Explicit Boundary Guided Semi-Push-Pull Contrastive Learning for Supervised Anomaly Detection

Xincheng YaoRuoqi LiJing ZhangJun SunChongyang Zhang

article2023CVPR114 citations

Proposes a supervised anomaly detection framework that leverages normal feature distributions to establish explicit decision boundaries and applies a semi-push-pull contrastive loss, enabling models to exploit limited known anomalies without losing generalization to unseen defect types.

Listen

Automated anomaly detection is essential across industrial defect inspection and medical imaging. Most existing systems rely strictly on unsupervised learning from normal data, which results in blurry decision boundaries and poor detection accuracy. While incorporating a small number of available abnormal samples can sharpen performance, existing supervised approaches tend to overfit to known defect types, failing when confronted with new, unseen anomalies.

The article demonstrates an explicit boundary guided framework for supervised anomaly detection that exploits limited anomaly data while preventing model bias against unseen defect types.

The authors develop a two-stage method combining normalizing flows with a boundary-guided semi-push-pull learning mechanism. First, the model maps normal image features into a standard distribution to define a clear boundary derived entirely from normal data. Next, a targeted training objective pulls near-boundary normal features inward while pushing abnormal features across an explicit safety margin. To address anomaly rarity, the approach also synthesizes realistic irregular patterns on normal images. The method was evaluated across six public benchmarks covering industrial manufacturing (MVTecAD, BTAD, AITEX, ELPV) and medical diagnostics (BrainMRI, HeadCT) under multi-class and unseen defect settings.

Across the evaluations, the proposed method achieved top performance, reaching 99.3% image-level and 99.2% pixel-level accuracy on the MVTecAD benchmark. When identifying entirely new, unseen anomaly classes, the framework outperformed competing supervised approaches by 3.3% to 12.9% in detection accuracy. On challenging and subtle defect subsets, the method increased detection performance by 3.6% and localization precision by 8.6% compared to the baseline. Visual assessments confirmed that the approach reliably eliminates ambiguous scoring regions, producing cleaner and more precise anomaly localization maps.

These results show that establishing separating boundaries anchored strictly in the normal data distribution prevents models from becoming biased toward known defects. For operational decision-makers, this translates to higher inspection reliability, reduced false alarms, and lower risk of overlooking novel defects in safety-critical manufacturing and clinical workflows.

Organizations deploying visual quality control or medical diagnostic systems should consider incorporating explicit boundary guidance and synthetic defect generation into their pipelines when small sets of historical defects are available. Before full-scale operational rollout, engineering teams should conduct pilot testing on domain-specific datasets to fine-tune margin parameters and assess computational overhead during live image processing.

While the reported performance gains are statistically robust across diverse benchmarks, confidence remains highest in standardized visual domains with clear foreground objects. Practitioners should exercise caution when applying the model to highly variable backgrounds or uncalibrated image streams until dedicated validation is complete.

anomaly detection methods contrastive learning detection algorithm supervised anomaly detection semi-supervised anomaly detection contrastive learning

arXiv: 2207.01463

Paper: Catching Both Gray and Black Swans: Open-set Supervised Anomaly Detection, Choubo Ding et al. (2022). This paper establishes the problem formulation of open-set supervised anomaly detection using few labeled anomalies while generalizing to unseen anomalies, which the source directly builds upon.
Paper: Deep One-Class Classification, Lukas Ruff et al. (2018). Deep SVDD introduces the foundational deep one-class boundary optimization concept that encloses normal feature distributions to detect anomalies.
Paper: Support Vector Data Description, DAVID M.J. TAX et al. (2004). This classic work develops Support Vector Data Description for learning tight, compact separating boundaries around normal data, forming the conceptual basis for explicit boundary guidance.
Paper: Latent Outlier Exposure for Anomaly Detection with Contaminated Data, Chen Qiu et al. (2022). This work introduces paired pull-normal and push-abnormal objectives in feature space to isolate anomalies, directly preceding the source paper's semi-push-pull contrastive learning mechanism.
Paper: Deep Anomaly Detection with Outlier Exposure, Dan Hendrycks et al. (2019). Outlier Exposure provides the essential conceptual baseline of leveraging known auxiliary anomaly samples to improve generalization to unseen out-of-distribution instances.
Paper: Deep Learning for Anomaly Detection, Guansong Pang et al. (2020). This survey provides a comprehensive taxonomy of deep anomaly score learning and weakly supervised few-shot normality representations that motivate the source's methodology.
Paper: MVTec AD — A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection, Paul Bergmann et al. (2019). This paper introduces the standard MVTec AD benchmark and evaluation protocol utilized for visual anomaly detection models.

