Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

Center-aware Residual Anomaly Synthesis for Multi-class Industrial Anomaly Detection

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read CRAS claims a single unified model detects industrial defects across many categories by learning in a center-aware residual space, reporting 98.3% image-level AUROC on MVTec AD.

desk verdict Solid multi-class anomaly detection method with a genuinely useful center-residual idea; the headline MVTec numbers are likely optimistic because key hyperparameters are tuned directly on the test set. read the letter →

arxiv 2505.17551 v1 pith:OJUWIPTI submitted 2025-05-23 cs.CV

Qiyu Chen , Huiyuan Luo , Haiming Yao , Wei Luo , Zhen Qu , Chengkan Lv , Zhengtao Zhang This is my paper
classification cs.CV
keywords multi-classanomalydetectioncenter-awareresiduallearningdistance-guidedsynthesisindustrialvisualinspectionMVTecADunifiedmodel
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Most industrial anomaly detectors train one model per product category, so deploying them across a factory floor multiplies memory and complexity. This paper claims that a single unified model can detect defects across many categories at once, without a per-category accuracy penalty, by learning in a shared 'residual' space rather than in raw feature space. The proposed method, CRAS, couples every category's normal features to a common contextual center and synthesizes training anomalies whose noise is tuned by how far each normal patch sits from that center. On the MVTec AD benchmark it reports 98.3% image-level and 98.0% pixel-level AUROC, and it reports the best multi-class numbers on VisA and MPDD. If the claim holds, a factory can inspect an entire product line with one model, fewer false alarms, and a five-fold speedup over naive center search.

What carries the argument

The load-bearing object is the center-aware residual feature, defined as the concatenation $[u_i \mid u_i - p_i]$, where $u_i$ is the adapted feature of an image and $p_i$ is a recomposed center obtained by matching $u_i$ against per-class memory centers (MCMC) using cosine similarity, first globally and then patch-locally (HPI). The matching reduces center retrieval to $O(1)$ per patch compared with $O(|K|)$ for a brute-force nearest-neighbor search. A second mechanism, DAFS, synthesizes anomaly features $v_i = u_i + \alpha_i \odot g_i$ by scaling Gaussian noise $g_i$ at each position with a distance ratio $\alpha$ that is proportional to the noise norm and inversely proportional to the residual norm $r_i = \|u_i - p_i\|$, so that homogeneous normal patches receive stronger perturbations and heterogeneous ones receive weaker perturbations. The discriminator is trained with binary cross-entropy on the paired normal and synthetic residuals, and at test time the same residual concatenation is scored to produce the anomaly map.

What would settle it

Run CRAS on a category whose normal set contains two visually distinct modes, such as a metal part photographed from both the front and the back, and measure the pixel-level AUROC restricted to the second mode. If the model systematically raises high anomaly scores on the second mode's normal images, the single-center assumption is violated; a direct comparison would be the drop in per-category AUROC relative to a one-centre-per-mode variant.

Watch

Extended reading notes

Core claim

CRAS's central discovery is that the residual between a patch feature and its class's contextual center is a better substrate for multi-class anomaly detection than the raw feature itself. The paper shows that raw normal features from different classes sit far apart, so a unified boundary either swallows anomalies (missed detections) or fragments into per-class regions. By matching each image to its class center through a global-to-local cosine search and recomposing a per-patch center, CRAS forms residual vectors $u_i - p_i$ that are compact and class-agnostic; the discriminator then trains to separate normal residuals from residuals of synthetically perturbed features $v_i - p_i$. The synthetic perturbations are generated by adding Gaussian noise whose per-location variance is scaled inversely to the residual norm of the normal feature and proportionally to the noise norm, so that anomalies are placed just outside each normal region rather than overlapped with it or far away. The paper reports that with this design a single model achieves 98.3% / 98.0% image/pixel AUROC on MVTec AD, outperforming both specialized multi-class methods and the SimpleNet baseline by 11+ points.

Load-bearing premise

The method assumes that each category's normal appearance is concentrated around one representative center, so that every normal patch's distance to that center is small; a class with two clearly distinct normal appearances would push one normal mode far from the center and cause it to be flagged as defective.

