Pith. sign in

REVIEW 3 major objections 5 minor 49 references

SwinAD: Multi-stage feature reconstruction for unsupervised industrial anomaly detection

T0 review · 3 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read A reconstruction model with two complementary decoder branches produces the sharpest pixel-level anomaly maps reported on three multi-class industrial benchmarks, beating a higher-resolution baseline at lower compute.

desk verdict A solid, incremental MUAD engineering paper with honest ablations and consistent pixel-level gains, but the dual-branch 'manifold compatibility' score is actually a norm-weighted average that could be sensitive to arbitrary branch magnitudes, and there is no code to verify the numbers. read the letter →

arxiv 2607.14534 v1 pith:IYJNFHEK submitted 2026-07-16 cs.CV

Huong Ninh , Chien Thai , Mai Xuan Trang , Vu-Minh Le , Thanh Ha Le , Long Tran This is my paper
classification cs.CV
keywords IndustrialanomalydetectionMulti-classunsupervisedReconstruction-basedFeaturediversitySwinTransformerV2Multi-scaleaggregationlocalizationPixel-levelevaluation
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

SwinAD is a reconstruction-based model for multi-class unsupervised anomaly detection, a setting where one model must spot defects across many product categories using only normal images for training. The central claim is that a frozen hierarchical Swin Transformer V2 encoder feeding two complementary decoder branches, with a loss that keeps the branches apart, yields sharper pixel-level anomaly maps than a single-branch reconstruction. On MVTec AD, VisA, and Real-IAD, the framework reports the highest pixel-level Average Precision and F1 among the compared methods at 256×256 input resolution, beating the strongest baseline even though that baseline runs at 392×392. If the claim holds, precise defect localization across many classes is achievable with one efficient model, which matters for industrial inspection pipelines.

What carries the argument

Feature diversity-preserving dual reconstruction: two lightweight decoder branches each reconstruct the frozen encoder's multi-scale features, while a cosine-based divergence loss max(0, cos(D1, D2) - τ_s)^2 with stage-dependent margins and a ramp weight λ_div = min(λ·t/1000, λ) keeps the branches complementary. The anomaly score is 1 - cos(F⊕F, D1⊕D2), turning point-wise similarity into a manifold-compatibility check in the doubled embedding space; scale-wise maps are resized, weighted, averaged, and Gaussian-smoothed into the final anomaly map.

What would settle it

Train the same architecture with the diversity loss disabled (λ=0) but keep two branches and concatenated scoring; if pixel AP on MVTec AD stays at 74.4, the divergence term is not the active ingredient. A complementary check is to record the cosine similarity between D1 and D2 on held-out normal images at convergence: if most tokens sit at or below the margins τ_s, the constraint is never binding and the reported gains must come from extra capacity or multi-stage weighting instead of diversity.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that deterministic single-prototype reconstruction is the bottleneck in feature-space anomaly detection: when a decoder learns one projection for all normal local structures, it collapses the normal manifold and becomes either over-tolerant (reconstructing anomalies) or over-sensitive (firing on normal variations). SwinAD avoids this by maintaining two manifold-consistent reconstruction hypotheses per location, scoring a pixel by the cosine distance between the encoder feature (concatenated with itself) and the concatenation of the two reconstructed features, and aggregating this across four Swin stages. The paper reports a consistent +4.8 to +6.6 p

Load-bearing premise

The diversity mechanism depends on the assumption that pushing two decoder branches apart with hand-set per-stage margins and a ramped weight produces genuinely complementary reconstructions of the normal manifold, rather than two marginally different copies or a constraint that is never binding.

Editorial extensions

If this is right

  • One frozen encoder plus two lightweight decoder branches can localize anomalies across 15–30 categories with a single model, removing per-category training cost.
  • At 256×256, SwinAD reports higher pixel AP/F1 than a 392×392 baseline while using roughly a fifth of the FLOPs, indicating that resolution is not the only lever for localization quality.
  • The ablation shows intermediate stages (2 and 3) carry most of the localization signal, and that combining all stages beats any single stage, so multi-scale aggregation is load-bearing.
  • The dual-branch gain is consistent across all three datasets and largest on Real-IAD, matching the claim that diversity helps most when normal patterns vary widely.