Paper: AnomalyCLIP: Object-agnostic Prompt Learning for Zero-shot Anomaly Detection, Qihang Zhou et al. (2024). Extends the detection of unseen anomalies by utilizing vision-language foundation models with object-agnostic prompt learning in a zero-shot regime.

Cover for Explicit Boundary Guided Semi-Push-Pull Contrastive Learning for Supervised Anomaly Detection

Abstract

Most anomaly detection (AD) models are learned using only normal samples in an unsupervised way, which may result in ambiguous decision boundary and insufficient discriminability. In fact, a few anomaly samples are often available in real-world applications, the valuable knowledge of known anomalies should also be effectively exploited. However, utilizing a few known anomalies during training may cause another issue that the model may be biased by those known anomalies and fail to generalize to unseen anomalies. In this paper, we tackle supervised anomaly detection, i.e., we learn AD models using a few available anomalies with the objective to detect both the seen and unseen anomalies. We propose a novel explicit boundary guided semi-push-pull contrastive learning mechanism, which can enhance model’s discriminability while mitigating the bias issue. Our approach is based on two core designs: First, we find an explicit and compact separating boundary as the guidance for further feature learning. As the boundary only relies on the normal feature distribution, the bias problem caused by a few known anomalies can be alleviated. Second, a boundary guided semi-push-pull loss is developed to only pull the normal features together while pushing the abnormal features apart from the separating boundary beyond a certain margin region. In this way, our model can form a more explicit and discriminative decision boundary to distinguish known and also unseen anomalies from normal samples more effectively. Code will be available at https://github.com/xcyao00/BGAD.

Topics

anomaly detection methods contrastive learning detection algorithm

Keywords

supervised anomaly detection semi-supervised anomaly detection contrastive learning deep anomaly detection novelty detection pseudo anomaly generation outlier exposure medical lesion detection anomaly scoring

1. Introduction
2. Related Work
3. Our Proposed Approach
3.1. Learning Normal Feature Distribution by Normalizing Flow
3.2. Finding an Explicit and Compact Separating Boundary
3.3. Learning More Discriminative Features by Boundary Guided Semi-Push-Pull
3.4. Generalization Capability to Unseen Anomalies
3.5. RandAugment-based Pseudo Anomaly Generation
4. Experiments
4.1. Datasets and Metrics
4.2. Experimental Settings
4.3. Results under the Multi-Class Setting
4.4. Results under the One-Class Setting
4.5. Ablation Study
4.6. Qualitative Results
5. Conclusion
Acknowledgements
References

Knowls

Knowl 1 — BGAD for supervised anomaly detection
model/method
Boundary Guided Anomaly Detection (BGAD) addresses supervised anomaly detection when training data contain many normal images and only a few known anomalies. Let $\mathcal I_n$ and $\mathcal I_a$ denote normal and known-abnormal training images, respectively; known anomalies come from seen classes $S_s$ , while the complete anomaly-class set is $S=S_s\cup S_u$ , with $S_u$ denoting unseen classes. BGAD learns an anomaly-scoring function $m:\mathcal I\to\mathbb R$ that assigns larger scores to both seen and unseen anomalies than to normal images.

BGAD consists of a feature extractor $f$ , a conditional normalizing flow $\phi_\theta$ that maps extracted $d$ -dimensional features to a latent space, an explicit-boundary-generating phase, and a boundary-guided-optimizing phase. First, the flow is trained using normal features only, and a normal log-likelihood boundary is estimated from their distribution. Second, normal and known-abnormal features are optimized using the boundary-guided semi-push-pull loss. At inference, the flow converts each feature's log-likelihood into an anomaly score. The boundary is determined only from normal features, while known anomalies are used to make the representation more discriminative without defining the final decision boundary from the particular known anomaly classes.
Knowl 2 — Normal-feature modeling with a conditional normalizing flow
model/method
BGAD models normal feature density with an invertible flow $\phi_\theta:\mathbb R^d\to\mathbb R^d$ composed of $L$ coupling layers, $\phi_\theta=\phi_L\circ\cdots\circ\phi_1$ . For an input feature $x\in\mathbb R^d$ , let $y_0=x$ and $y_l=\phi_l(y_{l-1})$ for $l=1,\ldots,L$ ; $J_{\phi_l}(y_{l-1})$ is the Jacobian of layer $\phi_l$ at $y_{l-1}$ . The latent prior is the standard Gaussian $p_Z(z)=\mathcal N(0,I_d)$ , where $I_d$ is the $d\times d$ identity matrix.