Editorial extensions

If this is right

  • On MVTec AD, image-level and pixel-level AUROC improve by about 11 points over the SimpleNet baseline, and CRAS also reports the best multi-class results on VisA and MPDD.
  • A single model covers 15 MVTec categories with 6.7 GB memory and about 77 FPS, versus 10+ GB and comparable or lower speed for several per-class competitors.
  • The global-to-local matching cuts per-patch center retrieval from $O(|K|)$ to $O(1)$, so inference time does not grow with the number of categories.
  • On the MPDD dataset, which contains parts photographed from varied angles, CRAS still reports the strongest multi-class numbers, indicating the residual representation tolerates moderate intra-class pose variation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The single-center assumption could be relaxed by assigning each class multiple centers, for example one per visual mode; CRAS's global-to-local matching would extend naturally to this, but the paper does not explore it.
  • Because the residual $u - p$ is computed with a frozen backbone plus lightweight adapter, the center memory could in principle be updated with new classes without retraining the discriminator, though the paper only evaluates a fixed class set.
  • The distance-guided noise rule could be reused as a data-augmentation module inside other feature-based anomaly detectors; the paper does not claim this generality.
  • The robustness results under illumination and vibration disturbances suggest the residual representation filters out style-like variation; an explicit test would compare CRAS against a raw-feature baseline under the same perturbations while keeping the discriminator architecture fixed.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes CRAS, a unified multi-class industrial anomaly detection method built on a pretrained frozen backbone with a trainable feature adapter. It introduces four modules: MCMC (spatial-averaged contextual centers), HPI (global-to-local center matching), DAFS (distance-guided Gaussian anomaly synthesis), and CRD (center-aware residual discrimination with a binary discriminator). The manuscript reports state-of-the-art or competitive results on MVTec AD, VisA, and MPDD under a multi-class setting, as well as on a newly collected textile defect dataset, with image-level AUROC 98.3% and pixel-level AUROC 98.0% on MVTec AD.

Significance. If the reported results are reproducible and the evaluation protocol is fair, CRAS is a meaningful contribution to unified multi-class anomaly detection, offering a compact single-model solution with competitive inference speed (77 FPS) and memory footprint. The release of source code and a new industrial dataset is a positive aspect. However, the central claim of superiority over prior methods rests on a small margin over PNPT (+1.3 I-AUROC) and on hyperparameter choices that appear to be selected directly on the test set, so the significance can only be assessed after these issues are resolved.