Reading between the lines

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

  • A natural transfer test would attach the same two-branch, mutually-repelled decoder to a frozen convolutional encoder; if the pixel-AP gain disappears, the effect is tied to Swin/attention features rather than to manifold diversity per se.
  • The fixed scale weights (0.75/0.75/1.25/0.25) and the per-stage margins τ_s are set empirically without an analysis of the cosine distribution between branches, so an adaptive schedule for these hyperparameters is a direct extension the paper leaves open.
  • The reported P-AUROC is slightly below the strongest baseline while AP/F1 are higher, implying the method trades broad pixel ranking for spatial precision; applications that threshold pixel scores downstream should test operating-point sensitivity.
  • Because only two hypotheses are used, a stress test with a third branch and a larger margin would show whether gains saturate quickly, which the paper's own capacity argument predicts.
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

3 major / 5 minor

Summary. The paper proposes SwinAD, a multi-class unsupervised anomaly detection method built on a frozen Swin Transformer V2 encoder and a dual-branch reconstruction decoder. Multi-scale encoder features are aligned with stage-wise bottleneck modules; two decoder branches are trained with cosine reconstruction losses and a cosine diversity loss; anomaly maps are formed by concatenating encoder and decoder features per scale and aggregating across stages. Experiments on MVTec AD, VisA, and Real-IAD report competitive image-level AUROC/AP/F1 and the highest pixel-level AP/F1 among the compared reconstruction-based multi-class methods, with P-AP gains of +5.1/+4.8/+6.6 over Dinomaly at 392x392. Ablations indicate intermediate encoder stages are most informative and dual-decoder aggregation improves over a single decoder.

Significance. The empirical contribution is potentially useful: the paper demonstrates consistent, if modest, pixel-level AP/F1 gains (+1.4/+2.4/+1.5 over the single-decoder ablation in Table 5) from two reconstruction branches, and reports a large efficiency advantage over Dinomaly (54.32 vs 268.92 GFLOPs at 392x392, Table 4). The evaluation is standard and uses held-out test labels; the authors disclose that baseline numbers come from original publications and reproduce Dinomaly at 224x224. The main novelty, however, is framed as a 'manifold-compatible' diversity mechanism, and this framing is not supported by the implemented anomaly score, which is a norm-weighted average of two per-branch cosine similarities. The missing ablation of the diversity loss and its hyperparameters means the causal role of the proposed mechanism is not established. These issues are addressable with additional experiments and a revised conceptual framing.

major comments (3)
  1. [§3.4/§3.6, anomaly-map equation] The anomaly map is A_ij = 1 - cos(F⊕F, D1⊕D2). Expanding the cosine gives (||D1|| c1 + ||D2|| c2) / (sqrt(2) * sqrt(||D1||^2 + ||D2||^2)), where c1,c2 are the per-branch cosine similarities. Thus the score is a norm-weighted average of two point-wise similarities, not a 'manifold compatibility' measure. Because all training losses are scale-invariant in D1,D2, branch magnitudes are unconstrained; a branch with larger norm can dominate the score. The paper reports no normalization of D1/D2 and no statistics of their magnitudes. This is a load-bearing mismatch between the central claim and the implemented score; please either revise the conceptual claim, add feature normalization, or provide evidence that the norm ratio is stable across classes and stages.
  2. [§3.5, Table 5] The paper credits the 'feature diversity-preserving reconstruction' mechanism for the improvements, but the ablation only compares a single decoder with the full dual-decoder system. There is no ablation that removes the diversity loss L_div, varies tau_s, varies lambda_div, or changes the number of branches. The observed gains could therefore be due to an ensemble effect of two independent decoders or to the hard-mining reconstruction loss, rather than to the proposed diversity regularization. Since the diversity mechanism is the paper's main novelty, this missing analysis is needed to support the causal claim. Please report ablation without L_div and the evolution of cos(D1,D2) by stage.
  3. [§4.2, Tables 1-3] Tables 1-3 compare SwinAD's pixel-level AP/F1 against Dinomaly at 392x392 (e.g., 74.4 vs 69.3, 58.0 vs 53.2, 49.4 vs 42.8) while SwinAD runs at 256x256. The authors state that baseline numbers are directly adopted from original publications and reproduce Dinomaly only at 224x224. Since P-F1 is threshold-dependent and input resolution changes the operating point, the claimed SOTA gains are not based on a controlled comparison. To support the headline claim, please reproduce Dinomaly (and ideally the other 256x256 baselines) at 256x256 or 392x392 under the same evaluation protocol, and report the threshold selection rule for all methods.
minor comments (5)
  1. [Abstract / Table 3 caption] The abstract says 'pixel-level AP and 1' where 'F1' is intended; the Table 3 caption spells 'RealiAD' without the hyphen.
  2. [§3.6] The text states that the aggregation weights satisfy sum_i w_i = N, but the implementation reports w_i in {0.75,0.75,1.25,0.25}, which sum to 3.0. Please correct either the normalization statement or the reported values.
  3. [§4.2] The 'optimal dataset-level threshold' used for F1 should be defined precisely: specify whether it is chosen on the test set and confirm that the same protocol is applied to every compared method.
  4. [§4.4] The ~20 FPS inference speed is reported without measurement methodology (number of runs, batch size, GPU clock locking). Please state the protocol or soften the claim.
  5. [Reproducibility] The paper does not mention code availability. Releasing code, seeds, and details of the two-branch initialization would facilitate verification of the hard-mining and diversity-loss implementation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method is benchmark-validated with held-out test labels and the anomaly score is a defined formula, not a fitted target.

