CoChat
Sign Up
Cited by 38Anomaly Detection Techniques and ApplicationsIndustrial Vision Systems and Defect DetectionAdversarial Robustness in Machine Learning

Self-Supervised Learning for Anomaly Detection With Dynamic Local Augmentation

Open access
Yoa, Seungdong, Lee, Seungjun, Kim, Chiyoon, Kim, Hyunwoo J·2021·IEEE Access
DOI 10.1109/access.2021.3124525
Open this paper in CoChat — free accountOpen PDF

Abstract

Anomaly detection is an important problem for recent advances in machine learning. To this end, many attempts have emerged to detect unknown anomalies of the images by learning representations and designing score functions. In this paper, we propose a simple yet effective framework for unsupervised anomaly detection using self-supervised learning. We extend conventional self-supervised learning for an anomaly detection problem. In anomaly detection, anomalous patterns appear in the local regions of an image, so we employ dynamic local augmentation to generate a negative pair of the images from the normal training dataset. Specifically, in addition to learning the global representation of an image, our framework contrasts a normal sample to a locally augmented sample. To effectively apply the local augmentations regardless of a category or a random location of an image, we use dynamically weighted local augmentations to generate more suitable negative samples. We also present a novel scoring function for detecting unseen anomalous patterns. Our experiment demonstrates the effectiveness of our method, and we show that our framework achieves competitive performance compared to state-of-the-art methods on MVTec Anomaly Detection dataset.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper proposes a self-supervised anomaly detection framework that uses dynamic local augmentation to create pseudo-abnormal negative samples and two pixel-level scoring functions based on regression error and MC-dropout uncertainty. On MVTec AD, the method achieves competitive pixel-level anomaly localization, with an average AUROC of 93.04%.

Key findings

  • The proposed method achieves the best average performance on MVTec AD, with AUROC averages of 91.42% for texture, 93.85% for object, and 93.04% overall.
  • It outperforms the second-highest baseline by 12.66% in the texture category and by 2.53% in the object category.
  • Dynamic local augmentation (DLA) outperforms static local augmentation (LA): average AUROC 93.04% vs. 82.92%, which the paper reports as an average 9.90% improvement.
  • Ablation shows dense loss and negative pairs matter: training without local augmentation is more effective than training with static local augmentation, while DLA gives the best performance.

Methods

  • The framework is based on SimSiam-style self-supervised learning with only positive pairs, plus a dense self-supervised loss adapted from DenseCL; it adds a negative-pair loss using dynamic local augmentation.
  • Dynamic local augmentation combines hard local augmentations—Rotation, ColorJitter, Perm, and Grayscale—weighted by their MSE difference from the original image, so augmentation strength adapts to image category and location.
  • Anomaly scores are computed per pixel using regression error between test feature maps and average training feature maps, and MC-dropout variance-map differences, across three ResNet-18 stages (C2, C3, C4), upsampled to 224×224.
  • Experiments use MVTec AD (5 texture and 10 object categories, over 5000 high-resolution images resized to 256×256), ResNet-18 trained from scratch, batch size 200, SGD with momentum 0.9, weight decay 5×10⁻⁴, 200 epochs, base learning rate 0.02, local region size 50–128, λ₁=1, λ₂=0.1, λ₃=0.5, α₁=1, α₂=0.1; results are average AUROC over three runs.

Limitations

  • Stated: hyperparameter tuning is difficult in anomaly detection because the training set contains only normal samples; the authors tune using normal samples and pseudo-abnormal samples generated by local augmentations.
  • Stated as future work: extending to wide-scene anomaly detection with many objects in one image or video where uncertainty is high, and studying how the shape, size, and variety of dynamic local augmentations affect abnormal sample quality.
  • Evident from the text: evaluation is limited to MVTec AD, with models trained and tested separately per category, and comparisons are mainly against baselines from the MVTec AD paper plus U-Student.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1109/access.2021.3124525

Cites

52 of 52 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 13 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    Some hard local augmentations have no effect at all when applied at certain random locations of an image, for example rotation applied at a local region without shape produces little change.

