Pith. sign in

REVIEW 4 major objections 4 minor 53 references

Advancing Metallic Surface Defect Detection via Anomaly-Guided Pretraining on a Large Industrial Dataset

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Anomaly-guided pretraining of both backbone and detector on 120,000 unlabeled industrial images consistently improves metallic surface defect detection over ImageNet and COCO initialization.

desk verdict Solid anomaly-guided pretraining study with a real dataset, but the undisclosed prompt source and overclaimed COCO comparison need fixing. read the letter →

arxiv 2509.18919 v2 pith:4GX2JJXN submitted 2025-09-23 cs.CV

Chuni Liu , Hongjie Li , Jiaqi Du , Yangyang Hou , Qian Sun , Lei Jin , Ke Xu This is my paper
classification cs.CV
keywords metallicsurfacedefectdetectionself-supervisedpretraininganomaly-guidedanomalymapdistillationpseudo-defectboxesindustrialdatasetfew-shotobjectCLIP-based
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

Metallic surface defects are rare, subtle, and visually close to background texture, so pretraining on ImageNet leaves a domain gap, and naive self-supervised pretraining on in-domain data does not help—the authors find it often hurts. This paper tries to establish that the missing ingredient is explicit anomaly guidance: pixel-wise heatmaps that highlight defects. It proposes AGSSP, a two-stage pretraining method that first distills anomaly maps into the backbone and then pretrains the detector on pseudo-defect boxes derived from the same maps. On a 120,000-image industrial pretraining corpus and small downstream datasets, the resulting weights beat ImageNet- and COCO-initialized models across backbones and detectors, with the largest gains under extreme data scarcity.

What carries the argument

The central object is the anomaly map: a pixel-wise score heatmap indicating where defects are likely. It carries the whole argument. KeAD generates the maps by comparing CLIP patch features against normal and anomalous text prompts, using detailed defect descriptions instead of generic words and a V-V attention variant that preserves local spatial information. The maps then do double duty: the backbone stage distills them into high-level feature maps (L2 loss within ViTs, cosine loss across architectures), and the detector stage converts them into pseudo-defect boxes through a category-specific dynamic threshold and connected-component extraction. This single signal unifies feature learning

What would settle it

Run AGSSP with KeAD prompts constructed only from the pretraining corpus—no representative images from Casting Billet, Steel Pipe, or GC10—and finetune on those datasets. If the large gains in Tables 2 and 3 mostly disappear, the effect depends on target-informed prompt construction; if the gains persist, the anomaly-guidance prior is genuinely general.

Watch

Extended reading notes

Core claim

The paper's central claim is that anomaly maps can act as a pretraining supervision signal that aligns representation learning with the defect-detection task better than either natural-image pretraining or naive in-domain self-supervision. To produce these maps at scale, the authors introduce KeAD, a training-free extension of the CLIP-based WinCLIP method that uses defect-specific text descriptions and a local-attention variant, and they collect 120,000 unlabeled industrial images across 61 categories. In the first pretraining stage, the backbone is trained so its high-level attention maps match the anomaly maps through a distillation loss; in the second, pseudo-defect boxes extracted from

Load-bearing premise

The load-bearing premise is that the anomaly maps used for supervision are a transferable prior and not a way of leaking target-dataset knowledge: for the downstream categories absent from the pretraining data, the text prompts are built from representative images, and if those come from the target datasets themselves, the observed gains could reflect the teacher already knowing those defect types.

Editorial extensions

If this is right

  • AGSSP-pretrained weights can replace ImageNet or COCO initialization for metallic surface defect detectors, improving accuracy without changing inference cost.
  • The benefit is concentrated in low-data regimes: on a 50-image training split, gains reach +41% mAP@0.5 and +24.5% mAP@0.5:0.95 over ImageNet/SimMIM baselines.
  • The method transfers across backbone architectures (CSPDarknet, ResNet50, Swin-Base) and detectors (YOLOv8, Faster R-CNN, DINO), suggesting the anomaly-guidance principle is not tied to one architecture.
  • It also improves a non-metallic fabric defect dataset, indicating some generality beyond metal surfaces.
  • Anomaly guidance can be added on top of existing pretraining objectives such as classification, contrastive learning, and masked image modeling, rather than replacing them.

