The DOI matches its Crossref record; 205 of the 227 references it cites were found in a registry or an index, and 8 are not checked yet; 25 of the 64 sources its claims cite were checked for support.
Deep Industrial Image Anomaly Detection: A Survey
Open accessAbstract
The recent rapid development of deep learning has laid a milestone in industrial image anomaly detection (IAD). In this paper, we provide a comprehensive review of deep learning-based image anomaly detection techniques, from the perspectives of neural network architectures, levels of supervision, loss functions, metrics and datasets. In addition, we extract the promising setting from industrial manufacturing and review the current IAD approaches under our proposed setting. Moreover, we highlight several opening challenges for image anomaly detection. The merits and downsides of representative network architectures under varying supervision are discussed. Finally, we summarize the research findings and point out future research directions.
Verified 2026-10-10 · 8 references not checked yet
Overview
CoChat Summary
TL;DR
This survey reviews deep learning-based industrial image anomaly detection (IAD) across network architectures, supervision levels, loss functions, metrics, and datasets, and proposes an industrial manufacturing setting for evaluation. It finds memory-bank methods strongest for image-level detection, reconstruction-based methods stronger for pixel-level segmentation, and highlights open challenges in datasets, metrics, efficiency, and real-line deployment.
Key findings
- The survey covers 20 IAD datasets and 6 metrics, compared with 9 datasets/3 metrics in Tao et al. and 7 datasets/1 metric in Cui et al. (Table 1); it also states that most deep learning IAD methods are more than 97% accurate.
- Memory bank-based approaches are the most effective neural network design for image-level anomaly detection, but they are inadequate at pixel-level anomalies; ensemble learning can dramatically improve state-of-the-art methods, and SSPCAB can be seamlessly integrated into cutting-edge methods to significantly enhance reconstruction-based methods.
- Without ensemble learning, MemSeg achieved the SOTA result on image-level anomaly classification, attributed mainly to its U-Net framework; DRAEM also uses U-Net and outperformed other methods on pixel-level anomaly segmentation. AU-PR is more valuable than AU-ROC for segmentation tasks, and reconstruction-based methods outperform other methods on pixel AU-PR.
- The gap between few-shot IAD and vanilla IAD is narrowing; segmentation modules are beneficial for anomaly detection even when trained with artificially generated anomalies. Future directions include multi-modality IAD datasets, inference speed, industrial pretrained models, fully supervised pixel-level datasets, abnormal synthesis, lightweight models, and unifying industrial and medical IAD.
Methods
- Literature review/survey; no new primary experiments. It reviews deep learning IAD from the perspectives of neural network architectures, levels of supervision, loss functions, metrics, and datasets, and extracts a promising industrial manufacturing setting.
- Taxonomy: unsupervised anomaly detection is divided into feature-embedding methods (teacher-student, one-class classification, distribution map, memory bank) and reconstruction-based methods (autoencoder, GAN, transformer, diffusion); supervised anomaly detection is also reviewed. Industrial settings covered are few-shot anomaly detection, noisy anomaly detection, 3D anomaly detection, and anomaly synthesis.
- Performance analysis: statistical comparison of current IAD methods on MVTec AD using image AUROC and pixel AUROC/AUPR. Results are reported from original papers, with some reproduced by the authors and marked with *.
Limitations
- The paper states that the public cannot access real-world anomalous datasets due to privacy concerns; most open-source IAD datasets generate anomalies from anomaly-free products, so their abnormalities may not occur in actual production lines, and multiple-domain IAD datasets are lacking.
- The paper states that there is no uniform assessment for image-level and pixel-level IAD performance; most metrics shrink the anomalous mask to the feature-map size, reducing assessment precision, and the validity of common metrics for IAD remains to be explored. It also notes that most noisy anomaly detection methods have not been verified on real industrial image datasets, and most 3D IAD methods are specialized for RGB-D images while real-world 3D manufacturing data consists of point clouds.
- Evident from the text: the survey’s performance tables contain many missing entries and rely largely on reported results, with only some reproduced; no new benchmark experiments are conducted. The paper also states that current algorithms often focus on detection accuracy while ignoring storage size and efficiency.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
18 Valid · 21 Not valid · 26 Not checked · 62 Own results and interpretations
Every source the paper cites for this claim was checked and supports it.
In the paper
“FCDD [35] is an unsupervised method that synthesizes abnormal samples for training the OCC model.”