    “Since the hard augmentation is applied at a random location, we observe that some hard local augmentations have no effect at all in certain locations as shown in the bottom row of Figure 2.”
    • Also stated in: § Local Augmentation. · plain
      “For example, if rotation is applied at a local region that is without shape (e.g., the center of metal nut and bottle classes, the smooth surface of capsule and leather classes), we observe that there is little change (e.g., the bottom row of the”
    Quote in the paperSubstantial claim
  • Findingown data

    The model trained with self-supervised learning using only global and dense loss (without dynamic local augmentation) tends to have poor anomaly detection at test time.

    “In Table 2, we observed that the model trained with self-supervised learning using only global and dense loss (the first∼second row in Table 2) tends to have poor detecting anomalies at test time.”
    Quote in the paperSubstantial claim
  • Findingown data

    The model with dynamic local augmentation (DLA) performs an average of 9.90% higher than with fixed local augmentation (LA) in Table 2.

    “The fourth row and the last row (Ours) in Table 2 show that the model with DLA (the last row) performs an average of 9.90% higher than with LA.”
    Quote in the paperSubstantial claim
  • Findingown data

    Training without local augmentation is more effective than training using static local augmentation (LA).

    “Furthermore, training without local augmentation is more effective than training using static local augmentation (LA) as shown in the second and forth row in Table 2.”
    Quote in the paperSubstantial claim
  • Findingown data

    The proposed framework achieves competitive performance compared to state-of-the-art methods on the MVTec Anomaly Detection dataset.

    “Our experiment demonstrates the effectiveness of our method, and we show that our framework achieves competitive performance compared to state-of-the-art methods on MVTec Anomaly Detection dataset.”
    • Also stated in: § VOLUME xx, 20xx · boosted
      “Lastly, our experiments demonstrate that the proposed method which is trained from scratch achieves the competitive performance on the public benchmark MVTec Anomaly Detection dataset.”
    Quote in the paperSubstantial claim
  • May combine two claims. 11
  • Classified as probably not a substantial claim. 9

  • May combine two claims.

    Findingown data

    The proposed method outperforms the second-highest method for wood by 16.4% and for zipper by 7.5%.

    “Our framework especially outperforms the secondhighest method for wood in texture category by 16.4% and for zipper in object category by 7.5%.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The average performance gap between DLA and LA on grid, tile, bottle, and metal nut is 25.4%, and the gap on texture category (14.02%) is larger than on object category (8.17%).

    “The average performance gap on these categories is 25.4%, and the performance gap between our framework and LA on texture category (14.02%) is averagely larger than the gap on object category (8.17%).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    On MVTec AD, the proposed method achieves the best average AUROC among compared baselines for most categories and overall average.

    “In Table 1, we show that our method achieves the best performance for most categories as well as the average of all categories.”
    • Also stated in: § Setup. · plain
      “Our framework that is trained from scratch achieves the best performance on average.”
    • Also stated in: § V. CONCLUSION · hedged
      “Our method achieved competitive performance for pixelwise anomaly segmentation.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The proposed method outperforms the second-highest baseline by 12.66% in texture category and by 2.53% in object category on MVTec AD.

    “Our method outperforms the second-highest baseline by 12.66% in texture category, by 2.53% in object category.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The average AUROCs of baselines in wood and zipper are 68.52% and 80.92%, while the proposed method achieves 89.4% and 95.5% respectively.

    “The average AUROCs of the baselines in wood and zipper classes are 68.52% and 80.92% respectively, however the performance of our method is 89.4% and 95.5% respectively.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    Hyper-parameter tuning was performed on the validation set using normal samples and pseudo abnormal samples generated by local augmentations.

    “In contrast to other baselines, our framework can generate the pseudo abnormal samples using local augmentations, so we did hyper-parameter tuning on the validation set using normal samples and pseudo abnormal samples.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: In anomaly detection, the training set consists of only normal samples, making hyper-parameter tuning using a validation set difficult.

  • May combine two claims.