Reading between the lines

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

  • If the anomaly prior is the true driver, substituting KeAD with any strong zero- or few-shot anomaly detector should preserve most of the gain; this is directly testable.
  • The authors' own bad-case analysis flags noise-mimicking textures and extremely subtle defects as failure modes, which points to a next step: adding a noise-invariance or texture-robustness term to the distillation loss.
  • Because the pseudo-boxes are derived from dense maps, the same pipeline could plausibly pretrain segmentation models by converting maps to pseudo-masks, not just boxes.
  • The prompt-construction step uses representative images for the downstream categories, so an important open check is whether the gains survive when prompts are built without any access to the validation datasets; that check separates a general pretraining prior from target-specific knowledge.
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

4 major / 4 minor

Summary. The paper proposes AGSSP (Anomaly-Guided Self-Supervised Pretraining), a two-stage pretraining framework for metallic surface defect detection. In Stage 1 (AGBP), anomaly maps generated by a CLIP-based teacher (KeAD) are distilled into the backbone's high-level feature maps via an attention-map loss, optionally combined with existing SSL objectives. In Stage 2 (AGDP), the same anomaly maps are binarized with a dynamic threshold and converted to pseudo-defect boxes, which supervise detector pretraining. The authors collect a 120,000-image, 61-category industrial dataset, introduce two new pixel-level labeled defect datasets (Casting Billet, Steel Pipe), and evaluate across three backbones, multiple pretraining methods, and several downstream datasets including GC10-DET and a fabric dataset. They report consistent gains over ImageNet/COCO baselines, up to 10% mAP@0.5 and 11.4% mAP@0.5:0.95, with particularly large gains in extreme few-shot settings.

Significance. If the reported gains are attributable to the proposed anomaly-guidance mechanism rather than to target-informed prompt construction, this is a significant contribution to industrial defect detection: it provides a general paradigm for in-domain pretraining, a large public pretraining corpus, and two new labeled datasets, with code and models to be released. The experimental scope is extensive—three backbones, three pretraining tasks, multiple detectors, and six evaluation settings—and the paper explicitly compares against ImageNet and COCO initialization. The main scientific risk is leakage via the text-prompt construction: the teacher's prompts are built from representative images whose provenance is not disclosed, which could confound the claimed transferable anomaly prior. This issue is load-bearing for the central claim and must be addressed with additional disclosure and ablation.

major comments (4)
  1. [Sec. 4.1.1, Sec. 3.1] KeAD's text prompts are constructed from 'one representative image' per object category, described by GPT-4o and expert knowledge. The paper does not state whether the validation categories (Casting Billet, Steel Pipe, GC10) are among the 61 pretraining categories, nor where the representative images for those categories were taken from. If these representative images come from the validation sets themselves, then the anomaly maps used in AGBP/AGDP are implicitly tuned to downstream defect types, and the reported gains (up to 10% mAP@0.5, 11.4% mAP@0.5:0.95) could reflect target-informed prompt engineering rather than a general, transferable anomaly prior. This is the main load-bearing assumption of the paper. Please disclose the source of all representative images and add a control experiment using generic prompts (e.g., WinCLIP's 'damaged/defected') or prompts constructed from categori
  2. [Sec. 5.1.2, Sec. 1] KeAD's few-shot mode uses 'four normal samples' as reference images, yet the large-scale pretraining dataset is described as unlabeled, with 'no labels distinguishing normal from defective samples' (Sec. 1). The paper does not explain how these four normal samples are selected. If they are chosen by human inspection, this is a form of supervision that should be reported; if they are selected at random, some may contain defects, which could corrupt the anomaly maps used as pseudo-labels. Please specify the selection protocol and evaluate the sensitivity of the results to the number and quality of normal reference samples.
  3. [Sec. 5.2.1, Table 3, Sec. 1] The extreme few-shot results on Casting Billet-Mini50 (gains of +20 to +41 mAP@0.5 over ImageNet and COCO baselines) are presented as evidence of the method's effectiveness. However, if the KeAD prompts already encode the target defect categories through representative images from the target domain (see first major comment), these large gains may be due to the teacher already knowing the downstream defect types, rather than to a sample-efficient anomaly-guidance mechanism. To separate these explanations, the authors should compare AGSSP pretraining with prompts built from the target categories against prompts built from unrelated categories; if the latter also produces large gains, the transferability claim is supported. This ablation is missing.
  4. [Sec. 4.2, Eq. (9)-(11)] The pseudo-box generation pipeline involves several hyperparameters: the threshold offset δ=0.1, the top-10 box cutoff, and NMS. The paper only ablates δ with a mask-IoU metric (Fig. 7), not with respect to the final detection performance of AGDP. The sensitivity of the two-stage pretraining to the number of pseudo-boxes and NMS thresholds should be documented, since these choices directly affect the supervisory signal in the detector pretraining stage.
minor comments (4)
  1. [Sec. 5.2.1, Table 2] The table formatting is inconsistent (e.g., 'GC10-Det' vs 'GC10-DET', missing delimiters in some rows). Also, clarify what 'Detection COCO' means for each backbone: is it the COCO-pretrained version of the same CSPDarknet/ResNet/Swin architecture, and are the COCO weights obtained from the original papers or MMDetection?
  2. [Sec. 3.2] The training/validation split for Casting Billet and Steel Pipe is described as '1:1 ratio.' Please specify whether the split is random, whether it is stratified by defect type, and the number of defect instances in each split. The unusually large validation fraction affects the reliability of the reported metrics.
  3. [Sec. 4.1.1, Eq. (1)] The temperature τ in the scoring function is not specified. Report the value used for the CLIP text-image similarity, as it affects the anomaly map calibration and the 0.5 image-level threshold in Sec. 4.2.
  4. [Sec. 4.1.2, Eq. (8)] The hyperparameter λ is set from the ratio of loss values at the first iteration. Specify whether this is computed on the first batch of each training run, and whether λ is held fixed throughout pretraining or adjusted in later iterations.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: AGSSP is an empirical distillation method whose supervision signal is not the downstream labels by construction.

