico di Milano · Thales Alenia Space

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

Deep Convolutional Large-Margin $\ell_p$ -SVDD for Visual Anomaly Detection

Alireza Dastmalchi Saei, Shervin Rahimzadeh Arashloo

Department of Computer Engineering, Bilkent University, 06800 Ankara, Türkiye

Kernel Methods One-Class Classification

Open paper AskPDF arXiv

Visual anomaly detection requires adaptive representations and reliable decision boundaries, particularly when anomalous training samples are scarce and class distributions are highly imbalanced. Classical kernel-based methods yield principled geometric decision regions but typically operate on fixed features, while deep detectors learn task-specific representations but often fail to provide an explicit margin-aware kernel boundary. In this study, we propose DLM-SVDD, a deep large-margin novelty-detection framework that jointly learns convolutional features and an explicit kernel-based decision boundary. By drawing on the large-margin

\ell_p

-Support Vector Data Description (

\ell_p

-SVDD) approach, the proposed method performs explicit margin maximization and nonlinear slack penalization while adapting the representation to the target task. To train the proposed model, we present an optimization scheme that alternates between a Frank--Wolfe--based update of the convex dual boundary and a CNN update step operating on a smooth margin-violation loss induced by the recovered boundary. To improve scalability, we analyze the efficiency--accuracy trade-offs for different kernel approximation strategies, deriving practical propositions for large-scale anomaly detection. Extensive experiments on multiple standard benchmarks show consistent performance improvements over the baseline and strong overall performance compared with state-of-the-art methods while illustrating that the proposed joint representation--boundary learning scheme remains effective under severe imbalanced class distributions.

Abstract

Dimensionality Reduction

Explore similar work

Deep Convolutional Large-Margin ℓp\ell_pℓp​-SVDD for Visual Anomaly Detection

SubspaceAD: Training-Free Few-Shot Anomaly Detection via Subspace Modeling

Augmenting Visual Anomaly Detection with Automated Interpretability

Deep Convolutional Large-Margin $\ell_p$ -SVDD for Visual Anomaly Detection