pith. sign in

arxiv: 2508.04818 · v2 · submitted 2025-08-06 · 💻 cs.CV · eess.IV· stat.ML

Single-Step Reconstruction-Free Anomaly Detection and Segmentation via Diffusion Models

Mehrdad Moradi , Marco Grasso , Bianca Maria Colosimo , Kamran Paynabar This is my paper

Pith reviewed 2026-05-18 23:52 UTC · model grok-4.3

classification 💻 cs.CV eess.IVstat.ML
keywords anomaly detectiondiffusion modelsreconstruction-freeattention mechanismsimage segmentationMVTec-ADreal-time detection3D-printed materials
0
0 comments X

The pith

RADAR generates anomaly maps directly from diffusion models without reconstructing the input image.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper proposes a method called RADAR that detects and segments anomalies using diffusion models trained only on normal data. Existing diffusion approaches perturb an image and then perform many reverse sampling steps to reconstruct a normal version, which is slow and can produce a wrong normal pattern for subtle defects. RADAR instead uses the model's internal attention to output an anomaly map in one step, avoiding the need to choose a noise level. This matters for practical uses such as inspecting manufactured parts where speed and precision both count. If the approach holds, diffusion models become viable for real-time anomaly tasks on standard benchmarks like MVTec-AD.

Core claim

The authors claim that their RADAR method directly produces anomaly maps from the attention mechanisms of a diffusion model trained solely on normal images, eliminating the reconstruction step entirely. This sidesteps the computational expense of multiple sampling steps, the risk that reconstruction yields a different normal pattern, and the difficulty of selecting the right noise level without knowing the anomalies in advance. On the MVTec-AD and 3D-printed material datasets, the method outperforms previous diffusion-based and statistical models in accuracy, precision, recall, and F1 score, delivering 7% and 13% improvements in F1 score respectively.

What carries the argument

Attention mechanisms inside the diffusion model that isolate anomalous regions directly from the input without any reverse sampling or noise-level selection.

If this is right

  • Anomaly detection runs in a single forward pass, enabling real-time industrial inspection.
  • No application-specific tuning of intermediate noise levels is required.
  • Reconstruction errors for complex or subtle anomalies are avoided by design.
  • The same trained model supports both detection and pixel-level segmentation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The direct attention approach could transfer to other generative models if their internal representations can be read out similarly.
  • Resource-limited settings such as edge devices would benefit from the reduced sampling cost.
  • Temporal extensions to video anomaly detection might be feasible by adding attention across frames.

Load-bearing premise

The attention mechanism inside the diffusion model can reliably isolate anomalous regions from normal training data alone without any reconstruction step or explicit choice of noise level at inference time.

What would settle it

A head-to-head test on MVTec-AD or the 3D-printed dataset in which RADAR's direct anomaly maps produce lower F1 scores than a reconstruction-based diffusion baseline that uses the same backbone network.

Figures

Figures reproduced from arXiv: 2508.04818 by Bianca Maria Colosimo, Kamran Paynabar, Marco Grasso, Mehrdad Moradi.

Figure 1
Figure 1. Figure 1: Overview of the RADAR anomaly detection framework. After training [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Forward and backward diffusion processes. The forward process [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: To handle low-data scenarios in engineering applications such as additive manufacturing, we adopt a patch-based train￾ing strategy by extracting small patches from the training images. This increases the training dataset size, reduces over￾fitting, and significantly lowers computational cost. For ex￾ample, dividing a 500×500 image into 25×25 patches reduces GPU memory usage to only 0.25% (1/400) of what wo… view at source ↗
Figure 3
Figure 3. Figure 3: DDPM as a noise prediction model. A Gaussian noise and time [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Single-step anomaly map generation. For normal (in-control) data, [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: U-Net architecture. The network consists of a downsampling block, [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative results on the Tile dataset. From left to right: defective [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative results on the 3D Print dataset. The left column displays [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
read the original abstract

