Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 35 of the 38 references it cites were found in a registry or an index; 3 of the 13 sources its claims cite were checked for support.

VT-ADL: A Vision Transformer Network for Image Anomaly Detection and Localization

Open access
Mishra, Pankaj, Verk, Riccardo, Fornasier, Daniele, Piciarelli, Claudio, Foresti, Gian Luca · 2021 · 2021 IEEE 30th International Symposium on Industrial Electronics (ISIE)

Abstract

We present a transformer-based image anomaly detection and localization network. Our proposed model is a combination of a reconstruction-based approach and patch embedding. The use of transformer networks helps preserving the spatial information of the embedded patches, which is later processed by a Gaussian mixture density network to localize the anomalous areas. In addition, we also publish BTAD, a real-world industrial anomaly dataset. Our results are compared with other state-of-the-art algorithms using publicly available datasets like MNIST and MVTec.

Overview

CoChat Summary

TL;DR

VT-ADL combines a Vision Transformer patch encoder, a reconstruction decoder, and a Gaussian mixture density network for semi-supervised image anomaly detection and localization. It also releases BTAD, a real-world industrial anomaly dataset, and reports competitive results on MNIST, MVTec, and BTAD.

Key findings

  • On MNIST, VT-ADL achieves a mean AUC of 0.984, compared with 0.95 for OC SVM, 0.97 for VAE, 0.97 for LSA, and 0.97 for Pyr. AE.
  • On MVTec, VT-ADL reaches a mean PRO score of 0.807, performing at par with recent state-of-the-art methods and outperforming them in 7 product categories; the best comparator, Uni. Stud., has a mean of 0.857.
  • On the newly published BTAD dataset, VT-ADL reports a mean PRO score of 0.89 and mean PR AUC of 0.90, compared with convolutional autoencoders using MSE loss (0.78) and MSE+SSIM loss (0.79).
  • Adding Gaussian noise to transformer-encoded features improves generalization: with 150 Gaussians, PRO score is 0.897 with noise versus 0.807 without noise.

Methods

  • The model is trained semi-supervised using only normal data. Images are split into patches, embedded with positional information, encoded by a Vision Transformer, decoded for reconstruction, and modeled by a Gaussian mixture density network for anomaly localization.
  • Training combines MSE, SSIM, and negative conditional log-likelihood losses, with Gaussian noise N(0, 0.2) added to transformer-embedded features for regularization. Hyperparameters include Adam learning rate 0.0001, weight decay 0.0001, batch size 8, 400 epochs, 150 Gaussians, and patch dimension P = 64.
  • Experiments use MNIST, MVTec (5,354 high-resolution images, scaled to 550×550 then center cropped to 512×512), and BTAD (2,830 real-world images of 3 industrial products, with 400, 1000, and 399 training images respectively). Evaluation uses AUC for MNIST and PRO for MVTec and BTAD.

Limitations

  • The paper does not include a dedicated limitations section. It states that dropout was not used because it causes instability in the Gaussian approximation network.
  • Evident from the reported results: on MVTec, VT-ADL’s mean PRO (0.807) is below Uni. Stud. (0.857), and it underperforms that method in several categories, such as Capsule (0.672 vs. 0.916), Pill (0.705 vs. 0.935), and Metal Nut (0.726 vs. 0.895).
  • Evident from the text: BTAD results are described as the first results on the new dataset, and the comparison is limited to basic convolutional autoencoders with MSE and MSE+SSIM losses.

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

Integrity signals

Registry

Crossref · 10.1109/isie45552.2021.9576231

Cites

35 of 35 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 The three products of BTAD dataset. First column shows an example of normal images, second column shows anomalous images, third column shows the anomalous image with pixel-level ground truth labels, fourth column shows the predicted heat map by our proposed method.
Fig. 1 The three products of BTAD dataset. First column shows an example of normal images, second column shows anomalous images, third column shows the anomalous image with pixel-level ground truth labels, fourth column shows the predicted heat map by our proposed method.From the PDF
Fig. 2 Left image: model overview. Image is split into patches, which are augmented with positional embedding. The resulting sequence is fed to the Transformer encoder. Then encoded features are summed into a reconstruction vector which is fed to decoder. The transformer encoded features are also fed into an Gaussian approximation network [10], which is later used to localize the anomaly. Right image: detailed structure of the tranformer encoder (image from [9]).
Fig. 2 Left image: model overview. Image is split into patches, which are augmented with positional embedding. The resulting sequence is fed to the Transformer encoder. Then encoded features are summed into a reconstruction vector which is fed to decoder. The transformer encoded features are also fed into an Gaussian approximation network [10], which is later used to localize the anomaly. Right image: detailed structure of the tranformer encoder (image from [9]).From the PDF
Fig. 3 Anomaly detection on MVTec dataset. First row shows the actual anomalous image of bottle, cable, capsule, metal nut and brush. second row shows the actual ground truth and third row shows the generated anomaly score and anomaly localization by our method
Fig. 3 Anomaly detection on MVTec dataset. First row shows the actual anomalous image of bottle, cable, capsule, metal nut and brush. second row shows the actual ground truth and third row shows the generated anomaly score and anomaly localization by our methodFrom the PDF
Fig. 4 Plot shows the PRO score for the different no of Gaussians used in the Gaussian approximation.
Fig. 4 Plot shows the PRO score for the different no of Gaussians used in the Gaussian approximation.From the PDF

Verification

Claims the paper makes

2 Valid · 2 Not valid · 6 Not checked · 14 Own results and interpretations

8 more claims

Create a free account to see all 24 claims and to verify the 3 references still open.

See all claims — free account
  • It may state two separate claims as one. 10
  • The quote only partly states this claim. 1
  • Probably not a substantial claim. 2

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Citations

Cites

Show all 38 references

Research works printed but not linked: 1

  1. goodfellow, 2016, Deep Learning

Cited by

See all 462

Metadata

Authors
Mishra, P., Verk, R., Fornasier, D., Piciarelli, C., & Foresti, G. L.
Affiliations
University of Udine (IT)
Venue
2021 IEEE 30th International Symposium on Industrial Electronics (ISIE)
Publisher
IEEE
Published
2021-06-20
Pages
01-06
OpenAlex
W3154719617
Cited by
462
References
38
Citations by year
2026: 91 · 2025: 139 · 2024: 124 · 2023: 81 · 2022: 26 · 2021: 1
FWCI
26.74
Type
conference paper
Language
en
Open access
Locations
  • IEEE International Symposium on Industrial Electronics (ISIE) (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • Institutional Research Information System (University of Udine) (submittedVersion)
  • arXiv (Cornell University)
Funders
ONRG (N62909-20-1-2075)
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 › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
Keywords
image anomaly detection, anomaly localization, vision transformer, patch embedding, reconstruction-based 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.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

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

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: VT-ADL: A Vision Transformer Network for Image Anomaly Detection and Localization. https://app.cochat.ai/papers/10.1109/isie45552.2021.9576231
The paper
Mishra, P., Verk, R., Fornasier, D., Piciarelli, C., & Foresti, G. L. (2021). VT-ADL: A Vision Transformer Network for Image Anomaly Detection and Localization. 2021 IEEE 30th International Symposium on Industrial Electronics (ISIE). https://doi.org/10.1109/isie45552.2021.9576231

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.