full rationale

SwinAD's anomaly map is defined in Section 3.6 as A_s,ij = 1 - cos((F_s,ij ⊕ F_s,ij), (D1_s,ij ⊕ D2_s,ij)) and aggregated across scales with fixed weights. The training losses in Section 3.5 are per-branch cosine reconstruction losses plus a cosine diversity term; no parameter is fitted to the reported pixel-level AP/F1 values. The reported numbers are computed on the held-out test labels of MVTec AD, VisA, and Real-IAD, so the predictions are not equivalent by construction to any training target. The bottleneck and decoder blocks are directly adopted from Dinomaly [13], but [13] is an external baseline with no author overlap, and the paper additionally reproduces Dinomaly at 224×224, so this is architecture reuse rather than a self-citation chain. The concatenated-cosine score does algebraically reduce to a norm-weighted average of the two single-branch cosine similarities, so the 'manifold compatibility' label overstates what the formula implements; however, the formula is the method's operational definition, not a derived first-principles result whose conclusion is its own input. Manual hyperparameters such as w_i = {0.75,0.75,1.25,0.25}, τ_s = {0.4,0.8,0.95,0.99}, top-k = 10%, and the Gaussian smoothing kernel are empirical tuning choices; this is a benchmark-comparison and generalization risk, not circularity, because the evaluations are held out and the metrics are not structurally forced to match these settings.

Assumptions & free parameters 6 free parameters · 3 assumptions · 1 invented entities

The method rests on a frozen pretrained backbone (standard for the field), standard reconstruction assumptions, and several hand-set hyperparameters that are not varied or analyzed. The only invented conceptual entity is the two-branch 'manifold hypothesis' framing, whose independent support is limited to the paper's own ablations.

free parameters (6)
  • Scale aggregation weights w_i = {0.75, 0.75, 1.25, 0.25}
    Set empirically across the three datasets; the paper does not report a sensitivity analysis.
  • Diversity margins tau_s = {0.4, 0.8, 0.95, 0.99}
    Stage-dependent cosine margins for the diversity loss, given as a hand-set schedule.
  • Diversity ramp lambda = not given (lambda*t/1000 capped at lambda)
    The ramp weight for the diversity loss; the final lambda value is not reported.
  • Top-k hard-mining ratio = 10% of spatial tokens
    Chosen without ablations reported for other values.
  • Gaussian smoothing kernel/scale = 9x9, sigma=2
    Applied to the final anomaly map; no ablation or sensitivity check.
  • Number of decoder branches = 2
    Chosen as 'a practical balance between feature diversity and optimization stability'; no ablation with 3+ branches is shown.
assumptions (3)
  • domain assumption Frozen ImageNet-pretrained Swin Transformer V2 features generalize to industrial anomaly detection.
    The whole method depends on the pretrained features being semantically aligned with industrial normal/defect structure; this is consistent with prior work (UniAD, Dinomaly) but is not derived or tested.
  • domain assumption Normal features are reconstructible by a learnable decoder with dropout, while anomalous features are not.
    The standard reconstruction assumption stated in Section 3.1 ('normal features are reconstructible, while anomalous patterns typically result in higher reconstruction errors').
  • ad hoc to paper Two reconstruction hypotheses can cover the multi-modal local normal manifold, and concatenating them turns point-wise similarity into 'manifold compatibility'.
    The core theoretical framing of Section 3.4; the paper provides intuition but no formal argument for why concatenation captures manifold membership.
invented entities (1)
  • Two parallel reconstruction hypotheses D^1 and D^2
    purpose: To preserve feature diversity and avoid deterministic feature collapse in reconstruction-based anomaly detection.
    These are internal decoder branches; their only evidence is the reported benchmark improvements in the same paper. No ablation quantifies whether the gains come from the diversity loss vs the extra decoder parameters.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SwinAD: Multi-stage feature reconstruction for unsupervised industrial anomaly detection." pith.science (2026). https://pith.science/paper/IYJNFHEK

