CoZAD: Zero-Shot Anomaly Detection Framework

• CoZAD is a zero-shot anomaly detection framework that integrates soft confident learning, MAML-style meta-learning, and contrastive feature representation to detect anomalies from nominal-only training data.
• It employs a modified SimpleNet with a frozen WideResNet-50 backbone and Gaussian noise perturbations to synthesize negative examples for both image-level detection and pixel-level localization.
• The framework achieves state-of-the-art performance on multiple industrial and medical benchmarks by preserving uncertain samples with graded weights and effectively modeling data and model uncertainties.

Searching arXiv for CoZAD and closely related zero-shot anomaly detection work. CoZAD denotes Contrastive Learning-Guided Confident Meta-learning for Zero Shot Anomaly Detection, a zero-shot anomaly detection framework for industrial and medical imaging that integrates soft confident learning, MAML-style meta-learning, and contrastive feature representation. Its operating assumption is that training data contain only normal / nominal samples, while anomalies are absent during training, yet the deployed model must support both image-level anomaly detection and pixel-level anomaly localization under domain shift. The framework is explicitly positioned against prompt-heavy vision-language methods and ensemble-based designs: it weights all training samples rather than discarding uncertain ones, models both data uncertainty and model uncertainty, and builds a discriminative normality representation without relying on vision-language alignment or multi-model ensembles (Aqeel et al., 25 Aug 2025).

1. Problem setting and conceptual position

CoZAD addresses zero-shot anomaly detection (ZSAD) in settings where labels are expensive or unavailable, target domains may differ from source domains, and anomalous samples are not available for training. The motivating observation is that nominal datasets are often not perfectly clean: they may contain rare-but-valid normal variations, borderline samples, or samples that resemble anomalies. In that setting, hard sample filtering can remove informative boundary cases rather than improving robustness (Aqeel et al., 25 Aug 2025).

The framework therefore treats uncertainty as a first-class training signal. Its central claim is not merely that uncertain samples exist, but that they should remain in optimization with graded importance. This distinguishes CoZAD from traditional confident learning formulations that discard suspicious samples. It also distinguishes the method from prior ZSAD approaches such as WinCLIP, AnomalyCLIP, APRIL-GAN, CLIP-AD, AdaCLIP, and Bayes-PFL, which are described as depending more heavily on vision-language alignment, prompt engineering, or multiple foundation models (Aqeel et al., 25 Aug 2025).

A common misconception is that “zero-shot” in this context implies text-driven or prompt-driven transfer. In CoZAD, zero-shot refers instead to training on nominal data only and transferring across domains without anomalous examples. The framework’s mechanism is confidence-guided discriminative learning, not prompt alignment.

2. Architecture and anomaly scoring

CoZAD is built on a modified SimpleNet architecture with a frozen WideResNet-50 backbone, a feature adaptor $G_{\theta_1}$, a discriminator $D_{\theta_2}$, synthetic anomalies generated by Gaussian noise injection, soft confident learning, MAML meta-learning, and contrastive learning. Features are extracted from the 2nd and 3rd intermediate layers of WideResNet-50, giving a feature dimension of 1536 (Aqeel et al., 25 Aug 2025).

The feature pipeline is defined by

$q_i = G_{\theta_1}(o_i) = G_{\theta_1}(F_*(x_i)),$

where $F_*$ is the frozen feature extractor. Because no real anomalies are available during training, the method synthesizes negative examples by perturbing adapted normal features: $$q_i^- = q_i + \epsilon,\qquad \epsilon \sim \mathcal{N}(0,\sigma^2),$$ with $\sigma = 0.015$. The discriminator is a small MLP comprising a linear layer, batch normalization, LeakyReLU with slope $0.2$, and a final linear layer (Aqeel et al., 25 Aug 2025).

The anomaly detection loss is margin-based: $$\mathcal{L}_{AD}(x|\theta) = \max(0, th_+ - D_{\theta_2}(q_i)) + \max(0, D_{\theta_2}(q_i^-) + th_-),$$ with $th_+ = th_- = 0.5$. The intended separation is explicit: normal features should score above $th_+$, while synthetic anomalies should score below $D_{\theta_2}$0 or otherwise be separated by the margin. At inference, the synthetic anomaly generator is removed, and the anomaly score becomes

$D_{\theta_2}$1

Higher $D_{\theta_2}$2 indicates greater abnormality (Aqeel et al., 25 Aug 2025).

This formulation makes the representation pipeline asymmetric in a useful way: training uses synthetic perturbations to shape a decision boundary, whereas inference uses only the backbone, adaptor, and discriminator. A plausible implication is that the deployment graph remains comparatively lightweight because the auxiliary anomaly generator is not needed at test time.