The normalizing-flow log-likelihood is
$\log p_\theta(x)= -\frac d2\log(2\pi)-\frac12\|\phi_\theta(x)\|_2^2+\sum_{l=1}^{L}\log\left|\det J_{\phi_l}(y_{l-1})\right|.$
With $\mathcal X_n$ denoting the set or distribution of normal features, BGAD trains the flow by minimizing the normal-feature maximum-likelihood loss
$\mathcal L_{\mathrm{ml}}=\mathbb E_{x\sim\mathcal X_n}\left[\frac d2\log(2\pi)+\frac12\|\phi_\theta(x)\|_2^2-\sum_{l=1}^{L}\log\left|\det J_{\phi_l}(y_{l-1})\right|\right].$
Because ordinary fully connected coupling layers can destroy spatial relationships when feature maps are flattened, BGAD adds two-dimensional positional embeddings before flow processing.
Knowl 3 — Explicit normal and abnormal likelihood boundaries
model/method
After normal-flow training, BGAD estimates the log-likelihoods of all normal features, $\mathcal P_n=\{\log p_i\}_{i=1}^{N}$ , and sorts them from low to high. The explicit normal boundary $b_n$ is the $\beta$ -th percentile of this normal log-likelihood distribution. Since low likelihood indicates low normality, choosing $b_n$ at the lower tail makes the normal false-positive rate no greater than approximately $\beta\%$ under the empirical normal distribution. The abnormal boundary is placed below the normal boundary by a margin $\tau$ :
$b_a=b_n-\tau.$
Here $N$ is the number of normal features used to estimate the distribution, $\beta$ controls how close the boundary is to the normal-distribution center or tail, and $\tau>0$ defines the ambiguous margin region $(b_a,b_n)$ . The paper gives $\beta=1$ and $\tau=0.1$ as example settings and reports that performance is not very sensitive to these hyperparameters. Crucially, both boundaries are computed from normal features only; known-abnormal samples do not influence their locations.
Knowl 4 — Likelihood-based anomaly scoring
equation
For a feature $x\in\mathbb R^d$ , BGAD converts the normalizing-flow log-likelihood into an anomaly score using
$s(x)=1-\exp\bigl(\log p(x)\bigr),$
where $p(x)$ is the density estimated by the trained normalizing flow and $\exp$ is the exponential function. Higher likelihood corresponds to greater normality, so lower likelihood produces a larger anomaly score. Because the exponential is monotonic, thresholding the log-likelihood at $b_n$ is equivalent to thresholding the anomaly score at the corresponding transformed value; BGAD therefore performs boundary construction and feature optimization in log-likelihood space while reporting anomaly scores at test time.
Knowl 5 — Boundary-guided semi-push-pull optimization
model/method
BGAD uses the explicit normal boundary $b_n$ and abnormal boundary $b_a=b_n-\tau$ as targets for contrastive feature learning. Let $\ell_i$ be the normalized log-likelihood of normal feature $i$ , $i=1,\ldots,N$ , and let $\ell_j$ be the normalized log-likelihood of abnormal feature $j$ , $j=1,\ldots,M$ . The normalization uses a sufficiently large normalizer, with $\alpha_n=10$ as the paper's example, to place log-likelihoods on an approximately $[-1,0]$ scale; extremely low values below $-1$ can be excluded because they are already readily identifiable as anomalous. The BG-SPP loss is
$\mathcal L_{\mathrm{bg\text{-}spp}}= \sum_{i=1}^{N}\left|\min(\ell_i-b_n,0)\right| + \sum_{j=1}^{M}\left|\max(\ell_j-b_n+\tau,0)\right|.$
The first term acts only on normal features whose log-likelihood is below $b_n$ , raising them toward the normal side rather than forcing every normal feature to move. The second term acts only on abnormal features whose log-likelihood is above $b_a=b_n-\tau$ , lowering them beyond the margin instead of pushing them arbitrarily far from the normal distribution. The complete second-phase objective is
$\mathcal L=\mathcal L_{\mathrm{ml}}+\lambda\mathcal L_{\mathrm{bg\text{-}spp}},$
where $\lambda\ge 0$ weights boundary-guided optimization. Minimizing the loss reduces the ambiguous likelihood region $(b_a,b_n)$ while avoiding the strong bias that can result from pushing known anomalies as far away as possible.
Knowl 6 — RandAugment-based pseudo-anomaly generation
algorithm
BGAD increases the quantity and diversity of abnormal training patterns with RandAugment-based Pseudo Anomaly Generation (RPAG). The input is a known abnormal image $I^a$ , a normal image $I^n$ , and an object-region mask derived from the normal image; the output is a simulated abnormal image $I^{sa}$ .