@misc{pith2026260714534,
  author       = {Pith},
  title        = {Pith review of: SwinAD: Multi-stage feature reconstruction for unsupervised industrial anomaly detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IYJNFHEK}},
  note         = {Machine review of arXiv:2607.14534}
}
read the original abstract

Industrial anomaly detection aims to identify and localize defective regions without relying on exhaustive annotations of all possible defect types. Although recent unsupervised methods have achieved strong performance, most are primarily designed for single-class settings and often struggle in multi-class scenarios, where diverse normal patterns may lead to over-generalization and reduce the discriminative capability between normal and anomalous regions. In this paper, we propose SwinAD, a reconstruction-based framework for multi-class unsupervised anomaly detection that leverages a frozen pretrained Swin Transformer V2 encoder and a feature diversity-preserving reconstruction decoder. The hierarchical encoder provides semantically rich multi-scale features, while stage-wise bottleneck modules with dropout prevent trivial identity mapping and encourage robust reconstruction of normal patterns. To further improve localization, we introduce a feature diversity-preserving reconstruction framework that maintains complementary reconstruction hypotheses instead of relying on a single decoding branch. The discrepancies between encoder features and the two reconstructed features are then aggregated across multiple scales to produce the final anomaly map. Experiments conducted on three industrial anomaly detection benchmarks, including MVTec AD, VisA, and Real-IAD, demonstrate that SwinAD achieves competitive image-level performance and strong pixel-level localization accuracy, with particularly notable improvements in pixel-level AP and 1 on MVTec AD. These results indicate that combining hierarchical Swin features with diverse multi-scale reconstruction substantially improve pixel-level localization in multi-class unsupervised anomaly setting.

Figures

Figures reproduced from arXiv: 2607.14534 by the authors.

Figure 1
Figure 1. Examples from the MVTec AD (top), VisA (middle), and Real-IAD (bottom) datasets with SwinAD segmentation results overlaid. Orange contours indicate anomaly boundaries, while the blue-to-orange heatmaps visualize anomaly intensity for defects such as broken glass, scratches, and structural changes. are trained exclusively on normal samples, and anomalies are identified as deviations from learned normal patterns. Exis… view at source ↗
Figure 2
Figure 2. Illustration of the proposed SwinAD framework. A frozen pretrained Swin Transformer V2 first extracts hierarchical multi-scale features from the input image. These features are then aligned by stage-wise bottleneck modules with dropout and passed to a feature diversity-preserving reconstruction decoder, which produces complementary reconstructions at different scales. The differences between the encoder and reconstr… view at source ↗
Figure 3
Figure 3. Overview of the anomaly map estimation process of SwinAD. For each stage, the encoder feature is compared with the two reconstructed features from the dual decoder using cosine distance. The scale-wise anomaly maps are first resized to a unified resolution and then aggregated into a single map. where 𝑡 is the training iteration. This progressive scheduling strategy stabilizes optimization during early training stage… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison of anomaly localization on MVTecAD samples. From top to bottom: input images with annotated defect regions, zoomed-in views of anomalies, anomaly maps (AMs) produced by our proposed SwinAD and Dinomaly, corresponding binary masks after thresholdi…
Figure 5
Figure 5. Figure 5: Qualitative comparison of anomaly localization on VisA samples. The rows show input images with annotated defect regions, zoomed-in anomaly regions, anomaly maps produced by SwinAD and Dinomaly, thresholded binary masks, and ground-truth masks. SwinAD produces more loc…
Figure 6
Figure 6. Figure 6: Qualitative comparison of anomaly localization on Real-IAD samples. From top to bottom, the figure presents input images, enlarged defect regions, anomaly maps from SwinAD and Dinomaly, the corresponding predicted masks, and ground-truth masks. Compared with Dinomaly, …
Figure 7
Figure 7. Figure 7: Visualization of the stage-wise ablation study. From left to right: Input image, anomaly map computed from individual stages, final anomaly map, and ground-truth anomaly mask. The figure compares anomaly responses obtained from individual Swin encoder stages and corres…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 7 linked inside Pith