major comments (5)
  1. [Section IV-D, Fig. 8] The parameter analyses in Section IV-D sweep individual hyperparameters and report I-AUROC/P-AUROC on the MVTec AD test set, and the configuration listed in Section IV-A2 (WRN50, p=3, levels 2+3, sigma=0.015, beta=0.3) matches the peaks of the corresponding curves. No held-out validation split is described for these choices. Because the main empirical claim depends on these test numbers, the authors must either demonstrate that the hyperparameters were selected on a validation split or re-evaluate with a fixed default configuration and report test performance. In addition, since all results appear to be single-run values, mean and standard deviation over multiple seeds should be reported for the main tables and ablations.
  2. [Section III-D, Eq. (5)] Equation (5) defines the distance ratio α using the quantities g'_i and r'_i, which are only referred to as 'norm matrices.' The formula is not reproducible without knowing whether these are per-position values, per-image means, per-class statistics, or global constants, and how they are computed. Please define these terms explicitly and describe the computation. Relatedly, clarify whether the Gaussian noise vector g_i consists of independent components each with variance σ^2, and how the Gaussian norm g^{h,w}_i is obtained.
  3. [Section III-B2 and Section III-E] The feature adapter A_θ is described as trainable in Section III-B1 and as being trained together with the discriminator in Section III-E, but Section III-B2 states that the contextual centers are computed using a 'frozen feature adapter.' The paper does not describe any center update after adapter training. If the centers are computed with the initial random adapter and then kept fixed while the adapter is trained, the test features and the centers live in different feature spaces, which undermines the residual matching in Section III-F. Please clarify whether the centers are updated during or after training, or justify why fixed initial centers remain a valid reference.
  4. [Section IV-C, Table V] The ablation removing MCMC (variant B) shows an I-AUROC drop from 95.0 to 80.2 on MPDD, while the corresponding drops on MVTec AD and VisA are 0.7 and 2.8 points. The text attributes this to replacing the averaged center with a single sample, but such a strong dataset-specific sensitivity is surprising and should be supported by repeated runs or a statistical analysis. Please report variance across seeds or explain why MPDD is so sensitive to this component.
  5. [Section IV-A2 and Section IV-B] The paper states that all compared methods were re-trained and tested under a unified multi-class anomaly detection setting, but it does not specify how the hyperparameters of each baseline were chosen or adapted for the multi-class setting. If the baselines were run with default single-class hyperparameters, the comparison may be biased in favor of CRAS. Please provide the exact training configurations for all baselines, or a reference to the benchmark protocol used, so that the fairness of Table I (and Tables III–IV) can be assessed.
minor comments (5)
  1. [Section III-C.1] The complexity claim 'reducing the time complexity for a single sample from O(|K|) to O(1)' is imprecise: the global matching step still requires an argmax over K centers, so the complexity is O(|K| + HWP) for the local alignment. Please rephrase or clarify.
  2. [Section IV-C, Fig. 8(e)] The text says β=0.3 is optimal but also that β=0.5 lags by -0.1% in I-AUROC and leads by +0.2% in P-AUROC. This is confusing; please clarify how the 'optimal' choice is defined when the two metrics disagree.
  3. [Section III-C.2] The recomposed center p_i is obtained by selecting the most similar patch from the matched center rather than by an averaging operation; the term 'recomposed center' might be misleading. Consider adding a sentence clarifying that it is a nearest-patch selection.
  4. [Section II-C] Reference [33] (a point cloud primitive segmentation paper) does not directly support the statement that raw feature distributions vary significantly across categories and lack clear boundaries; please cite a more relevant reference or clarify the connection.
  5. [Section IV-E] The ITDD dataset is summarized only by the number of training/test images. Please provide additional information such as defect types, annotation procedure, and image resolution so that the public dataset claim can be verified.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: CRAS's pipeline is self-contained end-to-end; its own prior works (GLASS, PBAS) appear only as baselines, and the reported gains are empirical, not derived from a premise that already contains the result.

full rationale

CRAS's derivation chain is self-contained. The contextual centers in Eq. 1 are spatial averages of normal features from a frozen extractor; the recomposed centers in Eqs. 2-3 are retrieved by cosine matching; residuals are computed as u - p; synthetic anomalies are v = u + alpha * g from Eqs. 4-6; the discriminator is trained with binary cross-entropy in Eq. 8; and the inference score in Eqs. 9-10 is the discriminator's confidence on [u | u - p]. No quantity is defined in terms of the reported AUROC, and there is no equation-level reduction (Eq. X = Eq. Y by construction). The method is an empirical architecture benchmarked against external baselines, which is the normal non-circular case. The authors' own prior works, GLASS [17] and PBAS [30], appear only as comparison baselines and as prior art for Gaussian feature synthesis, and that premise is also supported by external works (SimpleNet [16], UniAD [31]), so the self-citations are not load-bearing. No uniqueness theorem is imported from the authors. The main caveat is evaluation integrity rather than circularity: Section IV-D and Fig. 8 sweep the pretrained backbone, neighborhood size p, hierarchy levels, Gaussian variance sigma, and adjustment magnitude beta directly on the MVTec AD test set with no described held-out validation split, and the configuration in Section IV-A.2 (WRN50, p=3, levels 2+3, sigma=0.015, beta=0.3) matches the optima in Fig. 8. The headline 98.3/98.0 I-AUROC/P-AUROC may therefore be optimistic due to selection on the evaluation set, and the +1.3 margin over PNPT could shrink under a proper validation protocol. This does not make the derivation circular, because the AUROC values are measured outcomes rather than constructed quantities, and the VisA, MPDD, and ITDD results use the same fixed hyperparameters, giving partially independent support. The paper also states its own key assumption, 'the initial center cannot adequately handle samples with large intra-class variance' (Section III-C), which is consistent with the largest ablation drops on MPDD's dispersed parts; this is an acknowledged limitation rather than hidden circularity. Score 2 reflects the minor non-load-bearing self-citations and the flagged test-set selection risk, not a circular derivation.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The method rests on several untested domain assumptions: transferred ImageNet features are informative, residual features are more consistent across classes, and per-class centers are sufficient. It adds no new physical entities. Free hyperparameters sigma, beta, p, and feature levels are tuned on MVTec AD, and the norm matrices g'/r' in Eq. 5 are left undefined.