Generative models have demonstrated significant success in anomaly detection and segmentation over the past decade. Recently, diffusion models have emerged as a powerful alternative, outperforming previous approaches such as GANs and VAEs. In typical diffusion-based anomaly detection, a model is trained on normal data, and during inference, anomalous images are perturbed to a predefined intermediate step in the forward diffusion process. The corresponding normal image is then reconstructed through iterative reverse sampling. However, reconstruction-based approaches present three major challenges: (1) the reconstruction process is computationally expensive due to multiple sampling steps, making real-time applications impractical; (2) for complex or subtle patterns, the reconstructed image may correspond to a different normal pattern rather than the original input; and (3) Choosing an appropriate intermediate noise level is challenging because it is application-dependent and often assumes prior knowledge of anomalies, an assumption that does not hold in unsupervised settings. We introduce Reconstruction-free Anomaly Detection with Attention-based diffusion models in Real-time (RADAR), which overcomes the limitations of reconstruction-based anomaly detection. Unlike current SOTA methods that reconstruct the input image, RADAR directly produces anomaly maps from the diffusion model, improving both detection accuracy and computational efficiency. We evaluate RADAR on real-world 3D-printed material and the MVTec-AD dataset. Our approach surpasses state-of-the-art diffusion-based and statistical machine learning models across all key metrics, including accuracy, precision, recall, and F1 score. Specifically, RADAR improves F1 score by 7% on MVTec-AD and 13% on the 3D-printed material dataset compared to the next best model. Code available at: https://github.com/mehrdadmoradi124/RADAR

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces RADAR, a reconstruction-free anomaly detection and segmentation method that uses attention maps extracted from a diffusion model trained solely on normal data. It claims to directly generate anomaly maps in a single step, bypassing the iterative reverse sampling, potential reconstruction mismatches for subtle anomalies, and application-dependent noise-level selection that affect prior diffusion-based approaches. Evaluations on the MVTec-AD and 3D-printed material datasets report F1-score gains of 7% and 13% over state-of-the-art diffusion and statistical baselines, with code released for reproducibility.

Significance. If the central claims hold after clarification, the work would offer a practical advance for real-time anomaly detection by removing reconstruction overhead while improving accuracy. The explicit code release supports reproducibility and allows direct verification of the attention-based extraction procedure.

major comments (2)
  1. [Abstract and §3] Abstract and §3 (method description): the claim that anomaly maps are produced without any explicit noise-level choice at inference is load-bearing, yet the precise extraction procedure (including whether a fixed timestep t is used or how attention is isolated from the UNet) is not specified; this leaves open whether the method implicitly reintroduces challenge (3) via a schedule-dependent choice.
  2. [§4] §4 (experiments): quantitative details on the training protocol, attention architecture within the diffusion UNet, and the exact formula for converting attention features to anomaly maps are absent, preventing verification that the reported F1 gains are independent of hyperparameter choices made on the test sets.
minor comments (2)
  1. [Abstract] Abstract: expand the description of the three challenges to include a brief reference to how RADAR specifically resolves each one with a pointer to the relevant section or equation.
  2. [§5] §5 (results): include a table or figure showing runtime comparisons to confirm the real-time efficiency claim relative to multi-step reconstruction baselines.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We are grateful to the referee for their insightful comments, which have helped us improve the clarity of our manuscript. Below, we provide point-by-point responses to the major comments. We have made revisions to address the concerns raised regarding the method description and experimental details.

read point-by-point responses

  1. Referee: [Abstract and §3] Abstract and §3 (method description): the claim that anomaly maps are produced without any explicit noise-level choice at inference is load-bearing, yet the precise extraction procedure (including whether a fixed timestep t is used or how attention is isolated from the UNet) is not specified; this leaves open whether the method implicitly reintroduces challenge (3) via a schedule-dependent choice.