  1. [1]

    Efficientad: Accurate visual anomaly detection at millisecond-level latencies

    Kilian Batzner, Lars Heckler, and Rebecca König. Efficientad: Accurate visual anomaly detection at millisecond-level latencies. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 128–138, 2024

    2024

  2. [2]

    Improving unsupervised defect segmentation by applying structural similarity to autoencoders.arXiv preprint arXiv:1807.02011, 2018

    Paul Bergmann, Sindy Löwe, Michael Fauser, David Sattlegger, and Carsten Steger. Improving unsupervised defect segmentation by applying structural similarity to autoencoders.arXiv preprint arXiv:1807.02011, 2018

    arXiv 2018

  3. [3]

    Mvtec ad–a comprehensive real-world dataset for unsupervised anomaly detection

    Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Ste- ger. Mvtec ad–a comprehensive real-world dataset for unsupervised anomaly detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592–9600, 2019

    2019

  4. [4]

    Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings

    Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Ste- ger. Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4183– 4192, 2020

    2020

  5. [5]

    InProceedingsoftheIEEE/CVF internationalconferenceoncomputervision,pages9650–9660,2021

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal,PiotrBojanowski,andArmandJoulin.Emergingpropertiesin self-supervisedvisiontransformers. InProceedingsoftheIEEE/CVF internationalconferenceoncomputervision,pages9650–9660,2021

    2021

  6. [6]

    Detecting anomalous structures by convolutional sparse models

    Diego Carrera, Giacomo Boracchi, Alessandro Foi, and Brendt Wohlberg. Detecting anomalous structures by convolutional sparse models. In2015 International Joint Conference on Neural Networks (IJCNN), pages 1–8. IEEE, 2015

    2015

  7. [7]

    Padim: a patch distribution modeling framework for anomaly detection and localization

    Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: a patch distribution modeling framework for anomaly detection and localization. InInternational conference on pattern recognition, pages 475–489. Springer, 2021

    2021

  8. [8]

    Anomaly detection via reverse distil- lation from one-class embedding

    Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distil- lation from one-class embedding. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9737– 9746, 2022

    2022

Show all 49 references
  1. [9]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weis- senborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arX...

    2010 arXiv

  2. [10]

    Few-shot defect image generation via defect-aware feature manipulation

    Yuxuan Duan, Yan Hong, Li Niu, and Liqing Zhang. Few-shot defect image generation via defect-aware feature manipulation. In Proceedings of the AAAI conference on artificial intelligence, pages 571–578, 2023

    2023

  3. [11]

    Fastrecon: Few-shot industrial anomaly detection via fast feature reconstruction

    Zheng Fang, Xiaoyang Wang, Haocheng Li, Jiejie Liu, Qiugui Hu, and Jimin Xiao. Fastrecon: Few-shot industrial anomaly detection via fast feature reconstruction. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 17481–17490, 2023

    2023

  4. [12]

    Recon- trast: Domain-specific anomaly detection via contrastive reconstruc- tion.AdvancesinNeuralInformationProcessingSystems,36:10721– 10740, 2023

    Jia Guo, Shuai Lu, Lize Jia, Weihang Zhang, and Huiqi Li. Recon- trast: Domain-specific anomaly detection via contrastive reconstruc- tion.AdvancesinNeuralInformationProcessingSystems,36:10721– 10740, 2023

    2023

  5. [13]

    Dinomaly: The less is more philosophy in multi-class unsuper- visedanomalydetection