3. Soft confident learning and uncertainty-aware meta-learning

The core of CoZAD is its soft confident learning strategy. Rather than deleting uncertain samples, it assigns each sample $D_{\theta_2}$3 a confidence weight $D_{\theta_2}$4 derived from its anomaly score: $D_{\theta_2}$5 If $D_{\theta_2}$6 is small, the sample is treated as highly normal and $D_{\theta_2}$7; if $D_{\theta_2}$8 is large, the weight decreases, but the sample remains in the training set. The confidence-weighted anomaly objective is

$D_{\theta_2}$9

This preserves boundary information while emphasizing prototypical normal patterns (Aqeel et al., 25 Aug 2025).

To set the reference threshold $q_i = G_{\theta_1}(o_i) = G_{\theta_1}(F_*(x_i)),$0, CoZAD uses Interquartile Range (IQR) thresholding on anomaly scores: $q_i = G_{\theta_1}(o_i) = G_{\theta_1}(F_*(x_i)),$1 where $q_i = G_{\theta_1}(o_i) = G_{\theta_1}(F_*(x_i)),$2 and $q_i = G_{\theta_1}(o_i) = G_{\theta_1}(F_*(x_i)),$3 are the first and third quartiles. The reported best sensitivity setting is $q_i = G_{\theta_1}(o_i) = G_{\theta_1}(F_*(x_i)),$4. Scores above $q_i = G_{\theta_1}(o_i) = G_{\theta_1}(F_*(x_i)),$5 are interpreted as highly uncertain or anomalous-looking, but the threshold is not used as a strict discard rule. This matters because the framework’s design objective is not outlier excision; it is soft normalization of sample influence (Aqeel et al., 25 Aug 2025).

CoZAD also models model uncertainty using the covariance structure of training and validation losses inside a Model-Agnostic Meta-Learning scheme. The determinant $q_i = G_{\theta_1}(o_i) = G_{\theta_1}(F_*(x_i)),$6 of the train/validation loss covariance matrix acts as a scalar uncertainty proxy, and regularization strength is adapted as

$q_i = G_{\theta_1}(o_i) = G_{\theta_1}(F_*(x_i)),$7

The unified soft confident learning objective is

$q_i = G_{\theta_1}(o_i) = G_{\theta_1}(F_*(x_i)),$8

Within MAML, each task uses an inner-loop update

$q_i = G_{\theta_1}(o_i) = G_{\theta_1}(F_*(x_i)),$9

followed by an outer-loop meta-update

$F_*$0

The meta-objective re-evaluates both sample weights and uncertainty regularization after task adaptation, which is the mechanism by which CoZAD attempts to improve cross-domain generalization (Aqeel et al., 25 Aug 2025).

A key empirical clarification is that the two uncertainty terms are not equivalent. The ablations report that data uncertainty weighting matters more than model uncertainty regularization, and that soft weighting beats hard filtering. This is one of the main substantive claims of the method rather than an implementation detail.

4. Contrastive representation learning

The contrastive component is designed to make normal patterns form compact clusters and to separate them from non-matching patterns. Given normal feature embeddings $F_*$1, CoZAD constructs an augmented view

$F_*$2

and uses cosine similarity

$F_*$3

The basic contrastive loss is

$F_*$4

This explicitly pulls $F_*$5 toward its noisy augmentation while pushing it away from other embeddings (Aqeel et al., 25 Aug 2025).

The framework extends this with nearest-neighbor positive mining, treating the $F_*$6-nearest neighbors as additional positives in order to capture the manifold of normality more effectively. The reported best setting is $F_*$7. To keep memory usage constant, the contrastive loss is also computed in chunks: $F_*$8 The total training objective is then

$F_*$9

with best reported $$q_i^- = q_i + \epsilon,\qquad \epsilon \sim \mathcal{N}(0,\sigma^2),$$0 (Aqeel et al., 25 Aug 2025).

In ablation, the paper attributes the largest and most consistent gains to Confident Learning, identifies Meta-Learning as especially important for cross-domain generalization, and reports that Contrastive Learning contributes smaller but stable improvements. On MVTec-AD the progression is 89.7 $$q_i^- = q_i + \epsilon,\qquad \epsilon \sim \mathcal{N}(0,\sigma^2),$$1 91.8 $$q_i^- = q_i + \epsilon,\qquad \epsilon \sim \mathcal{N}(0,\sigma^2),$$2 92.6 $$q_i^- = q_i + \epsilon,\qquad \epsilon \sim \mathcal{N}(0,\sigma^2),$$3 92.9, and on VisA it is 80.0 $$q_i^- = q_i + \epsilon,\qquad \epsilon \sim \mathcal{N}(0,\sigma^2),$$4 81.8 $$q_i^- = q_i + \epsilon,\qquad \epsilon \sim \mathcal{N}(0,\sigma^2),$$5 84.9 $$q_i^- = q_i + \epsilon,\qquad \epsilon \sim \mathcal{N}(0,\sigma^2),$$6 85.3, corresponding to Baseline SimpleNet, then adding Confident Learning, then Meta-Learning, then Contrastive Learning (Aqeel et al., 25 Aug 2025).