full rationale

AGSSP's derivation chain is: KeAD (a training-free adaptation of WinCLIP with an external CLIP encoder) produces anomaly maps from 120k unlabeled industrial images; AGBP aligns student high-level attention maps to these maps via Eq. (7); AGDP derives pseudo-boxes from the same maps via Eqs. (9)-(11) and pretrains the detector. At no point is the downstream mAP encoded in the pretraining objective. The teacher (CLIP) and the attention-transfer technique ([49]) are external, and the validation images are stated to be disjoint from the pretraining set (Sec. 3.1). The only arguable weakness is that KeAD's text prompts are built from 'one representative image' plus expert knowledge (Sec. 4.1.1), whose source is not disclosed. If that image came from an evaluation domain, it would be a leakage confound, not a mathematical reduction of the reported gain to a fitted parameter. No self-citation is load-bearing; the two self-citations in Sec. 2.3 are related-work context. Hence no circular step can be quoted or exhibited.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper's central claim rests on the availability of a reliable anomaly-prior teacher. The teacher is an external CLIP model, but its prompts are created using domain knowledge and a representative defect image, which is a form of target-domain supervision. The hyperparameters (delta, epochs, layers, box count) are tuned on the target datasets, but they are not the main scientific claim. No new physical or theoretical entities are introduced.

free parameters (6)
  • delta = 0.1
    Threshold offset in Eq. 9; set by ablation study (Fig. 7) to balance mask accuracy across datasets.
  • backbone_pretraining_epochs = 200
    Ablation in Fig. 6(b) shows 200 epochs optimal on Casting Billet.
  • distillation_layers = last two layers
    Table 5 shows distilling into last two layers outperforms only the last layer.
  • frozen_lower_layers = first two layers
    Freezing lower layers during pretraining and finetuning is based on experiments in Fig. 6(a) and Table 5.
  • lambda_loss_weight = ratio of initial losses
    Set as L0_distill/L0_task at first iteration; a heuristic that is not constant.
  • pseudo_box_count = top 10
    Top 10 connected components kept as pseudo-boxes, a design choice not ablated.
assumptions (5)
  • domain assumption CLIP's zero-shot image-text alignment generalizes to industrial surface images
    The entire KeAD teacher relies on CLIP embeddings for anomaly scoring; if CLIP does not distinguish defects from background textures in metallic images, the maps are noise. Section 4.1.1.
  • domain assumption Hand-crafted defect text prompts are a transferable prior
    Prompts from GPT-4o and expert knowledge are assumed to describe defect features across the dataset, but they are constructed per category and may encode target-specific information. Section 4.1.1.
  • domain assumption Anomaly maps are a good proxy for defect locations
    The whole pretraining distills the teacher's predicted anomaly regions into the student; if the teacher's maps are noisy or biased, the student learns those biases. Sections 4.1 and 4.2.
  • domain assumption Low-level features are domain-agnostic and can be kept frozen from ImageNet
    The method initializes lower layers from ImageNet and only distills into high-level layers, based on observation rather than proof. Section 4.1.2.
  • domain assumption The 120k industrial dataset is representative of downstream defect distributions
    No overlap with validation sets is asserted, but the diversity across 61 categories is assumed to cover the defect types in Casting Billet and Steel Pipe. Section 3.1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Advancing Metallic Surface Defect Detection via Anomaly-Guided Pretraining on a Large Industrial Dataset." pith.science (2026). https://pith.science/paper/4GX2JJXN