free parameters (6)
  • Gaussian noise variance sigma = 0.015
    Set via parameter analysis on MVTec AD (Fig. 8d); no separate validation split is described.
  • Variance adjustment magnitude beta = 0.3
    Tuned on MVTec AD (Fig. 8e); the ablation shows minor sensitivity.
  • Neighborhood patch size p = 3
    Chosen as optimal in Fig. 8b for feature aggregation in the MCMC module.
  • Hierarchy levels used = 2 and 3
    Selected in Fig. 8c as the best concatenation of feature levels.
  • Learning rates for adapter and discriminator = 1e-4 and 2e-4
    Reported in Sec IV-A.2 as standard optimizer choices; not analyzed for sensitivity.
  • Training epochs and batch size = 100 epochs, batch 32
    Chosen per ADer survey recommendation and computational constraints in Sec IV-A.2.
assumptions (4)
  • domain assumption Pretrained ImageNet features, after a linear adapter, retain enough information to detect industrial anomalies across classes.
    Section III-B freezes WideResNet50 and relies on transfer; no analysis of domain shift is provided.
  • ad hoc to paper Residual features are more consistent across categories than raw features.
    Section I states this as the key insight but does not prove or measure it on normal-only data; the t-SNE evidence comes after training with synthetic anomalies.
  • domain assumption Each category's normal patch distribution is well summarized by a single contextual center found by cosine matching.
    Section III-C, Eq. 2-3; the paper itself notes the center cannot handle large intra-class variance, and MPDD ablation shows larger drops.
  • ad hoc to paper Distance-guided Gaussian noise, with variance inversely proportional to residual norm, approximates the real anomaly distribution.
    Section III-D introduces this heuristic; its validity is only indirectly tested through final AUROC scores.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Center-aware Residual Anomaly Synthesis for Multi-class Industrial Anomaly Detection." pith.science (2026). https://pith.science/paper/OJUWIPTI

@misc{pith2026250517551,
  author       = {Pith},
  title        = {Pith review of: Center-aware Residual Anomaly Synthesis for Multi-class Industrial Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OJUWIPTI}},
  note         = {Machine review of arXiv:2505.17551}
}
read the original abstract

Anomaly detection plays a vital role in the inspection of industrial images. Most existing methods require separate models for each category, resulting in multiplied deployment costs. This highlights the challenge of developing a unified model for multi-class anomaly detection. However, the significant increase in inter-class interference leads to severe missed detections. Furthermore, the intra-class overlap between normal and abnormal samples, particularly in synthesis-based methods, cannot be ignored and may lead to over-detection. To tackle these issues, we propose a novel Center-aware Residual Anomaly Synthesis (CRAS) method for multi-class anomaly detection. CRAS leverages center-aware residual learning to couple samples from different categories into a unified center, mitigating the effects of inter-class interference. To further reduce intra-class overlap, CRAS introduces distance-guided anomaly synthesis that adaptively adjusts noise variance based on normal data distribution. Experimental results on diverse datasets and real-world industrial applications demonstrate the superior detection accuracy and competitive inference speed of CRAS. The source code and the newly constructed dataset are publicly available at https://github.com/cqylunlun/CRAS.

Figures

Figures reproduced from arXiv: 2505.17551 by the authors.