    Interpretationown data

    From experiment results, the importance of an effective anomaly sample and dense features is shown.

    “From the experiment results, we have shown the importance of an effective anomaly sample and dense features.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    The paper has first explored uncertainty scores for pixel-level anomaly detection with self-supervised learning.

    “Also, we have first explored uncertainty scores for pixel-level anomaly detection with self-supervised learning.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    One-class classifiers have shown competitive results on image-level classification with high-level representations.

    “For example, One-class classifiers [1], [2] have shown competitive results on imagelevel classification with high-level representations.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [1]

      Not checked in the free run.

    • cites [2]

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    The performance drop on texture category is bigger than object class because static local augmentation is not effective in texture category causing wrong decision boundary.

    “The performance drop on the texture category is bigger than the object class, since static local augmentation is not effective in the texture category causing the wrong decision boundary.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationnone given

    The proposed method is a simple yet effective framework by self-supervised learning with dynamic local augmentation in anomaly detection.

    “We proposed a simple yet effective framework by selfsupervised learning with dynamic local augmentation in anomaly detection.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    GAN-based models detect anomalies when the confidence score is below a certain threshold.

    “GAN-based models detect anomalies when the confidence score is below a certain threshold.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [3]-[5]

      Not checked in the free run.

  • May combine two claims.

    Interpretationnone given

    The paper presents a novel augmentation method called dynamic local augmentation to generate pseudo abnormal images and effectively detect anomalies at test time.

    “We present a novel augmentation method called dynamic local augmentation to generate pseudo abnormal images and effectively detect the anomalies at test time.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    The reported average AUROC is based on three independent runs with different random seeds.

    “We report the average AUROC (%) of three independent runs with different random seeds in Table 1.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation

    AE-based models are trained to reconstruct normal input data and determine anomalies when reconstruction error is high at test time.

    “AEbased models are trained to reconstruct the input data which is normal, and determine anomalies when reconstruction error is high at test time.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [6]-[8]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Self-supervised learning on only normal training data has limitations because the model has never been trained with abnormal data.

    “Self-supervised learning on only normal training data has limitations, since the model has never been trained with abnormal data.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: In anomaly detection tasks, only normal data is given during training.

  • May combine two claims.

    Interpretationreasoning

    The model's uncertainty will be high when the input is abnormal test data, because the model is trained on only normal data.

    “The intuition is that since the model is trained on only normal data, the model's uncertainty will be high when the input is abnormal test data.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: In anomaly detection tasks, only normal data is given during training.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    A few works have shown the effectiveness of contrastive learning in image-level anomaly detection.

    “A few works [15] have shown the effectiveness of contrastive learning in image-level anomaly detection.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [15]

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Both GAN-based and AE-based anomaly detection methods consider pixelwise reconstruction error.

    “Since both methods consider pixelwise reconstruction error, the methods lack to capture the high-level features and semantic information [9], [10].”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [9]

      Not checked in the free run.

    • cites [10]

      Not checked in the free run.

19 of 19

  • The authors state this as an assumption.

    In anomaly detection tasks, only normal data is given during training.

    “In most anomaly detection tasks, we assume that only the normal data is given during the training and the model predicts whether a test sample is normal or not during the test time.”

  • The authors state this as an assumption.

    Anomalies rarely occur and anomaly patterns are not given in the training dataset.

    “In general, one of the assumptions in anomaly detection is that anomalies rarely occur and anomaly patterns are not given in the training dataset.”

  • The paper relies on this without support.

    Anomalous patterns appear in the local regions of an image.

    “In anomaly detection, anomalous patterns appear in the local regions of an image, so we employ dynamic local augmentation to generate a negative pair of the images from the normal training dataset.”

  • The paper relies on this without support.

    The category of the training data is the same in the anomaly detection task.

    “Note that the category of the training data is the same in the anomaly detection task, so we set v i and v j as positive pairs although they are transformed from either the same image or different images.”

  • The paper relies on this without support.

    MVTec AD is not a large-scale dataset.