@misc{pith2026250918919,
  author       = {Pith},
  title        = {Pith review of: Advancing Metallic Surface Defect Detection via Anomaly-Guided Pretraining on a Large Industrial Dataset},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4GX2JJXN}},
  note         = {Machine review of arXiv:2509.18919}
}
read the original abstract

The pretraining-finetuning paradigm is a crucial strategy in metallic surface defect detection for mitigating the challenges posed by data scarcity. However, its implementation presents a critical dilemma. Pretraining on natural image datasets such as ImageNet, faces a significant domain gap. Meanwhile, naive self-supervised pretraining on in-domain industrial data is often ineffective due to the inability of existing learning objectives to distinguish subtle defect patterns from complex background noise and textures. To resolve this, we introduce Anomaly-Guided Self-Supervised Pretraining (AGSSP), a novel paradigm that explicitly guides representation learning through anomaly priors. AGSSP employs a two-stage framework: (1) it first pretrains the model's backbone by distilling knowledge from anomaly maps, encouraging the network to capture defect-salient features; (2) it then pretrains the detector using pseudo-defect boxes derived from these maps, aligning it with localization tasks. To enable this, we develop a knowledge-enhanced method to generate high-quality anomaly maps and collect a large-scale industrial dataset of 120,000 images. Additionally, we present two small-scale, pixel-level labeled metallic surface defect datasets for validation. Extensive experiments demonstrate that AGSSP consistently enhances performance across various settings, achieving up to a 10\% improvement in mAP@0.5 and 11.4\% in mAP@0.5:0.95 compared to ImageNet-based models. All code, pretrained models, and datasets are publicly available at https://clovermini.github.io/AGSSP-Dev/.

Figures

Figures reproduced from arXiv: 2509.18919 by the authors.

Figure 1
Figure 1. Motivation of our method. (a) Most previous pretrained models rely [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Examples of defect images with annotations in Casting Billet and [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Overview of the proposed Anomaly-Guided Self-Supervised Pretraining Framework. This framework is composed of two key phases: anomaly map guided [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Comparative analysis of ImageNet vs. AGSSP pretraining: Good and [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Visualization of anomaly maps: Comparison between our proposed [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: (a) Effects of freezing layers 1–4 during finetuning. (b) Impact of different backbone pretraining epochs. (c) mAP and training time comparison: freezing vs. not freezing the backbone during detector pretraining. (d) Perfor￾mance comparison: Cosine Distance vs. L2 loss…
Figure 7
Figure 7. Figure 7: An ablation study of δ parameter choices. Model mAP@.5 mAP@.5:95 Param GPU Load FPS YOLOv8s 75.1 43.9 11.1M 1851MiB 94.5 YOLOv11s 74.7 45.2 36.0M 531MiB 111.4 YOLOv12s 77.2 47.1 35.3M 535Mib 70.5 RT-DETR 73.1 45.4 125.2M 2797Mib 26.5 SLF-YOLO 72.6 38.6 10.2M 368Mib 78.…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 1 canonical work pages

  1. [1]

    X. Lv, F. Duan, J.-j. Jiang, X. Fu, L. Gan, Deep metallic surface defect detection: The new benchmark and detec- tion network, Sensors 20 (6) (2020) 1562