    Authors: We thank the referee for highlighting this critical aspect of our contribution. We agree that the original description in the abstract and §3 would benefit from greater precision to fully substantiate the claim. In the revised manuscript, we have expanded §3 to explicitly detail the anomaly map extraction procedure: a single forward pass is performed through the pre-trained diffusion UNet on the input image with no added noise (fixed timestep equivalent to the clean data regime), and attention maps are isolated from designated layers of the UNet. The resulting anomaly map is generated directly from these attention features via a fixed aggregation operation that does not involve any input-dependent or anomaly-dependent selection of noise levels. This fixed procedure is set once based on the training distribution and does not reintroduce challenge (3). We have also added a clarifying figure and pseudocode. revision: yes

  2. Referee: [§4] §4 (experiments): quantitative details on the training protocol, attention architecture within the diffusion UNet, and the exact formula for converting attention features to anomaly maps are absent, preventing verification that the reported F1 gains are independent of hyperparameter choices made on the test sets.

    Authors: We agree that the experimental section requires additional quantitative details to support reproducibility and to demonstrate that the reported improvements are not sensitive to test-set-specific choices. In the revised §4, we have included the full training protocol (optimizer, learning rate schedule, number of epochs, and batch size), the specific attention layers and heads within the diffusion UNet architecture, and the exact mathematical formula used to derive anomaly maps from the extracted attention features. These details confirm that all hyperparameters were selected using only the normal training data (with a small validation split from the training set) and were not tuned on the test sets. The publicly released code further documents the implementation. revision: yes

Circularity Check

0 steps flagged

No circularity: RADAR presents a novel inference procedure independent of fitted parameters or self-citation chains

full rationale

The paper defines RADAR as a direct anomaly-map extraction method via attention in a diffusion model trained solely on normal data, explicitly avoiding reconstruction and explicit timestep selection at inference. No equations or claims in the abstract or described method reduce the reported F1 gains or anomaly isolation to a parameter fitted on the same data or to a prior self-citation that is itself unverified. The three challenges are addressed by construction of the new procedure rather than by re-deriving inputs. This is the common case of a self-contained methodological contribution evaluated on external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The approach rests on standard unsupervised anomaly-detection assumptions plus the novel claim that attention features inside a diffusion model suffice for direct map generation. No new physical entities or ad-hoc constants are introduced in the abstract.

axioms (2)
  • domain assumption Normal training images contain no anomalies and the model learns a distribution over defect-free data only.
    Stated in the description of training on normal data and the three challenges of reconstruction methods.
  • ad hoc to paper Attention maps extracted from the diffusion process at a single step are sufficient to localize anomalies without iterative sampling.
    This is the central modeling choice that enables the reconstruction-free claim.

pith-pipeline@v0.9.0 · 5859 in / 1359 out tokens · 37218 ms · 2026-05-18T23:52:48.844531+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Backbone-Equated Diffusion OOD via Sparse Internal Snapshots

    cs.LG 2026-05 unverdicted novelty 6.0

    Sparse internal snapshots at canonical low-noise levels from frozen diffusion backbones suffice for competitive out-of-distribution detection without full trajectories or large heads.

Reference graph

Works this paper leans on

44 extracted references · 44 canonical work pages · cited by 1 Pith paper · 2 internal anchors

  1. [1]

    Estimation and monitoring of product aesthetics: application to manufacturing of “engineered stone

    J. J. Liu and J. F. MacGregor, “Estimation and monitoring of product aesthetics: application to manufacturing of “engineered stone” counter- tops,” Machine Vision and Applications , vol. 16, no. 6, pp. 374–383, 2006

    work page 2006

  2. [2]

    Computer-aided visual inspection of surface defects in ceramic capacitor chips,

    H. Lin, “Computer-aided visual inspection of surface defects in ceramic capacitor chips,” Journal of Materials Processing Technology, vol. 189, no. 1–3, pp. 19–25, 2007

    work page 2007

  3. [3]

    Softwood lumber grading through on-line multivariate image analysis techniques,

    M. H. Bharati, J. F. MacGregor, and W. Tropper, “Softwood lumber grading through on-line multivariate image analysis techniques,” Indus- trial & Engineering Chemistry Research, vol. 42, no. 21, pp. 5345–5353, 2003

    work page 2003

  4. [4]

    Random heterogeneous materials via texture synthesis,

    X. Liu and V . Shapiro, “Random heterogeneous materials via texture synthesis,” Computational Materials Science , vol. 99, pp. 177–189, 2015

    work page 2015

  5. [5]