    “Since the training data, e.g., MVTec AD, is not a large-scale dataset, our framework does not rely on large mini-batch training, unlike BYOL.”

  • The paper relies on this without support.

    Hard augmentations, such as rotation, are harmful and unused for standard contrastive learning.

    “Previous works observed that”

  • The paper relies on this without support.

    Defects are often revealed as locally changed color.

    “We observe that the defects are often revealed as the locally changed color.”

  • The paper relies on this without support.

    Defects such as hole, crack, and gray stroke are usually gray.

    “Fourth, we add Grayscale transformation to easily detect a defect, e.g., hole, crack, and gray stroke, which is usually gray.”

  • The paper relies on this without support.

    Learning local features of an image is important in anomaly detection.

    “Second, we observe that learning local features of an image is important in anomaly detection.”

  • The paper relies on this without support.

    The category and random location of an image affect the effectiveness of local augmentations to generate negative samples.

    “Our intuition is that the category and the random location of the image affect the effectiveness of the local augmentations to generate the negative sample.”

  • The paper relies on this without support.

    Not many works have been proposed for pixel-level anomaly detection, which requires accurate and reliable pixel-level representation.

    “However, not many works have been proposed for pixel-level anomaly detection which requires accurate and reliable pixellevel representation.”

  • The paper relies on this without support.

    Feature maps of training images differ from feature maps of abnormal test images.

    “So the intuition is that the feature maps of training images from each layer differs from the feature maps of abnormal images on test data.”

  • The paper relies on this without support.

    In anomaly detection, the training set consists of only normal samples, making hyper-parameter tuning using a validation set difficult.

    “In anomaly detection, the training set consists of only normal samples, so the hyper-parameter tuning using the validation set is difficult in the anomaly detection task.”
13 of 13

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Ablation improvement over fixed augmentation is misstated

    Minor

    The reported total averages for DLA and LA are 93.04% and 82.92%. DLA is 10.12 percentage points higher, not 9.90.

    Reported 9.9 · recomputed 10.12

  • Zipper result in prose differs from Table 1

    Minor

    The prose assigns 95.5% to the proposed method for zipper, while Table 1 reports 95.4%.

  • Wood result in prose differs from Table 1

    Minor

    The prose assigns 89.4% to the proposed method for wood, while Table 1 reports 89.2%.

3 of 3

Cites

Flagged first, then by citations

Research works printed but not linked: 3

  1. akcay, 2018, GANomaly: Semi-supervised anomaly detection via adversarial training, Proc Asian Conf Comput Vis, 622
  2. zong, 2018, Deep autoencoding Gaussian mixture model for unsupervised anomaly detection, Proc Int Conf Learn Represent, 2
  3. lakshminarayanan, 2017, Simple and scalable predictive uncertainty estimation using deep ensembles, Proc NIPS, 1

Cited by

Sorted by citations

Related

Authors
Yoa, S., Lee, S., Kim, C., & Kim, H. J.
Affiliations
Korea University (KR)
ORCID
Venue
IEEE Access
Publisher
Institute of Electrical and Electronics Engineers (IEEE)
Published
2021
Volume
9
Pages
147201-147211
ISSN
2169-3536 (ISSN-L)
DOI
10.1109/access.2021.3124525
OpenAlex
W3208975336
Semantic Scholar
f215dde930dce032d682692f54cb5c77af694d0f
Cited by
38
References
52
Citations by year
2026: 6 · 2025: 9 · 2024: 11 · 2023: 11 · 2022: 1
FWCI
2.03
Type
journal article
Language
en
Open access
Licence
Locations
  • IEEE Access (publishedVersion, cc-by-nc-nd)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
Samsung
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Anomaly Detection Techniques and Applications
  • Physical Sciences › Engineering › Industrial and Manufacturing Engineering › Industrial Vision Systems and Defect Detection
  • Physical Sciences › Computer Science › Artificial Intelligence › Adversarial Robustness in Machine Learning
Keywords
unsupervised anomaly detection, self-supervised learning, contrastive learning, negative sample generation, image anomaly detection

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-09.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-09.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review