    2020

  2. [2]

    Ameri, C.-C

    R. Ameri, C.-C. Hsu, S. S. Band, A systematic review of deep learning approaches for surface defect detection in industrial applications, Engineering Applications of Arti- ficial Intelligence 130 (2024) 107717. 11

    2024

  3. [3]

    Saberironaghi, J

    A. Saberironaghi, J. Ren, M. El-Gindy, Defect detection methods for industrial products using deep learning tech- niques: A review, Algorithms 16 (2) (2023) 95

    2023

  4. [4]

    C.-L. Li, K. Sohn, J. Yoon, T. Pfister, Cutpaste: Self- supervised learning for anomaly detection and localiza- tion, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 9664– 9674

    2021

  5. [5]

    Y . Tai, K. Yang, T. Peng, Z. Huang, Z. Zhang, Defect im- age sample generation with diffusion prior for steel sur- face defect recognition, arXiv preprint arXiv:2405.01872 (2024)

    arXiv 2024

  6. [6]

    J. Liu, F. Guo, H. Gao, M. Li, Y . Zhang, H. Zhou, Defect detection of injection molding products on small datasets using transfer learning, Journal of manufacturing pro- cesses 70 (2021) 400–413

    2021

  7. [7]

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, L. Fei- Fei, Imagenet: A large-scale hierarchical image database, in: 2009 IEEE conference on computer vision and pattern recognition, Ieee, 2009, pp. 248–255

    2009

  8. [8]

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, C. L. Zitnick, Microsoft coco: Common objects in context, in: Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, Springer, 2014, pp. 740–755

    2014

Show all 53 references
  1. [9]

    X. Chen, S. Xie, K. He, An empirical study of training self-supervised vision transformers, in: Proceedings of the IEEE/CVF international conference on computer vi- sion, 2021, pp. 9640–9649

    2021

  2. [10]

    Z. Xie, Z. Zhang, Y . Cao, Y . Lin, J. Bao, Z. Yao, Q. Dai, H. Hu, Simmim: A simple framework for masked im- age modeling, in: Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition, 2022, pp. 9653–9663

    2022

  3. [11]

    K. Tian, Y . Jiang, Q. Diao, C. Lin, L. Wang, Z. Yuan, Designing bert for convolutional networks: Sparse and hierarchical masked modeling, arXiv preprint arXiv:2301.03580 (2023)

    2023 arXiv

  4. [12]

    Oquab, T

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haz- iza, F. Massa, A. El-Nouby, et al., Dinov2: Learning robust visual features without supervision, arXiv preprint arXiv:2304.07193 (2023)

    2023 arXiv

  5. [13]

    X. Hu, J. Yang, F. Jiang, A. Hussain, K. Dashtipour, M. Gogate, Steel surface defect detection based on self-supervised contrastive representation learning with matching metric, Applied Soft Computing 145 (2023) 110578

    2023

  6. [14]

    Torpey, L

    D. Torpey, L. Pratt, R. Klein, A large-scale evaluation of pretraining paradigms for the detection of defects in elec- troluminescence solar cell images (2024)

    2024

  7. [15]

    X. Liu, X. Miao, H. Jiang, J. Chen, M. Wu, Z. Chen, Tower masking mim: A self-supervised pretraining method for power line inspection, IEEE Transactions on Industrial Informatics 20 (1) (2023) 513–523

    2023

  8. [16]

    M. Li, J. Wu, X. Wang, C. Chen, J. Qin, X. Xiao, R. Wang, M. Zheng, X. Pan, Aligndet: Aligning pre-training and fine-tuning in object detection, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 6866–6876

    2023

  9. [17]

    Huang, Y

    Z. Huang, Y . Feng, Q. Liu, Y . Wang, Mutdet: Mutu- ally optimizing pre-training for remote sensing object de- tection, in: European Conference on Computer Vision, Springer, 2025, pp. 1–17

    2025

  10. [18]

    Jeong, Y

    J. Jeong, Y . Zou, T. Kim, D. Zhang, A. Ravichandran, O. Dabeer, Winclip: Zero-/few-shot anomaly classifica- tion and segmentation, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 19606–19616

    2023

  11. [19]