    Statistical description of microstructures,

    S. Torquato, “Statistical description of microstructures,” Annual Review of Materials Research , vol. 32, no. 1, pp. 77–111, 2002

    work page 2002

  6. [6]

    A monitoring and diagnostic approach for stochastic textured surfaces,

    A. T. Bui and D. W. Apley, “A monitoring and diagnostic approach for stochastic textured surfaces,” Technometrics, vol. 60, no. 1, pp. 1–13, 2017

    work page 2017

  7. [7]

    In-situ monitoring of image texturing via random forests and clustering with applications to additive manufacturing,

    F. Caltanissetta, L. Bertoli, and B. M. Colosimo, “In-situ monitoring of image texturing via random forests and clustering with applications to additive manufacturing,” IISE Transactions, vol. 56, no. 10, pp. 1070– 1084, 2024

    work page 2024

  8. [8]

    Online automatic anomaly detection for photovoltaic systems using thermography imaging and low rank matrix decomposition,

    Q. Wang, K. Paynabar, and M. Pacella, “Online automatic anomaly detection for photovoltaic systems using thermography imaging and low rank matrix decomposition,” Journal of Quality Technology , vol. 54, no. 5, pp. 503–516, 2022. [Online]. Available: https: //doi.org/10.1080/00224065.2021.1948372

    work page doi:10.1080/00224065.2021.1948372 2022

  9. [9]

    Anomaly detection in pv systems using constrained low-rank and sparse decomposition,

    W. Yang, D. Fregosi, M. Bolen, and K. Paynabar, “Anomaly detection in pv systems using constrained low-rank and sparse decomposition,” IISE Transactions , vol. 57, no. 6, pp. 607–620, 2025. [Online]. Available: https://doi.org/10.1080/24725854.2024.2339345

    work page doi:10.1080/24725854.2024.2339345 2025

  10. [10]

    A novel representation of periodic pattern and its application to untrained anomaly detection,

    P. Ye, C. Tao, and J. Du, “A novel representation of periodic pattern and its application to untrained anomaly detection,” IISE Transactions, 2025, published online: 10 Jul 2025. [Online]. Available: https://doi.org/10.1080/24725854.2025.2526776

    work page doi:10.1080/24725854.2025.2526776 2025

  11. [11]

    IISE Transactions52(11), 1204–1217 (2020) https://doi.org/10.1080/24725854.2019.1701753

    H. Yan, K. Paynabar, and J. Shi, “Akm2d: An adaptive framework for online sensing and anomaly quantification,” IISE Transactions , vol. 52, no. 9, pp. 1032–1046, 2020. [Online]. Available: https: //doi.org/10.1080/24725854.2019.1681606

    work page doi:10.1080/24725854.2019.1681606 2020

  12. [12]

    Robust principal component analysis?

    E. J. Cand `es, X. Li, Y . Ma, and J. Wright, “Robust principal component analysis?” Journal of the ACM (JACM) , vol. 58, no. 3, pp. 1–37, 2011. [Online]. Available: https://doi.org/10.1145/1970392.1970395

    work page doi:10.1145/1970392.1970395 2011

  13. [13]

    Modeling an Augmented Lagrangian for Blackbox Constrained Optimization

    H. Yan, K. Paynabar, and J. Shi, “Anomaly detection in images with smooth background via smooth-sparse decomposition,” Technometrics, vol. 59, no. 1, pp. 102–114, 2017. [Online]. Available: https: //doi.org/10.1080/00401706.2015.1102764

    work page doi:10.1080/00401706.2015.1102764 2017

  14. [14]

    A cnn-based adaptive surface monitoring system for fused deposition modeling,

    Y . Wang et al. , “A cnn-based adaptive surface monitoring system for fused deposition modeling,” IEEE/ASME Transactions on Mechatronics, vol. 25, no. 5, pp. 2287–2296, Oct. 2020

    work page 2020

  15. [15]

    Support vector machine and convolutional neural network based approaches for defect detection in fused filament fabrication,

    B. Narayanan, N. Powar, G. Loughnane, and K. Beigh, “Support vector machine and convolutional neural network based approaches for defect detection in fused filament fabrication,” 08 2019

    work page 2019

  16. [16]