What the sources say
- Supportedcites [35] P. Liznerski, L. Ruff, R. A. Vandermeulen, B. J. Franks, M. Kloft, K. R. Muller. Explainable deep one-class classification. In Proceedings of the 9th International Conference on Learning Representations, 2021. source
Section 4.2 (pp. 7–8) describes generating synthetic anomalies with confetti noise for FCDD and evaluates it in an unsupervised setting. Section 3 (p. 4) confirms that synthetic anomalies can be used for training when known anomalies are unavailable. Synthetic anomalies are an option, not a requirement of FCDD. (LLM, conf=0.87)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“PaDiM requires the batch-inverse of the multidimensional covariance tensor, which makes it challenging to scale up to larger CNNs due to the increased feature size.”
What the sources say
- Partly supportedcites [73] T. Defard, A. Setkov, A. Loesch, R. Audigier. PaDiM: A patch distribution modeling framework for anomaly detection and localization. In Proceedings of International Conference on Pattern Recognition - ICPR International Workshops and Challenges, Springer, Milano, Italy, pp. 475–489, 2021. DOI: https://doi.org/10.1007/978-3-030-68799-1_35. sourceExperimental — accuracy not yet measured
“Learning of the normality” (p. 3) says PaDiM estimates a covariance matrix for each patch and makes it invertible. “Ablative studies” (pp. 4–5) says larger embedding vectors increase computational and memory complexity. The copy does not say that PaDiM requires a batch-inverse operation or that this makes scaling to larger CNNs challenging; “Scalability gain” (p. 6) instead describes choosing the backbone and embedding size to fit resource limits. (LLM, conf=0.85)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“Since the relative positions of the features of each image layer do not change during the convolution process, FCDD yields more interpretable results than alternative methods.”
What the sources say
- Supportedcites [35] P. Liznerski, L. Ruff, R. A. Vandermeulen, B. J. Franks, M. Kloft, K. R. Muller. Explainable deep one-class classification. In Proceedings of the 9th International Conference on Learning Representations, 2021. source
Section 3 (p. 3) says convolutional layers preserve the relative positions of receptive fields and that FCDD’s fully convolutional network preserves spatial information. The Introduction (p. 1) reports that FCDD’s anomaly heatmaps are less noisy and more structured than those of the baselines, while Section 4.2 (p. 8) reports better quantitative explanation performance than its competitors. (LLM, conf=0.90)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“The performance of these methods is outstanding, but they depend on pre-trained models such as ResNet [7] VGG [8] and EfficientNet [9] .”
What the sources say
- Not checked — no full text of the sourcecites [7] K. M. He, X. Y. Zhang, S. Q. Ren, J. Sun. Deep residual learning for image recognition. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, USA, pp. 770–778, 2016. DOI: https://doi.org/10.1109/CVPR.2016.90. source
- Supportedcites [8] K. Simonyan. A. Zisserman. Very deep convolutional networks for large-scale image recognition. In Proceedings of the 3rd International Conference on Learning Representations, San Diego, USA, 2015. DOI: https://doi.org/10.48550/arXiv.1409.1556. source
The citation is attached to VGG as an example of a pre-trained model, not to the broader claim about teacher-student methods. In “Generalisation of Very Deep Features” (pp. 11–12), the paper identifies its method as VGG and describes using its ConvNets pre-trained on ILSVRC. (LLM, conf=0.92)
- Supportedcites [9] M. X. Tan, Q. Le. EfficientNet: Rethinking model scaling for convolutional neural networks. In Proceedings of the 36th International conference on machine learning, Long Beach, USA, pp. 6105–6114, 2019. source
Citation [9] is attached to EfficientNet as an example of a pre-trained model. Section 5.3 (p. 8) identifies the authors’ EfficientNet models and describes using ImageNet-pretrained checkpoints for transfer learning. (LLM, conf=0.90)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Few-shot anomaly detection (FSAD) [139, 140] is still in its infancy.”
What the sources say
- Not checked — no full text of the sourcecites [139] J. C. Wu, D. J. Chen, C. S. Fuh, T. L. Liu. Learning unsupervised metaformer for anomaly detection. In Proceedings of IEEE/CVF International Conference on Computer Vision, Montreal, Canada, pp. 4349–4358, 2021. DOI: https://doi.org/10.1109/ICCV48922.2021.00433. source
- Partly supportedcites [140] A. M. Kamoona, A. K. Gostar, A. Bab-Hadiashar, R. Hoseinnezhad. Anomaly detection of defect using energy of point pattern features within random finite set framework, [Online], Available: https://arxiv.org/abs/2108.12159, 2021. sourceExperimental — accuracy not yet measured
The few-shots learning experiments (p. 9) show that this paper studies anomaly detection with limited normal samples, supporting its citation as an example of FSAD. That passage does not assess whether FSAD is still in its infancy, so the maturity claim is not supported by this source. (LLM, conf=0.82)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
46 more claims · 2 with their cited sources checked
Create a free account to see all 127 claims and to verify the 17 references still open.