The augmentation set contains nine transformations: Flip, Rotate, Transpose, Noise, Distortion, Brightness, Sharpness, Translate, and Blur. RPAG randomly selects a subset of $S$ transformations from this set and applies them to the known abnormal image. It then cuts an anomalous region from the transformed image. To preserve object semantics, a grayscale binary-thresholding procedure obtains a foreground mask from $I^n$ , and a random paste location is selected only among foreground pixels. Finally, the cropped transformed abnormal region is pasted at that location in $I^n$ . Repeating this process creates local irregularities at varied locations and with varied appearances, providing additional pseudo-abnormal examples for feature learning without placing synthetic defects in the background.
Knowl 7 — Experimental protocol for known and unseen anomalies
experimental setup
BGAD is evaluated on six real-world anomaly-detection datasets: the industrial defect datasets MVTecAD, BTAD, AITEX, and ELPV, and the medical lesion datasets BrainMRI and HeadCT. Image-level AUROC measures image anomaly detection, pixel-level AUROC measures anomaly localization, and Per-Region-Overlap (PRO) evaluates localization while weighting ground-truth regions of different sizes equally.

The multi-class setting evaluates detection of known anomaly classes. A few abnormal images are randomly sampled from existing anomaly classes, removed from the test set, and used for training. The one-class setting evaluates generalization to unseen classes: training anomalies are sampled from only one anomaly class, and all samples of that class are removed from testing so that test anomalies belong to other classes. BGAD uses ten random abnormal samples per category by default, together with RPAG-generated pseudo anomalies; experiments on the other datasets follow the compared protocol of one known anomaly sample. All baseline methods are re-run under the same modified train/test splits for fair comparison.

Knowl 8 — MVTecAD and BTAD performance

data/table

Under the multi-class setting, BGAD is compared with unsupervised and supervised anomaly detectors using the same training and test protocol. The aggregate MVTecAD results are:

NFAD is a flow-based baseline with the same network structure but without explicit boundary guidance; BGADw/o uses the normal-only part of the boundary-guided loss and no abnormal samples. Relative to NFAD, full BGAD improves image-level AUROC by 2.5 percentage points, pixel-level AUROC by 1.3 points, and PRO by 3.0 points. Against supervised MVTecAD baselines, the image-level mean AUROCs are 0.965 for FCDD, 0.948 for DevNet, 0.961 for DRA, and $0.993\mathbin{\pm}0.0012$ for BGAD.

On BTAD, the mean pixel-level AUROC and PRO are $0.986\mathbin{\pm}0.0015$ and $0.824\mathbin{\pm}0.0163$ for BGAD, compared with 0.978 and 0.778 for NFAD. These results show that boundary guidance improves both image detection and localization, with particularly strong gains in region overlap.