    Q. Sun, K. Xu, H. Liu, Unsupervised surface defect detec- tion of aluminum sheets with combined bright-field and dark-field illumination, Optics and Lasers in Engineering 168 (2023) 107674

    2023

  12. [20]

    Y . Li, H. Wang, Y . Duan, X. Li, Clip surgery for better ex- plainability with enhancement in open-vocabulary tasks, arXiv preprint arXiv:2304.05653 (2023)

    2023 arXiv

  13. [21]

    S. Ren, K. He, R. Girshick, J. Sun, Faster r-cnn: Towards real-time object detection with region proposal networks, IEEE transactions on pattern analysis and machine intelli- gence 39 (6) (2016) 1137–1149

    2016

  14. [22]

    Jocher, A

    G. Jocher, A. Chaurasia, J. Qiu, Ultralytics yolov8 (2023). URLhttps://github.com/ultralytics/ ultralytics

    2023

  15. [23]

    Jocher, J

    G. Jocher, J. Qiu, Ultralytics yolo11 (2024). URLhttps://github.com/ultralytics/ ultralytics

    2024

  16. [24]

    Y . Tian, Q. Ye, D. Doermann, Yolov12: Attention- centric real-time object detectors, arXiv preprint arXiv:2502.12524 (2025)

    2025 arXiv

  17. [25]

    Zhang, F

    H. Zhang, F. Li, S. Liu, L. Zhang, H. Su, J. Zhu, L. M. Ni, H.-Y . Shum, Dino: Detr with improved denoising an- chor boxes for end-to-end object detection, arXiv preprint arXiv:2203.03605 (2022)

    2022 arXiv

  18. [26]

    W. Lv, Y . Zhao, Q. Chang, K. Huang, G. Wang, Y . Liu, Rtdetrv2: All-in-one detection transformer beats yolo and dino (2024).arXiv:2407.17140. 12

    2024 arXiv

  19. [27]

    Y . Liu, Y . Liu, X. Guo, X. Ling, Q. Geng, Metal surface defect detection using slf-yolo enhanced yolov8 model, Scientific Reports 15 (1) (2025) 11105

    2025

  20. [28]

    H. Feng, K. Song, W. Cui, Y . Zhang, Y . Yan, Cross po- sition aggregation network for few-shot strip steel surface defect segmentation, IEEE Transactions on Instrumenta- tion and Measurement 72 (2023) 1–10

    2023

  21. [29]

    Huang, Y

    J. Huang, Y . Wu, X. Zhou, J. Lin, Z. Chen, G. Zhang, L. Xia, J. Zhang, Multi-scale adaptive prototype trans- former network for few-shot strip steel surface defect segmentation, IEEE Transactions on Instrumentation and Measurement (2025)

    2025

  22. [30]

    Y . Ma, M. Liu, Y . Zhang, X. Wang, Y . Wang, Spdp-net: A semantic prior guided defect perception network for auto- mated aero-engine blades surface visual inspection, IEEE Transactions on Automation Science and Engineering 22 (2024) 2724–2733

    2024

  23. [31]

    Y . Fang, W. Wang, B. Xie, Q. Sun, L. Wu, X. Wang, T. Huang, X. Wang, Y . Cao, Eva: Exploring the limits of masked visual representation learning at scale, in: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 19358–19369

    2023

  24. [32]

    Caron, H

    M. Caron, H. Touvron, I. Misra, H. Jégou, J. Mairal, P. Bojanowski, A. Joulin, Emerging properties in self- supervised vision transformers, in: Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 9650–9660

    2021

  25. [33]

    Siméoni, H

    O. Siméoni, H. V . V o, M. Seitzer, F. Baldassarre, M. Oquab, C. Jose, V . Khalidov, M. Szafraniec, S. Yi, M. Ramamonjisoa, et al., Dinov3, arXiv preprint arXiv:2508.10104 (2025)

    2025 arXiv

  26. [34]

    H.-Y . Zhou, S. Yu, C. Bian, Y . Hu, K. Ma, Y . Zheng, Com- paring to learn: Surpassing imagenet pretraining on radio- graphs by comparing image representations, in: Medical Image Computing and Computer Assisted Intervention– MICCAI 2020: 23rd International Conference, Lima, P...