- It may state two separate claims as one. 26
- Probably not a substantial claim. 64
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.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 2.1.4 Memory bank · p. 8
“Npad [74] allows for the possibility of marginal misalignment”
§ 2.1.4 Memory bank · p. 8
“| N-pad [71] | Mahalanobis distance, log- likelihood | ResNet | The paper allows for possible edge misalignment by estimating a nominal distribution for each pixel using the pixel ′ s neighborhood features. |”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 2.1.4 Memory bank · p. 8
“The primary distinction between memory bank-based methods and OCC-based methods, is that memory-based methods, such as SVDD, require additional memory space to store image features.”
§ 2.1.2 One-class classification · p. 4
“SVDD [34] is a classic algorithm in the OCC problem”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 4.3 3D anomaly detection · p. 15
“Since the release of MVTec 3D-AD [6] dataset”
§ 5 Datasets and metrics · p. 16
“| MVTec 3D-AD [174] | 10 | 2 904 | 948 | 3 852 | Segmentation mask | Real |”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 2.1.2 One-class classification · p. 4
“with the exception of MemSeg [18] , the training of other methods relies on SVDD and cross-entropy loss.”
§ 2.1.2 One-class classification · p. 5
“| CPC-AD [33] | InfoNCE | - | The paper applies contrastive predictive coding (CPC) model to AD and gets an anomaly score through pixel-wise loss. |”
Citations
Cites
Deep Residual Learning for Image Recognition
He, K., Zhang, X., Ren, S., & Sun, J. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.90 · Cited by 235,474 · cited 20×
“CFA [69] SVDD ResNet The paper improves PatchCore so that image features are distributed on a hypersphere.”
InfluentialVery Deep Convolutional Networks for Large-Scale Image Recognition
Simonyan, K., & Zisserman, A. (2014) · arXiv · doi:10.48550/arxiv.1409.1556 · Cited by 114,148 · cited 1×
“The performance of these methods is outstanding, but they depend on pre-trained models such as ResNet [7] VGG [8] and EfficientNet [ ?”
U-Net: Convolutional Networks for Biomedical Image Segmentation
Ronneberger, O., Fischer, P., & Brox, T. (2015) · Lecture notes in computer science · doi:10.1007/978-3-319-24574-4_28 · Cited by 99,404 · cited 8×
“DREAM [91] L 2 , SSIM, Focal - ThepaperdesignsamethodtogenerateabnormalimagesandusesU-Net[92]todistinguishanomaliesafterreconstruction.”
InfluentialAn Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al. (2020) · arXiv · doi:10.48550/arxiv.2010.11929 · Cited by 47,472
Squeeze-and-Excitation Networks
Hu, J., Shen, L., & Sun, G. (2018) · 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2018.00745 · Cited by 35,667 · cited 3×
“This paper replaces the SE-layer in SSPCAB with transformer architecture.”
Swin Transformer: Hierarchical Vision Transformer using Shifted Windows
Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., & Guo, B. (2021) · 2021 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv48922.2021.00986 · Cited by 35,534 · cited 3×
“NF methods extract features from normal images from a pre-trained model, such as ResNet [59] or Swin Transformer [60], and transform the feature distribution as a Gaussian distribution during the training phase.”
CBAM: Convolutional Block Attention Module
Woo, S., Park, J., Lee, J.-Y., & Kweon, I. S. (2018) · Lecture Notes in Computer Science · doi:10.1007/978-3-030-01234-2_1 · Cited by 25,853
Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization
Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., & Batra, D. (2017) · 2017 IEEE International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2017.74 · Cited by 24,513
Unpaired image-to-image translation using cycle-consistent adversarial networks
Zhu, J.-Y., Park, T., Isola, P., & Efros, A. A. (2017) · 2017 IEEE International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2017.244 · Cited by 23,602
Xception: Deep Learning with Depthwise Separable Convolutions
Chollet, F. (2017) · 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2017.195 · Cited by 20,380
Research works printed but not linked: 2
- J. Yoon, K. Sohn, C. L. Li, S. Ö. Arik, C. Y. Lee, T. Pfister. Self-supervise, refine, repeat: Improving unsupervised anomaly detection. Transactions on Machine Learning Research, to be published.