Knowl 9 — Generalization to hard and unseen anomaly subsets

data/table

In the one-class setting, BGAD is trained with anomalies from one class and tested on other classes. On diverse application datasets, BGAD is reported as the best-performing method, with mean image-level AUROC gains of approximately 3.3%–12.9% over the strongest competing supervised detector. On MVTecAD, the aggregate results on the original, hard, and unseen subsets are:

Dataset subset	Metric	NFAD	BGAD
MVTecAD	Image AUROC	0.968	0.992( +2.5% )
MVTecAD	Pixel AUROC	0.979	0.992( +1.3% )
MVTecAD	PRO	0.946	0.976( +3.0% )
Hard subsets	Image AUROC	0.948	0.984( +3.6% )
Hard subsets	Pixel AUROC	0.960	0.986( +2.6% )
Hard subsets	PRO	0.863	0.949( +8.6% )
Unseen subsets	Image AUROC	0.948	0.971( +2.3% )
Unseen subsets	Pixel AUROC	0.960	0.982( +2.2% )
Unseen subsets	PRO	0.863	0.930( +6.7% )

The larger gains on hard and unseen subsets support the claim that BGAD generalizes beyond the few anomaly types observed during training, rather than merely fitting the known anomalies.

Knowl 10 — Ablation of semi-push-pull versus full pushing

data/table

The semi-push-pull mechanism is tested in the one-class setting against NFAD and a BGAD variant, denoted BGAD†, that removes the BG-SPP loss and uses a conventional contrastive loss with full abnormal-feature pushing. Each entry is image-level AUROC / pixel-level AUROC:

Category	NFAD	BGAD $^{\dagger}$	BGAD
Carpet	0.998/0.994	0.998/0.994	0.999/0.995
Metal nut	0.983/0.966	0.997/0.927	0.998/0.975
Capsule	0.941/0.990	0.868/0.963	0.988/0.991
Screw	0.885/0.989	0.823/0.980	0.947/0.991
Transistor	0.984/0.929	0.933/0.847	0.994/0.942

BGAD† is substantially worse than BGAD on complex categories and can even underperform NFAD. The paper attributes this to full pushing: forcing known abnormalities far from normal features can increase anomaly scores for normal samples and overfit the observed anomaly classes. BG-SPP instead modifies primarily the ambiguous likelihood region, preserving the broader normal and abnormal distributions and improving generalization.

Knowl 11 — Mechanism and evidence for reduced anomaly-class bias
empirical result
BGAD is designed to reduce bias toward the few known anomaly classes through three coupled mechanisms. First, its explicit separating boundary is estimated solely from the normal likelihood distribution, so the decision threshold is not directly displaced by the identities of known anomalies. Second, anomaly decisions remain based on compactness and likelihood under the learned normal distribution rather than on a boundary fitted between normal samples and the particular known anomalies. Third, BG-SPP pushes known anomalies only beyond the abnormal boundary $b_a=b_n-\tau$ ; it does not require them to move arbitrarily far away from normal features.

Experiments support this design. On AITEX and ELPV, BGAD obtains the best image-level AUROC among the compared recent methods, while on BrainMRI and HeadCT it obtains results comparable to the strongest methods. Qualitative feature distributions show that supervised DevNet can become biased toward known anomalies and fail to separate unseen anomalies, whereas BGAD produces more discriminative normal-versus-anomaly features. Its likelihood histograms also show a reduced ambiguous region between normal and abnormal distributions, and its localization maps are reported to be more accurate than those of the unsupervised MSFD baseline and the normal-only BGAD variant.

Coverage note — Detailed per-category MVTecAD tables, full hyperparameter-sensitivity results, and the appendix comparison of alternative synthetic-anomaly generators were omitted to keep the extraction to the ten most significant knowls; their aggregate effects are represented by the reported main results and ablation.