    Autonomous in-situ correction of fused deposition modeling printers using computer vision and deep learning,

    Z. Jin, Z. Zhang, and G. X. Gu, “Autonomous in-situ correction of fused deposition modeling printers using computer vision and deep learning,” Manufacturing Letters, vol. 22, pp. 11–15, 2019

    work page 2019

  17. [17]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” in Advances in Neural Information Processing Systems , vol. 33, 2020, pp. 6840–6851

    work page 2020

  18. [18]

    Generative adversarial nets,

    I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial nets,” Advances in neural information processing systems , vol. 27, 2014

    work page 2014

  19. [19]

    Auto-encoding variational bayes,

    D. P. Kingma, M. Welling et al. , “Auto-encoding variational bayes,” 2013

    work page 2013

  20. [20]

    Diffusion models beat gans on image synthesis,

    P. Dhariwal and A. Nichol, “Diffusion models beat gans on image synthesis,” Advances in neural information processing systems , vol. 34, pp. 8780–8794, 2021

    work page 2021

  21. [21]

    AnoDDPM: Anomaly Detection with Denoising Diffusion Probabilistic Models using Simplex Noise,

    J. Wyatt, A. Leach, S. M. Schmon, and C. G. Willcocks, “AnoDDPM: Anomaly Detection with Denoising Diffusion Probabilistic Models using Simplex Noise,” in 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) . New Orleans, LA, USA: IEEE, Jun. 2022, pp. 649–655. [Online]. Available: https://ieeexplore.ieee.org/document/9857019/

    work page arXiv 2022

  22. [22]

    Paedid: Patch autoencoder-based deep image decomposition for pixel-level defective region segmentation,

    S. Mou, M. Cao, H. Bai, P. Huang, J. Shi, and J. Shan, “Paedid: Patch autoencoder-based deep image decomposition for pixel-level defective region segmentation,” IISE Transactions, vol. 56, no. 9, pp. 917–931, 2024

    work page 2024

  23. [23]

    Industrial image anomaly localization based on gaussian clustering of pretrained feature,

    Q. Wan, L. Gao, X. Li, and L. Wen, “Industrial image anomaly localization based on gaussian clustering of pretrained feature,” IEEE Transactions on Industrial Electronics , vol. 69, no. 6, pp. 6182–6192, 2021

    work page 2021

  24. [24]

    Unsupervised image anomaly detection and localization in industry based on self-updated memory and center clustering,

    Y . Liu, X. Gao, J. Z. Wen, and H. Luo, “Unsupervised image anomaly detection and localization in industry based on self-updated memory and center clustering,” IEEE Transactions on Instrumentation and Measure- ment, vol. 72, pp. 1–10, 2023

    work page 2023

  25. [25]

    Unsupervised image anomaly detection and segmentation based on pretrained feature mapping,

    Q. Wan, L. Gao, X. Li, and L. Wen, “Unsupervised image anomaly detection and segmentation based on pretrained feature mapping,” IEEE Transactions on Industrial Informatics , vol. 19, no. 3, pp. 2330–2339, 2022

    work page 2022

  26. [26]

    Score-Based Generative Modeling through Stochastic Differential Equations

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differ- ential equations,” arXiv preprint arXiv:2011.13456 , 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  27. [27]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10 684–10 695

    work page 2022

  28. [28]

    Classifier-Free Diffusion Guidance

    J. Ho and T. Salimans, “Classifier-free diffusion guidance,” arXiv preprint arXiv:2207.12598, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  29. [29]

    Unsupervised surface anomaly detection with diffusion probabilistic model,

    X. Zhang, N. Li, J. Li, T. Dai, Y . Jiang, and S. Xia, “Unsupervised surface anomaly detection with diffusion probabilistic model,” in Pro- ceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 6759–6768

    work page 2023

  30. [30]

    Dzad: Diffusion-based zero-shot anomaly detection,

    T. Zhang, L. Gao, X. Li, and Y . Gao, “Dzad: Diffusion-based zero-shot anomaly detection,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 10, 2025, pp. 10 131–10 138

    work page 2025

  31. [31]