Figure 1
Figure 1. Conceptual illustration of Industrial Anomaly Detection (IAD) settings and our motivation. (a) The single-class setting trains separate models for each category independently. (b) The multi-class setting trains a unified model for all known categories. (c) Our method aims to address inter￾class interference and intra-class overlap in the multi-class setting through center-aware residual learning and distance-guided … view at source ↗
Figure 2
Figure 2. Schematic of the proposed CRAS. (a) Multi-class Contextual Memory Center (MCMC) extracts features and initializes the multi-class centers. (b) Hierarchical Pattern Integration (HPI) aligns normal features with the contextual centers. (c) Distance-guided Anomaly Feature Synthesis (DAFS) synthesizes anomaly features based on Gaussian noise. (d) Center-aware Residual Discrimination (CRD) enhances discriminative ability… view at source ↗
Figure 3
Figure 3. Overview of the DAFS module. (a) Pipeline of anomaly synthesis based on recomposed center and Gaussian noise. (b) Anomaly feature distribution synthesized by the distance-guided mechanism. (c) Residual feature distribution after center-aware residual learning. an initialization method based on spatial position averaging. This method uses a pair of frozen feature extractor Eϕ and frozen feature adapter Aθ to average … view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: t-SNE visualization of adapted patch-level feature distributions. (a) Raw features show significant overlap between normal and abnormal samples. (b) Center-aware residual features generated by CRD form more compact normal clusters and better separation from anomalies. …
Figure 5
Figure 5. Figure 5: Performance comparison of various methods on MPDD under the single-class setting, as measured by I-AUROC% and P-AUROC%. insight for imbalanced positive and negative samples. B. Comparative Experiments The proposed CRAS is compared with several typical and state-of-the-…
Figure 6
Figure 6. Figure 6: Qualitative comparison of CRAS with recent SOTA methods (OneNIP [32] and PBAS [30]) across different categories of several datasets. TABLE III PERFORMANCE COMPARISON OF VARIOUS METHODS ON VISA UNDER MULTI-CLASS SETTING, AS MEASURED BY I-AUROC%, I-AP%, P-AUROC%, AND P-A…
Figure 7
Figure 7. Figure 7: t-SNE visualization of adapted sample-level feature distributions on the MVTec AD dataset. All normal and abnormal samples are accurately matched to their category-specific global centers. C. Ablation Studies In this section, ablation experiments were conducted to veri…
Figure 8
Figure 8. Figure 8: Quantitative results of various parameters on MVTec AD measured by I-AUROC% and P-AUROC%. (a) Dependence on pretrained model. (b) Selection of neighborhood size. (c) Concatenation of hierarchy levels. (d) Variance of Gaussian noise. (e) Magnitude of variance adjustment…
Figure 9
Figure 9. Figure 9: Experimental verification of CRAS on real-world application. (a) Automatic optical inspection equipment. (b) Industrial Textile Defect Detection (ITDD) dataset. (c) Quantitative results. (d) Qualitative results. 6) Robustness under industrial disturbances: To evaluate …

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. INP-Former++: Advancing Universal Anomaly Detection via Intrinsic Normal Prototypes and Residual Learning

    cs.CV 2025-06 conditional novelty 6.0 of 10

    INP-Former++ detects image defects by extracting intrinsic normal prototypes from the test image itself and reconstructing only normal regions, achieving state-of-the-art results across single-class, multi-class, few-...

Reference graph

Works this paper leans on

41 extracted references · 30 canonical work pages · cited by 1 Pith paper

  1. [1]

    An anomaly feature-editing- based adversarial network for texture defect visual inspection,

    H. Yang, Q. Zhou, K. Song, and Z. Yin, “An anomaly feature-editing- based adversarial network for texture defect visual inspection,” IEEE Transactions on Industrial Informatics , vol. 17, no. 3, pp. 2220–2230, 2020

    work page 2020

  2. [2]

    Collaborative discrepancy opti- mization for reliable image anomaly localization,

    Y . Cao, X. Xu, Z. Liu, and W. Shen, “Collaborative discrepancy opti- mization for reliable image anomaly localization,” IEEE Transactions on Industrial Informatics , vol. 19, no. 11, pp. 10 674–10 683, 2023

    work page 2023

  3. [3]