References

1.Samet Akcay, Amir Atapour-Abarghouei, and Toby P. Breckon. Ganomaly: Semi-supervised anomaly detection via adversarial training. In ACCV, page 622–637, 2018. 1, 2
2.Lynton Ardizzone, Carsten Luth, Jakob Kruse, Carsten Rother, and Ullrich Kothe. Guided image generation with conditional invertible neural networks. arXiv preprint arXiv:1907.02392, 2019. 11, 12
3.Liron Bergman, Niv Cohen, and Yedid Hoshen. Deep nearest neighbor anomaly detection. arXiv preprint arXiv:2002.10445, 2020. 1, 2
4.Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad - a comprehensive real-world dataset for unsupervised anomaly detection. In CVPR, 2019. 1, 6, 11
5.Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. In CVPR, 2020. 1, 2, 6
6.Paul Bergmann, Sindy Lowe, Michael Fauser, David Sattlegger, and Carsten Steger. Improving unsupervised defect segmentation by applying structural similarity to autoencoders. In International Conference on Computational Vision Technologies and Applications, 2019. 1, 2, 6
7.Shuo Chen, Gang Niu, Chen Gong, Jun Li, Jian Yang, and Masashi Sugiyama 1 3. Large-margin contrastive learning with distance polarization regularizer. In International Conference on Machine Learning, 2017. 12
8.Niv Cohen and Yedid Hoshen. Sub-image anomaly detection with deep pyramid correspondences. arXiv preprint arXiv:2005.02357v3, 2020. 1, 2
9.Ekin D. Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V. Le. Randaugment: Pratical automated data augmentation with a reduced search space. In CVPR, 2020. 5
10.Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: a patch distribution modeling framework for anomaly detection and localization. In 1st International Workshop on Industrial Machine Learning, 2021. 1, 2, 6, 7
11.Sergiu Deitscha, Vincent Christlein, Stephan Berger, Claudia Buerhop-Lutz, Andreas Maier, Florian Gallwitza, and Christian Riess. Automatic classification of defective photovoltaic module cells in electroluminescence images. In Solar Energy, pages 455–468, 2019. 6, 11
12.Choubo Ding, Guansong Pang, and Chunhua Shen. Catching both gray and black swans: open-set supervised anomaly detection. In CVPR, 2022. 1, 3, 6, 7, 11, 14
13.Laurent Dinh, David Krueger, and Yoshua Bengio. Nice: Non-linear independent components estimation. In International Conference on Learning Representations, 2015. 3
14.Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. In International Conference on Learning Representations, 2017. 2, 3, 11, 12
15.Denis Gudovskiy, Shun Ishizaka, and Kazuki Kozuka. Cflow-ad: Real-time unsupervised anomaly detection with localization via conditional normalizing flows. In IEEE Winter Conference on Application of Computer Vision, 2022. 1, 2, 3, 11
16.Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. In International Conference on Learning Representations, 2019. 1, 2
17.Dan Hendrycks, Mantas Mazeika, Saurav Kadavath, and Dawn Song. Using self-supervised learning can improve model robustness and uncertainty. In Conference and Workshop on Neural Information Processing Systems, 2019. 1, 2, 3
18.Jorn-Henrik Jacobsen, Arnold Smeulders, and Edouard Oyallon. i-revnet: Deep invertible networks. In International Conference on Learning Representations, 2018. 11
19.Diederik P. Kingma and Prafulla Dhariwal. Glow: Generative flow with invertible 1x1 convolutions. In Conference and Workshop on Neural Information Processing Systems, 2019. 11, 12
20.Sungwook Lee, Seunghyun Lee, and Byung Cheol Song. Cfa: Coupled-hypersphere-based feature adaptation for target-oriented anomaly localization. arXiv preprint arXiv:2206.04325, 2022. 6, 7
21.Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In CVPR, 2021. 11
22.Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, 2017. 6
23.Wen Liu, Weixin Luo, Zhengxin Li, Peilin Zhao, and Shenghua Gao1. Margin learning embedding prediction for video anomaly detection with a few anomalies. In International Joint Conference on Artificial Intelligence, 2019. 6
24.Philipp Liznerski, Lukas Ruff, Robert A. Vandermeulen, Billy Joe Franks, Marius Kloft, and Klaus-Robert Muller. Explainable deep one-class classification. In International Conference on Learning Representations, 2021. 1, 2, 3, 6, 7
25.Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. Vt-adl: A vision transformer network for image anomaly detection and localization. arXiv preprint arXiv:2104.10036, 2021. 6, 11
26.Duc Tam Nguyen, Zhongyu Lou, Michael Klar, and Thomas Brox. Anomaly detection with multiple-hypotheses predictions. In International Conference on Machine Learning, 2019. 2
27.Guansong Pang, Choubo Ding, Chunhua Shen, and Anton van den Hengel. Explainable deep few-shot anomaly detection with deviation networks. arXiv preprint arXiv:2108.00462, 2021. 1, 3, 6, 7, 8
28.Guansong Pang, Chunhua Shen, and Anton van den Hengel. Deep anomaly detection with deviation networks. In Proc. ACM SIGKDD Int. Conf. Knowledge Discovery & Data Mining, 2019. 3