    JiaGuo,ShuaiLu,WeihangZhang,FangChen,HuiqiLi,andHongen Liao. Dinomaly: The less is more philosophy in multi-class unsuper- visedanomalydetection. InProceedingsoftheComputerVisionand Pattern Recognition Conference, pages 20405–20415, 2025

    2025

  6. [14]

    Con- trolling neural collapse enhances out-of-distribution detection and transfer learning.arXiv preprint arXiv:2502.10691, 2025

    Md Yousuf Harun, Jhair Gallardo, and Christopher Kanan. Con- trolling neural collapse enhances out-of-distribution detection and transfer learning.arXiv preprint arXiv:2502.10691, 2025

    2025

  7. [15]

    Mambaad: Exploring state space models for multi-class unsupervised anomaly detection.Advances in Neural Information Processing Systems, 37:71162–71187, 2024

    Haoyang He, Yuhu Bai, Jiangning Zhang, Qingdong He, Hongxu Chen, Zhenye Gan, Chengjie Wang, Xiangtai Li, Guanzhong Tian, and Lei Xie. Mambaad: Exploring state space models for multi-class unsupervised anomaly detection.Advances in Neural Information Processing Systems, 37:7116...

    2024

  8. [16]

    Adiffusion- based framework for multi-class anomaly detection

    Haoyang He, Jiangning Zhang, Hongxu Chen, Xuhai Chen, Zhishan Li,XuChen,YabiaoWang,ChengjieWang,andLeiXie. Adiffusion- based framework for multi-class anomaly detection. InProceedings of the AAAI conference on artificial intelligence, pages 8472–8480, 2024

    2024

  9. [17]

    Vlmdiff:Leveragingvision-languagemodelsformulti-classanomaly detection with diffusion, 2025

    Samet Hicsonmez, Abd El Rahman Shabayek, and Djamila Aouada. Vlmdiff:Leveragingvision-languagemodelsformulti-classanomaly detection with diffusion, 2025

    2025

  10. [18]

    Registrationbasedfew-shotanomaly detection

    Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling,andYan-FengWang. Registrationbasedfew-shotanomaly detection. InEuropean conference on computer vision, pages 303–

  11. [19]

    Adversarial discriminative attention for robust anomaly detection

    Daiki Kimura, Subhajit Chaudhury, Minori Narita, Asim Munawar, and Ryuki Tachibana. Adversarial discriminative attention for robust anomaly detection. InProceedings of the IEEE/CVF winter confer- ence on applications of computer vision, pages 2172–2181, 2020

    2020

  12. [20]

    Cutpaste: Self-supervised learning for anomaly detection and local- ization

    Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and local- ization. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9664–9674, 2021

    2021

  13. [21]

    Ipg-frn: Intrinsic prototype-guided feature reconstruction network for industrial anomaly detection.Expert Systems with Ap- plications, page 132147, 2026

    Lanxiao Li, Chuanxu Wang, Yishuo Liu, Binghui Wang, and Da Song. Ipg-frn: Intrinsic prototype-guided feature reconstruction network for industrial anomaly detection.Expert Systems with Ap- plications, page 132147, 2026

    2026

  14. [22]

    Swintransformerv2: Scaling up capacity and resolution

    ZeLiu,HanHu,YutongLin,ZhuliangYao,ZhendaXie,YixuanWei, JiaNing,YueCao,ZhengZhang,LiDong,etal. Swintransformerv2: Scaling up capacity and resolution. InProceedings of the IEEE/CVF conferenceoncomputervisionandpatternrecognition,pages12009– 12019, 2022

    2022

  15. [23]

    Sim- plenet: A simple network for image anomaly detection and localiza- tion

    Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Sim- plenet: A simple network for image anomaly detection and localiza- tion. InProceedingsoftheIEEE/CVFconferenceoncomputervision and pattern recognition, pages 20402–20411, 2023

    2023

  16. [24]

    Decoupled weight decay regular- ization.arXiv preprint arXiv:1711.05101, 2017

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regular- ization.arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

  17. [25]

    Patch distance based auto-encoder for industrial anomaly detection.Expert Systems with Applications, 270:126537, 2025

    Zeqi Ma, Jiaxing Li, and Wai Keung Wong. Patch distance based auto-encoder for industrial anomaly detection.Expert Systems with Applications, 270:126537, 2025

    2025

  18. [26]