    Prior normality prompt transformer for multiclass industrial image anomaly detection,

    H. Yao, Y . Cao, W. Luo, W. Zhang, W. Yu, and W. Shen, “Prior normality prompt transformer for multiclass industrial image anomaly detection,” IEEE Transactions on Industrial Informatics, vol. 20, no. 10, pp. 11 866– 11 876, 2024

    work page 2024

  4. [4]

    Memorizing normality to detect anomaly: Memory- augmented deep autoencoder for unsupervised anomaly detection,

    D. Gong, L. Liu, V . Le, B. Saha, M. R. Mansour, S. Venkatesh, and A. v. d. Hengel, “Memorizing normality to detect anomaly: Memory- augmented deep autoencoder for unsupervised anomaly detection,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 1705–1714

    2019

  5. [5]

    Reconstruction by inpainting for visual anomaly detection,

    V . Zavrtanik, M. Kristan, and D. Sko ˇcaj, “Reconstruction by inpainting for visual anomaly detection,” Pattern Recognition, vol. 112, p. 107706, 2021

    2021

  6. [6]

    Anomaly detection via reverse distillation from one-class embedding,

    H. Deng and X. Li, “Anomaly detection via reverse distillation from one-class embedding,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 9737–9746

    2022

  7. [7]

    Varad: Lightweight high- resolution image anomaly detection via visual autoregressive modeling,

    Y . Cao, H. Yao, W. Luo, and W. Shen, “Varad: Lightweight high- resolution image anomaly detection via visual autoregressive modeling,” IEEE Transactions on Industrial Informatics , vol. 21, no. 4, pp. 3246– 3255, 2025

    work page 2025

  8. [8]

    Exploring intrinsic normal prototypes within a single image for universal anomaly detection,

    W. Luo, Y . Cao, H. Yao, X. Zhang, J. Lou, Y . Cheng, W. Shen, and W. Yu, “Exploring intrinsic normal prototypes within a single image for universal anomaly detection,” arXiv preprint arXiv:2503.02424 , 2025

    arXiv 2025

Show all 41 references
  1. [9]

    Multiresolution knowledge distillation for anomaly detection,

    M. Salehi, N. Sadjadi, S. Baselizadeh, M. H. Rohban, and H. R. Rabiee, “Multiresolution knowledge distillation for anomaly detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 14 902–14 912

    2021

  2. [10]

    Towards total recall in industrial anomaly detection,

    K. Roth, L. Pemula, J. Zepeda, B. Sch ¨olkopf, T. Brox, and P. Gehler, “Towards total recall in industrial anomaly detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 14 318–14 328

    2022

  3. [11]

    Cfa: Coupled-hypersphere-based fea- ture adaptation for target-oriented anomaly localization,

    S. Lee, S. Lee, and B. C. Song, “Cfa: Coupled-hypersphere-based fea- ture adaptation for target-oriented anomaly localization,” IEEE Access, vol. 10, pp. 78 446–78 454, 2022

    2022

  4. [12]

    Pyramidflow: High-resolution defect contrastive localization using pyramid normalizing flow,

    J. Lei, X. Hu, Y . Wang, and D. Liu, “Pyramidflow: High-resolution defect contrastive localization using pyramid normalizing flow,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 14 143–14 152

    2023

  5. [13]

    Draem-a discriminatively trained reconstruction embedding for surface anomaly detection,

    V . Zavrtanik, M. Kristan, and D. Sko ˇcaj, “Draem-a discriminatively trained reconstruction embedding for surface anomaly detection,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 8330–8339

    2021

  6. [14]

    Masked swin transformer unet for industrial anomaly detection,

    J. Jiang, J. Zhu, M. Bilal, Y . Cui, N. Kumar, R. Dou, F. Su, and X. Xu, “Masked swin transformer unet for industrial anomaly detection,” IEEE Transactions on Industrial Informatics , vol. 19, no. 2, pp. 2200–2209, 2022

    2022

  7. [15]