29.Stanislav Pidhorskyi, Ranya Almohsen, Donald A. Adjeroh, and Gianfranco Doretto. Generative probabilities novelty detection with adversarial autoencoders. In Conference and Workshop on Neural Information Processing Systems, 2018. 2
30.Tal Reiss, Niv Cohen, Liron Bergman, and Yedid Hoshen. Panda: Adapting pretrained features for anomaly detection and segmentation. In CVPR, 2021. 1
31.Oliver Rippel, Patrick Mertens, and Dorit Merhof. Modeling the distribution of normal data in pre-trained deep features for anomaly detection. arXiv preprint arXiv:2005.14140, 2020. 2
32.Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Scholkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2, 6, 7
33.Marco Rudolph, Bastian Wandt, and Bodo Rosenhahn. Same same but differnet: Semi-supervised defect detection with normalizing flows. In IEEE Winter Conference on Application of Computer Vision, 2021. 1, 2
34.Lukas Ruff, Robert A. Vandermeulen, Billy Joe Franks, Klaus-Robert Muller, and Marius Kloft. Rethinking assumptions in deep anomaly detection. arXiv preprint arXiv:2006.00339, 2020. 1, 2, 3
35.Lukas Ruff, Robert A. Vandermeulen, Nico Gornitz, Lucas Deecke, and Shoaib A. Siddiqui. Deep one-class classification. In International Conference on Machine Learning, 2021. 2, 3
36.Lukas Ruff, Robert A. Vandermeulen, Nico Gornitz, Alexander Binder, Emmanuel Muller, Klaus-Robert Muller, and Marius Kloft. Deep semi-supervised anomaly detection. In International Conference on Learning Representations, 2021. 1, 2, 3
37.Mohammad Sabokrou, Mohammad Khalooei, Mahmood Fathy, and Ehsan Adeli. Adversarially learned one-class classifier for novelty detection. In IEEE Winter Conference on Application of Computer Vision, 2018. 2
38.Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H. Rohban, and Hamid R. Rabiee. Multiresolution knowledge distillation for anomaly detection. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. 1, 2, 6
39.Thomas Schlegl, Philipp Seebock, Sebastian M. Waldstein, Georg Langs, and Ursula Schmidt-Erfurthb. Fast unsupervised anomaly detection with generative adversarial networks. In Medical Image Analysis, 2017. 2
40.Thomas Schlegl, Philipp Seebock, Sebastian M. Waldstein, Ursula Schmidt-Erfurth, and Georg Langs. Unsupervised anomaly detection with generative adversarial networks to guide marker discovery. In International Conference on Information Processing in Medical Imaging, 2017. 2, 6
41.Hannah M. Schluter, Jeremy Tan, Benjamin Hou, and Bernhard Kainz. Natural synthetic anomalies for self-supervised anomaly detection and localization. In ECCV, 2022. 5, 13, 14
42.Bernhard Scholkopf, John C. Plattz, John Shawe-Taylory, Alex J. Smolax, and Robert C. Williamsonx. Estimating the support of a high-dimensional distribution. In Neural Computation, page 1443–1471, 2001. 2
43.Javier Silvestre-Blanes1, Teresa Albero-Albero1, Ignacio Miralles, Ruben Pérez-Llorens, and Jorge Moreno. A public fabric database for defect detection methods and results. In Autex Research Journal, 2019. 6, 11
44.Jihoon Tack, Sangwoo Mo, Jongheon Jeong, and Jinwoo Shin†. Csi: Novelty detection via contrastive learning on distributionally shifted instances. In Conference and Workshop on Neural Information Processing Systems, 2020. 6
45.Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, 2019. 11
46.David M.J. Tax and Robert P.W. Duin. Support vector data description. In Machine Learning, pages 45–66, 2004. 2
47.Guodong Wang, Shumin Han, Errui Ding, and Di Huang. Student-teacher feature pyramid matching for unsupervised anomaly detection. In British Machine Vision Conference, 2021. 1, 2, 6, 7, 8
48.Jie Yang, Yong Shi, and ZhiQuan Qi. Dfr: Deep feature reconstruction for unsupervised anomaly segmentation. arXiv preprint arXiv:2012.0712, 2020. 2
49.Xincheng Yao, Chongyang Zhang, Ruoqi Li, Jun Sun, and Zhenyu Liu. One-for-all: Proposal masked cross-class anomaly detection. In AAAI, 2023. 2
50.Jihun Yi and Sungroh Yoon. Patch svdd: Patch-level svdd for anomaly detection and segmentation. In Asian Conference on Computer Vision, 2021. 1, 2
51.Jiawei Yu, Ye Zheng, Xiang Wang, Wei Li, Yushuang Wu, Rui Zhao, and Liwei Wu. Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows. arXiv preprint arXiv:2111.07677, 2021. 1, 2
52.Vitjan Zavrtanik, Matej Kristan, and Danijel Skocaj. Draem: A discriminatively trained reconstruction embedding for surface anomaly detection. In International Conference on Computational Vision, 2021. 1, 5, 6, 7, 11, 13, 14
53.Houssam Zenati, Manon Romain, Chuan Sheng Foo, Bruno Lecouat, and Vijay Ramaseshan Chandrasekhar. Adversarially learned anomaly detection. In ICDM, pages 727–736, 2018. 2
54.Houssam Zenati, Manon Romain, Chuan Sheng Foo, Bruno Lecouat, and Vijay Ramaseshan Chandrasekhar. Pytorch image models. https://github.com/rwightman/pytorch-image-models, 2019. 11
55.Ev Zisselman and Aviv Tamar. Deep residual flow for out of distribution detection. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2