    Ocgan: One- classnoveltydetectionusingganswithconstrainedlatentrepresenta- tions.InProceedingsoftheIEEE/CVFconferenceoncomputervision and pattern recognition, pages 2898–2906, 2019

    Pramuditha Perera, Ramesh Nallapati, and Bing Xiang. Ocgan: One- classnoveltydetectionusingganswithconstrainedlatentrepresenta- tions.InProceedingsoftheIEEE/CVFconferenceoncomputervision and pattern recognition, pages 2898–2906, 2019

    2019

  19. [27]

    Towards total recall in industrial anomaly detection

    Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318–14328, 2022

    2022

  20. [28]

    Asymmetric student-teacher networks for industrial anomaly detection

    Marco Rudolph, Tom Wehrbein, Bodo Rosenhahn, and Bastian Wandt. Asymmetric student-teacher networks for industrial anomaly detection. InProceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2592–2602, 2023

    2023

  21. [29]

    Imagenet large scale visual recognition challenge.Internationaljournalofcomputervision,115(3):211–252, 2015

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh,SeanMa,ZhihengHuang,AndrejKarpathy,AdityaKhosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge.Internationaljournalofcomputervision,115(3):211–252, 2015

    2015

  22. [30]

    Multiresolutionknowledge distillation for anomaly detection

    Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mo- hammadHRohban,andHamidRRabiee. Multiresolutionknowledge distillation for anomaly detection. InProceedings of the IEEE/CVF conferenceoncomputervisionandpatternrecognition,pages14902– 14912, 2021. H. Ninh, C. Thai, M...

    2021

  23. [31]

    Natural synthetic anomalies for self-supervised anomaly detection andlocalization

    HannahMSchlüter,JeremyTan,BenjaminHou,andBernhardKainz. Natural synthetic anomalies for self-supervised anomaly detection andlocalization. InEuropeanConferenceonComputerVision,pages 474–489. Springer, 2022

    2022

  24. [32]

    Real-iad:Areal-worldmulti-viewdatasetforbenchmarkingver- satile industrial anomaly detection

    Chengjie Wang, Wenbing Zhu, Bin-Bin Gao, Zhenye Gan, Jiangning Zhang, Zhihao Gu, Shuguang Qian, Mingang Chen, and Lizhuang Ma. Real-iad:Areal-worldmulti-viewdatasetforbenchmarkingver- satile industrial anomaly detection. InProceedings of the IEEE/CVF Conference on Computer Vis...

    2024

  25. [33]

    Student- teacher feature pyramid matching for anomaly detection

    Guodong Wang, Shumin Han, Errui Ding, and Di Huang. Student- teacher feature pyramid matching for anomaly detection. InThe British Machine Vision Conference (BMVC), 2021

    2021

  26. [34]

    Student- teacher feature pyramid matching for anomaly detection.arXiv preprint arXiv:2103.04257, 2021

    Guodong Wang, Shumin Han, Errui Ding, and Di Huang. Student- teacher feature pyramid matching for anomaly detection.arXiv preprint arXiv:2103.04257, 2021

    2021 arXiv

  27. [35]

    Uninet: A contrastive learning-guidedunifiedframeworkwithfeatureselectionforanomaly detection

    Shun Wei, Jielin Jiang, and Xiaolong Xu. Uninet: A contrastive learning-guidedunifiedframeworkwithfeatureselectionforanomaly detection. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 9994–10003, 2025

    2025

  28. [36]

    Dfr: Deep feature recon- struction for unsupervised anomaly segmentation.arXiv preprint arXiv:2012.07122, 2020

    Jie Yang, Yong Shi, and Zhiquan Qi. Dfr: Deep feature recon- struction for unsupervised anomaly segmentation.arXiv preprint arXiv:2012.07122, 2020

    2012 arXiv

  29. [37]

    Lafite: Latent diffusion model with featureeditingforunsupervisedmulti-classanomalydetection.arXiv preprint arXiv:2307.08059, 2023

    Haonan Yin, Guanlong Jiao, Qianhui Wu, Borje F Karlsson, Biqing Huang, and Chin Yew Lin. Lafite: Latent diffusion model with featureeditingforunsupervisedmulti-classanomalydetection.arXiv preprint arXiv:2307.08059, 2023

    2023 arXiv

  30. [38]