    2020

  27. [35]

    Stuckner, B

    J. Stuckner, B. Harder, T. M. Smith, Microstructure seg- mentation with deep learning encoders pre-trained on a large microscopy dataset, npj Computational Materials 8 (1) (2022) 200

    2022

  28. [36]

    M. He, L. Qin, Y . Wang, X. Deng, Q. Liu, Y . Zhang, K. Liu, A weakly supervised contrastive learning pretrain- ing method for visual defect detection of transmission lines, IEEE Transactions on Instrumentation and Mea- surement 74 (2025) 1–15

    2025

  29. [37]

    Bouniot, R

    Q. Bouniot, R. Audigier, A. Loesch, A. Habrard, Proposal-contrastive pretraining for object detection from fewer data, in: The Eleventh International Conference on Learning Representations, 2023

    2023

  30. [38]

    Kirillov, E

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.- Y . Lo, et al., Segment anything, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 4015–4026

    2023

  31. [39]

    K. Roth, L. Pemula, J. Zepeda, B. Schölkopf, T. Brox, P. Gehler, Towards total recall in industrial anomaly de- tection, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 14318–14328

    2022

  32. [40]

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

    2022

  33. [41]

    Q. Sun, K. Xu, D. Zhao, H. Li, L. Jin, C. Liu, P. Xu, Png: an adaptive local-global hybrid framework for unsuper- vised material surface defect detection, Expert Systems with Applications (2025) 128711

    2025

  34. [42]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al., Learning transferable visual models from natural language supervision, in: International conference on ma- chine learning, PMLR, 2021, pp. 8748–8763

    2021

  35. [43]

    X. Chen, Y . Han, J. Zhang, A zero-/few-shot anomaly classification and segmentation method for cvpr 2023 vand workshop challenge tracks 1&2: 1st place on zero- shot ad and 4th place on few-shot ad, arXiv preprint arXiv:2305.17382 (2023)

    2023 arXiv

  36. [44]

    Q. Zhou, G. Pang, Y . Tian, S. He, J. Chen, Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection, in: The Twelfth International Conference on Learning Representations, 2023

    2023

  37. [45]

    W. Ma, X. Zhang, Q. Yao, F. Tang, C. Wu, Y . Li, R. Yan, Z. Jiang, S. K. Zhou, Aa-clip: Enhancing zero- shot anomaly detection via anomaly-aware clip, in: Pro- ceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 4744–4754

    2025

  38. [46]

    URLhttps://www.kaggle.com/competitions/ severstal-steel-defect-detection/data

    Severstal: Steel defect detection, retrieved 2 May 2022 (2019). URLhttps://www.kaggle.com/competitions/ severstal-steel-defect-detection/data

    2022

  39. [47]

    S. Chen, S. Jiang, X. Wang, P. Sun, C. Hua, J. Sun, An ef- ficient detector for detecting surface defects on cold-rolled steel strips, Engineering Applications of Artificial Intelli- gence 138 (2024) 109325

    2024

  40. [48]

    Vaswani, Attention is all you need, Advances in Neural Information Processing Systems (2017)

    A. Vaswani, Attention is all you need, Advances in Neural Information Processing Systems (2017)

    2017

  41. [49]

    Zagoruyko, N

    S. Zagoruyko, N. Komodakis, Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer, arXiv preprint arXiv:1612.03928 (2016). 13

    2016 arXiv

  42. [50]

    K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: Proceedings of the IEEE con- ference on computer vision and pattern recognition, 2016, pp. 770–778

    2016

  43. [51]

    Bochkovskiy, C.-Y

    A. Bochkovskiy, C.-Y . Wang, H.-Y . M. Liao, Yolov4: Optimal speed and accuracy of object detection (2020). arXiv:2004.10934

    2020 arXiv

  44. [52]

    Bradski, Learning opencv: Computer vision with the opencv library, O’REILLY google schola 2 (2008) 334– 352

    G. Bradski, Learning opencv: Computer vision with the opencv library, O’REILLY google schola 2 (2008) 334– 352

    2008

  45. [53]

    Neamah, Fabric defects object detection dataset (2024).doi:10.6084/m9.figshare.25546465.v2

    O. Neamah, Fabric defects object detection dataset (2024).doi:10.6084/m9.figshare.25546465.v2. 14

    2024 doi

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.