5. Training protocol, benchmarks, and quantitative performance

CoZAD is evaluated on 10 datasets spanning industrial and medical anomaly detection. The industrial benchmarks are MVTec-AD, VisA, BTAD, KSDD2, RSDD, DAGM, and DTD-Synthetic. The medical benchmarks are HeadCT, BrainMRI, and Br35H. The training protocol uses VisA as an auxiliary training dataset; the model is trained on one domain and evaluated on the others without domain-specific retraining, while VisA is also evaluated via models fine-tuned on MVTec-AD. The reported metrics are I-AUROC for image-level anomaly detection and P-AUROC for pixel-level localization (Aqeel et al., 25 Aug 2025).

Dataset	I-AUROC	P-AUROC
MVTec-AD	92.9	96.3
VisA	85.3	95.9
BTAD	97.2	—
KSDD2	94.2	97.8
RSDD	94.7	99.7
DAGM	97.9	99.5
DTD-Synthetic	99.2	97.9
HeadCT	96.8	94.3
BrainMRI	98.7	—
Br35H	99.3	79.9

The headline result is that CoZAD reports state-of-the-art performance, outperforming existing methods on 6 out of 7 industrial benchmarks and all 3 medical datasets in the image-level table. Particularly emphasized are 99.2% I-AUROC on DTD-Synthetic, 97.2% on BTAD, and 92.9% on MVTec-AD. For localization, the paper highlights 96.3 P-AUROC on MVTec-AD, 95.9 on VisA, 99.7 on RSDD, 99.5 on DAGM, 97.9 on DTD-Synthetic, and 94.3 on HeadCT. One caveat is explicit: on KSDD2, the reported 97.8 P-AUROC remains below some competing methods above 99 (Aqeel et al., 25 Aug 2025).

The confidence-learning ablation further sharpens the interpretation. For the paired MVTec-AD / VisA results, the paper reports Full CoZAD: 92.9 / 85.3, w/o Data Uncertainty: 91.4 / 83.7, w/o Model Uncertainty: 92.1 / 84.5, w/o Confident Learning: 90.8 / 81.2, and Hard CL: 91.6 / 83.1. This directly supports the claim that data uncertainty weighting contributes more than model uncertainty regularization and that soft weighting is preferable to hard filtering (Aqeel et al., 25 Aug 2025).

6. Limitations, implications, and relation to adjacent detection problems

The method’s stated advantages are straightforward: no anomaly labels are required; there is no heavy reliance on prompt engineering; no ensemble of foundation models is needed; soft confidence weighting preserves useful boundary samples; uncertainty is modeled explicitly; contrastive learning yields compact normal clusters; and the framework performs strongly in both industrial and medical domains while remaining suitable for resource-constrained deployment (Aqeel et al., 25 Aug 2025).

The paper also states several limitations. IQR thresholding may not adapt optimally to all distributions. Computational efficiency still matters for high-resolution real-time deployment. Future work is described in terms of explainability, few-shot adaptation, video / 3D extension, federated learning, and better threshold learning (Aqeel et al., 25 Aug 2025). These limitations are structurally important because they show where the framework’s current robustness depends on simple, unsupervised heuristics rather than task-specific calibration.

CoZAD is also best understood in contrast to nearby “co-” detection literatures that solve different problems. ConceptCoSOD frames co-salient object detection as an (image-text)-to-image problem and learns a shared textual concept token to guide segmentation across related images (Zhu et al., 2024). SCoSPARC is a self-supervised CoSOD method based on feature correspondence at patch and region levels, with confidence-based adaptive thresholding and region-level consensus filtering (Chakraborty et al., 2024). Those methods target group-wise common-object segmentation; CoZAD targets nominal-only anomaly modeling for image-level and pixel-level anomaly prediction. The shared vocabulary of “detection” can obscure this difference, but the underlying supervision assumptions, objectives, and inference outputs are distinct.

In that sense, CoZAD’s main contribution is not a new anomaly score alone. It is a composite training strategy in which soft confidence weighting, covariance-driven uncertainty regularization, MAML-style adaptation, and contrastive normality shaping are made to operate as a single zero-shot anomaly detection system. The practical message is specific: when nominal data are heterogeneous and imperfectly clean, retaining uncertain samples with calibrated weights can be more effective than removing them, particularly when cross-domain generalization is required (Aqeel et al., 25 Aug 2025).