    Memseg: A semi-supervised method for image surface defect detection using differences and commonalities,

    M. Yang, P. Wu, and H. Feng, “Memseg: A semi-supervised method for image surface defect detection using differences and commonalities,” Engineering Applications of Artificial Intelligence , vol. 119, p. 105835, 2023

    2023

  8. [16]

    Simplenet: A simple network for image anomaly detection and localization,

    Z. Liu, Y . Zhou, Y . Xu, and Z. Wang, “Simplenet: A simple network for image anomaly detection and localization,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 20 402–20 411

    2023

  9. [17]

    A unified anomaly synthesis strategy with gradient ascent for industrial anomaly detection and localization,

    Q. Chen, H. Luo, C. Lv, and Z. Zhang, “A unified anomaly synthesis strategy with gradient ascent for industrial anomaly detection and localization,” in European Conference on Computer Vision , 2024, pp. 37–54

    2024

  10. [18]

    Supersimplenet: Unifying unsu- pervised and supervised learning for fast and reliable surface defect detection,

    B. Rolih, M. Fu ˇcka, and D. Sko ˇcaj, “Supersimplenet: Unifying unsu- pervised and supervised learning for fast and reliable surface defect detection,” in International Conference on Pattern Recognition , 2025, pp. 47–65

    2025

  11. [19]

    Scalable industrial visual anomaly detection with partial semantics aggregation vision transformer,

    H. Yao, W. Luo, J. Lou, W. Yu, X. Zhang, Z. Qiang, and H. Shi, “Scalable industrial visual anomaly detection with partial semantics aggregation vision transformer,” IEEE Transactions on Instrumentation and Measurement, vol. 73, pp. 1–17, 2024

    2024

  12. [20]

    Masked au- toencoders are scalable vision learners,

    K. He, X. Chen, S. Xie, Y . Li, P. Doll ´ar, and R. Girshick, “Masked au- toencoders are scalable vision learners,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 16 000–16 009

    2022

  13. [21]

    Hierarchi- cal vector quantized transformer for multi-class unsupervised anomaly detection,

    R. Lu, Y . Wu, L. Tian, D. Wang, B. Chen, X. Liu, and R. Hu, “Hierarchi- cal vector quantized transformer for multi-class unsupervised anomaly detection,” Advances in Neural Information Processing Systems, vol. 36, pp. 8487–8500, 2023

    2023

  14. [22]

    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

    2024

  15. [23]

    Personalizing vision-language models with hybrid prompts for zero- shot anomaly detection,

    Y . Cao, X. Xu, Y . Cheng, C. Sun, Z. Du, L. Gao, and W. Shen, “Personalizing vision-language models with hybrid prompts for zero- shot anomaly detection,” IEEE Transactions on Cybernetics , vol. 55, no. 4, pp. 1917–1929, 2025

    1917

  16. [24]

    Unsupervised anomaly detection and localiza- tion with one model for all category,

    P. Tan and W. K. Wong, “Unsupervised anomaly detection and localiza- tion with one model for all category,” Knowledge-Based Systems , vol. 289, p. 111533, 2024

    2024

  17. [25]

    Hierarchical gaussian mixture normalizing flow modeling for unified anomaly detection,

    X. Yao, R. Li, Z. Qian, L. Wang, and C. Zhang, “Hierarchical gaussian mixture normalizing flow modeling for unified anomaly detection,” in European Conference on Computer Vision , 2024, pp. 92–108

    2024

  18. [26]

    Multi-confidence guided source-free domain adaption method for point cloud primitive seg- mentation,

    S. Wang, Y . Tong, X. Shang, and Z. Zhang, “Multi-confidence guided source-free domain adaption method for point cloud primitive seg- mentation,” in 2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 737–743

    2024

  19. [27]

    An incremental unified framework for small defect inspection,

    J. Tang, H. Lu, X. Xu, R. Wu, S. Hu, T. Zhang, T. W. Cheng, M. Ge, Y .-C. Chen, and F. Tsung, “An incremental unified framework for small defect inspection,” in European Conference on Computer Vision , 2024, pp. 307–324