- M. Bińkowski, D. J. Sutherland, M. Arbel, A. Gretton. Demystifying MMD GANs. In Proceedings of the 6th International Conference on Learning Representations, Vancouver, Canada, 2018.
Cited by
A Diffusion-Based Framework for Multi-Class Anomaly Detection
He, H., Zhang, J., Chen, H., Chen, X., Li, Z., Chen, X., Wang, Y., Wang, C., & Xie, L. (2024) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v38i8.28690 · Cited by 212
“Anomaly detection is a crucial task in computer vision and industrial applications (Tao et al. 2022; Salehi et al. 2022; Liu et al. 2023), which goal of visual anomaly detection is to determine anomalous images and locate the regions of anomaly accurately.”
BackgroundMethodSelf-citationReal-IAD: A Real-World Multi-View Dataset for Benchmarking Versatile Industrial Anomaly Detection
Wang, C., Zhu, W., Gao, B.-B., Gan, Z., Zhang, J., Gu, Z., Qian, S., Chen, M., & Ma, L. (2024) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr52733.2024.02159 · Cited by 353
Self-citationA survey of deep learning for industrial visual anomaly detection
Li, Z., Yan, Y., Wang, X., Ge, Y., & Meng, L. (2025) · Artificial Intelligence Review · doi:10.1007/s10462-025-11287-7 · Cited by 194
IM-IAD: Industrial Image Anomaly Detection Benchmark in Manufacturing
Xie, G., Wang, J., Liu, J., Lyu, J., Liu, Y., Wang, C., Zheng, F., & Jin, Y. (2024) · IEEE Transactions on Cybernetics · doi:10.1109/tcyb.2024.3357213 · Cited by 163
“Unsupervised IAD: There are two streams methods on unsupervised IAD [50], namely, feature embedding-based methods [4], [14], [67], and reconstruction-based meth-ods [86], [95], [100].”
MethodSelf-citationChen, Q., Luo, H., Lv, C., & Zhang, Z. (2024) · Lecture notes in computer science · doi:10.1007/978-3-031-72855-6_3 · Cited by 225
Metadata
- Authors
- Liu, J., Xie, G., Wang, J., Li, S., Wang, C., Zheng, F., & Jin, Y.
- Affiliations
- Southern University of Science and Technology (CN)
- University of Surrey (GB)
- Tencent (China) (CN)
- Bielefeld University (DE)
- Corresponding author
- Zheng, Feng
- Jin, Yaochu
- Venue
- Machine Intelligence Research
- Published
- 2024-01-15
- Volume
- 21
- Issue
- 1
- Pages
- 104-135
- ISSN
- 1476-8186 (ISSN-L)
- 2731-538X
- 2731-5398
- OpenAlex
- W4390875033
- Semantic Scholar
- 230c875f7563abf2e11bc79c0ae8855bfa52123c
- Cited by
- 463
- References
- 172
- Citations by year
- 2026: 193 · 2025: 178 · 2024: 81
- FWCI
- 92.07
- Type
- journal article
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- cc-by
- creativecommons.org/licenses/by/4.0 (tdm)
- creativecommons.org/licenses/by/4.0 (vor)
- Locations
- Machine Intelligence Research (publishedVersion, cc-by)
- arXiv (Cornell University) (submittedVersion, cc-by-nc-sa)
- PUB – Publications at Bielefeld University (Bielefeld University) (submittedVersion, cc-by)
- arXiv (Cornell University)
- Funders
- National Natural Science Foundation of China (62206122, 62122035)
- Bundesministerium für Bildung und Forschung
- Universität Bielefeld
- Topics
- Physical Sciences › Engineering › Industrial and Manufacturing Engineering › Industrial Vision Systems and Defect Detection
- Physical Sciences › Computer Science › Artificial Intelligence › Anomaly Detection Techniques and Applications
- Physical Sciences › Engineering › Civil and Structural Engineering › Infrastructure Maintenance and Monitoring
- Keywords
- deep learning, neural network architecture, supervised and unsupervised learning, loss functions, evaluation metrics, manufacturing industry, quality inspection
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: Deep Industrial Image Anomaly Detection: A Survey. https://app.cochat.ai/papers/10.1007/s11633-023-1459-z
- The paper
- Liu, J., Xie, G., Wang, J., Li, S., Wang, C., Zheng, F., & Jin, Y. (2024). Deep Industrial Image Anomaly Detection: A Survey. Machine Intelligence Research. https://doi.org/10.1007/s11633-023-1459-z
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.