    A diffusion-based framework for multi-class anomaly detection,

    H. He, J. Zhang, H. Chen, X. Chen, Z. Li, X. Chen, Y . Wang, C. Wang, and L. Xie, “A diffusion-based framework for multi-class anomaly detection,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 8, 2024, pp. 8472–8480

    work page 2024

  32. [32]

    Diffusionad: Norm-guided one-step denoising diffusion for anomaly detection,

    H. Zhang, Z. Wang, D. Zeng, Z. Wu, and Y .-G. Jiang, “Diffusionad: Norm-guided one-step denoising diffusion for anomaly detection,” IEEE Transactions on Pattern Analysis and Machine Intelligence , 2025

    work page 2025

  33. [33]

    Dynamic addition of noise in a diffusion model for anomaly detection,

    J. Tebbe and J. Tayyub, “Dynamic addition of noise in a diffusion model for anomaly detection,” in 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) , 2024, pp. 3940– 3949

    work page 2022

  34. [34]

    Feature prediction diffusion model for video anomaly detection,

    C. Yan, S. Zhang, Y . Liu, G. Pang, and W. Wang, “Feature prediction diffusion model for video anomaly detection,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , 2023, pp. 5504–5514

    work page 2023

  35. [35]

    Time series anomaly detection using diffusion-based models,

    I. Pintilie, A. Manolache, and F. Brad, “Time series anomaly detection using diffusion-based models,” in 2022 IEEE International Conference on Data Mining Workshops (ICDMW) , 2023, pp. 570–578

    work page 2022

  36. [36]

    Multimodal motion conditioned diffusion model for skeleton-based video anomaly detection,

    A. Flaborea, L. Collorone, G. M. D. Di Melendugno, S. D’Arrigo, B. Prenkaj, and F. Galasso, “Multimodal motion conditioned diffusion model for skeleton-based video anomaly detection,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , 2023, pp. 10 284–10 295

    work page 2023

  37. [37]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer, 2015, pp. 234–241

    work page 2015

  38. [38]

    Isolation forest,

    F. T. Liu, K. M. Ting, and Z.-H. Zhou, “Isolation forest,” in 2008 Eighth IEEE International Conference on Data Mining , 2008, pp. 413–422

    work page 2008

  39. [39]

    Lof: identifying density-based local outliers,

    M. M. Breunig, H.-P. Kriegel, R. T. Ng, and J. Sander, “Lof: identifying density-based local outliers,” in Proceedings of the 2000 ACM SIGMOD international conference on Management of data , 2000, pp. 93–104

    work page 2000

  40. [40]

    Support vector method for novelty detection,

    B. Sch ¨olkopf, R. C. Williamson, A. Smola, J. Shawe-Taylor, and J. Platt, “Support vector method for novelty detection,” Advances in neural information processing systems , vol. 12, 1999

    work page 1999

  41. [41]

    Approximate training of one-class sup- port vector machines using expected margin,

    S. Kang, D. Kim, and S. Cho, “Approximate training of one-class sup- port vector machines using expected margin,” Computers & Industrial Engineering, vol. 130, pp. 772–778, 2019

    work page 2019

  42. [42]

    A review of local outlier factor algorithms for outlier detection in big data streams,

    O. Alghushairy, R. Alsini, T. Soule, and X. Ma, “A review of local outlier factor algorithms for outlier detection in big data streams,” Big Data and Cognitive Computing , vol. 5, no. 1, p. 1, 2020

    work page 2020

  43. [43]

    MVTec AD — A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection,

    P. Bergmann, M. Fauser, D. Sattlegger, and C. Steger, “MVTec AD — A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection,” in 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . Long Beach, CA, USA: IEEE, Jun. 2019, pp. 9584–9592. [Online]. Available: https://ieeexplore.ieee.org/ document/8954181/

    work page arXiv 2019

  44. [44]

    A threshold selection method from gray-level histograms,

    N. Otsu, “A threshold selection method from gray-level histograms,” IEEE Transactions on Systems, Man, and Cybernetics , vol. 9, no. 1, pp. 62–66, 1979

    work page 1979