Citation

MLA

Yao, X., et al. “Explicit Boundary Guided Semi-Push-Pull Contrastive Learning for Supervised Anomaly Detection”. arXiv, 2022, http://arxiv.org/abs/2207.01463v2.

APA

Yao, X., Li, R., Zhang, J., Sun, J., & Zhang, C. (2022). Explicit Boundary Guided Semi-Push-Pull Contrastive Learning for Supervised Anomaly Detection. arXiv. http://arxiv.org/abs/2207.01463v2

Chicago

Yao, X., R. Li, J. Zhang, J. Sun, and C. Zhang. 2022. “Explicit Boundary Guided Semi-Push-Pull Contrastive Learning for Supervised Anomaly Detection”. arXiv. http://arxiv.org/abs/2207.01463v2.

Harvard

Yao, X. et al. (2022) “Explicit Boundary Guided Semi-Push-Pull Contrastive Learning for Supervised Anomaly Detection”, arXiv [Preprint]. Available at: http://arxiv.org/abs/2207.01463v2.

Vancouver

1. Yao X, Li R, Zhang J, Sun J, Zhang C (2022) Explicit Boundary Guided Semi-Push-Pull Contrastive Learning for Supervised Anomaly Detection. arXiv

BibTeX

@article{yao2022explicit,
  title = {Explicit Boundary Guided Semi-Push-Pull Contrastive Learning for Supervised Anomaly Detection},
  author = {Yao, Xincheng and Li, Ruoqi and Zhang, Jing and Sun, Jun and Zhang, Chongyang},
  year = {2022},
  journal = {arXiv},
  url = {http://arxiv.org/abs/2207.01463v2},
  eprint = {2207.01463}
}

Metadata:arXiv

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: IEEE

Metric	NFAD	BGADw/o	BGAD
Image-level AUROC	0.968	0.974	0.9930.0012
Pixel-level AUROC	0.979	0.982	0.9920.0007
PRO	0.946	0.955	0.9760.0006

Explicit Boundary Guided Semi-Push-Pull Contrastive Learning for Supervised Anomaly Detection

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — BGAD for supervised anomaly detection

Knowl 2 — Normal-feature modeling with a conditional normalizing flow

Knowl 3 — Explicit normal and abnormal likelihood boundaries

Knowl 4 — Likelihood-based anomaly scoring

Knowl 5 — Boundary-guided semi-push-pull optimization

Knowl 6 — RandAugment-based pseudo-anomaly generation

Knowl 7 — Experimental protocol for known and unseen anomalies

Knowl 8 — MVTecAD and BTAD performance

Knowl 9 — Generalization to hard and unseen anomaly subsets

Knowl 10 — Ablation of semi-push-pull versus full pushing

Knowl 11 — Mechanism and evidence for reduced anomaly-class bias

References

Citation

BibTeX

Access the Paper