    2024

  20. [28]

    Deep one-class classifi- cation via interpolated gaussian descriptor,

    Y . Chen, Y . Tian, G. Pang, and G. Carneiro, “Deep one-class classifi- cation via interpolated gaussian descriptor,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 36, no. 1, 2022, pp. 383–392

    2022

  21. [29]

    Dsr-a dual subspace re- projection network for surface anomaly detection,

    V . Zavrtanik, M. Kristan, and D. Sko ˇcaj, “Dsr-a dual subspace re- projection network for surface anomaly detection,” in European Con- ference on Computer Vision , 2022, pp. 539–554

    2022

  22. [30]

    Progressive bound- ary guided anomaly synthesis for industrial anomaly detection,

    Q. Chen, H. Luo, H. Gao, C. Lv, and Z. Zhang, “Progressive bound- ary guided anomaly synthesis for industrial anomaly detection,” IEEE CHEN et al.: CRAS FOR MULTI-CLASS INDUSTRIAL ANOMAL Y DETECTION 11 Transactions on Circuits and Systems for Video Technology , vol. 35, no. 2,...

    2025

  23. [31]

    A unified model for multi-class anomaly detection,

    Z. You, L. Cui, Y . Shen, K. Yang, X. Lu, Y . Zheng, and X. Le, “A unified model for multi-class anomaly detection,” Advances in Neural Information Processing Systems , vol. 35, pp. 4571–4584, 2022

    2022

  24. [32]

    Learning to detect multi-class anomalies with just one normal image prompt,

    B.-B. Gao, “Learning to detect multi-class anomalies with just one normal image prompt,” in European Conference on Computer Vision , 2024, pp. 454–470

    2024

  25. [33]

    Probabilistic boundary-guided point cloud primitive segmentation network,

    S. Wang, F. Qin, Y . Tong, X. Shang, and Z. Zhang, “Probabilistic boundary-guided point cloud primitive segmentation network,” IEEE Transactions on Instrumentation and Measurement , vol. 72, pp. 1–13, 2023

    2023

  26. [34]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in IEEE Conference on Computer Vision and Pattern Recognition , 2009, pp. 248–255

    2009

  27. [35]

    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 Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 9592–9600

    2019

  28. [36]

    Spot-the- difference self-supervised pre-training for anomaly detection and seg- mentation,

    Y . Zou, J. Jeong, L. Pemula, D. Zhang, and O. Dabeer, “Spot-the- difference self-supervised pre-training for anomaly detection and seg- mentation,” in European Conference on Computer Vision , 2022, pp. 392–408

    2022

  29. [37]

    Deep learning-based defect detection of metal parts: evaluating current meth- ods in complex conditions,

    S. Jezek, M. Jonak, R. Burget, P. Dvorak, and M. Skotak, “Deep learning-based defect detection of metal parts: evaluating current meth- ods in complex conditions,” in 2021 13th International Congress on Ultra Modern Telecommunications and Control Systems and Workshops, 2021, pp. 66–71

    2021

  30. [38]

    Ader: A comprehensive benchmark for multi-class visual anomaly detection,

    J. Zhang, H. He, Z. Gan, Q. He, Y . Cai, Z. Xue, Y . Wang, C. Wang, L. Xie, and Y . Liu, “Ader: A comprehensive benchmark for multi-class visual anomaly detection,” arXiv preprint arXiv:2406.03262 , 2024

    2024 arXiv

  31. [39]

    Wide residual networks,

    S. Zagoruyko and N. Komodakis, “Wide residual networks,” in British Machine Vision Conference, 2016

    2016

  32. [40]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 770–778. Qiyu Chen received the B.E. degree in Com- munication Engineering from Tongji University, S...

    2016

  33. [2023]

    degree with the Department of Precision Instrument, Tsinghua University, Beijing, China

    He is pursuing a Ph.D. degree with the Department of Precision Instrument, Tsinghua University, Beijing, China. His research interests include deep learning, anomaly detection, and machine vision. Zhen Qu received the B.Sc. degree from Xi- dian University, Xi’an, China, in 202...

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.