    A unified model for multi-class anomaly detection

    Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. Advances in Neural Information Processing Systems, 35:4571–4584, 2022

    2022

  31. [39]

    Draem - a dis- criminatively trained reconstruction embedding for surface anomaly detection

    Vitjan Zavrtanik, Matej Kristan, and Danijel Skocaj. Draem - a dis- criminatively trained reconstruction embedding for surface anomaly detection. InProceedingsoftheIEEE/CVFInternationalConference on Computer Vision (ICCV), pages 8330–8339, 2021

    2021

  32. [40]

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

    Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Dsr–a dual subspace re-projection network for surface anomaly detection. In European conference on computer vision, pages 539–554. Springer, 2022

    2022

  33. [41]

    A diverse embedding-based composite reconstruction encoder– decoder for color fabric defect detection.Expert Systems with Applications, 278:127261, 2025

    HongweiZhang,LiPingMeng,ShuaiLu,ZhihuanSong,andLvyuan Wu. A diverse embedding-based composite reconstruction encoder– decoder for color fabric defect detection.Expert Systems with Applications, 278:127261, 2025

    2025

  34. [42]

    Ader: A comprehensive benchmark for multi-class visual anomaly detection.arXiv preprint arXiv:2406.03262, 1(4), 2024

    Jiangning Zhang, Haoyang He, Zhenye Gan, Qingdong He, Yux- uan Cai, Zhucun Xue, Yabiao Wang, Chengjie Wang, Lei Xie, and Yong Liu. Ader: A comprehensive benchmark for multi-class visual anomaly detection.arXiv preprint arXiv:2406.03262, 1(4), 2024

    2024 arXiv

  35. [43]

    Exploring plain vit features for multi-class unsupervised visual anomaly detec- tion.ComputerVisionandImageUnderstanding,253:104308,2025

    Jiangning Zhang, Xuhai Chen, Yabiao Wang, Chengjie Wang, Yong Liu, Xiangtai Li, Ming-Hsuan Yang, and Dacheng Tao. Exploring plain vit features for multi-class unsupervised visual anomaly detec- tion.ComputerVisionandImageUnderstanding,253:104308,2025

    2025

  36. [44]

    Adaptive frequency modulated transformer for industrialsurfacedefectdetection.ExpertSystemswithApplications, page 132502, 2026

    Lei Zhang, Xiaoqian Zhang, Zejun Huang, Shijie Li, Huan Yuan, and Wenlong Qiu. Adaptive frequency modulated transformer for industrialsurfacedefectdetection.ExpertSystemswithApplications, page 132502, 2026

    2026

  37. [45]

    Destseg: Segmentation guided denoising student-teacher for anomaly detection

    Xuan Zhang, Shiyu Li, Xi Li, Ping Huang, Jiulong Shan, and Ting Chen. Destseg: Segmentation guided denoising student-teacher for anomaly detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3914–3923, 2023

    2023

  38. [46]

    Omnial: A unified cnn framework for unsupervised anomaly localization

    Ying Zhao. Omnial: A unified cnn framework for unsupervised anomaly localization. InProceedings of the IEEE/CVF conference oncomputervisionandpatternrecognition,pages3924–3933,2023

    2023

  39. [47]

    Fad:Featureaugmenteddistillationforanomaly detection and localization.Expert Systems with Applications, 288: 128249, 2025

    Qiyin Zhong, Xianglin Qiu, Xinqiao Zhao, Xiaowei Huang, Gang Liu,andJiminXiao. Fad:Featureaugmenteddistillationforanomaly detection and localization.Expert Systems with Applications, 288: 128249, 2025

    2025

  40. [48]

    Class- incremental learning via dual augmentation.Advances in neural information processing systems, 34:14306–14318, 2021

    Fei Zhu, Zhen Cheng, Xu-Yao Zhang, and Cheng-lin Liu. Class- incremental learning via dual augmentation.Advances in neural information processing systems, 34:14306–14318, 2021

    2021

  41. [49]

    Spot-the-difference self-supervised pre-training for anomaly detection and segmentation

    Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pre-training for anomaly detection and segmentation. InEuropean conference on computer vision, pages 392–408. Springer, 2022. H. Ninh, C. Thai, M.X. Trang, V. Le, T.H....

    2022

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.