학술

Online unsupervised industrial anomaly detection with adaptive resonance patch prototype memory and multi-level views

PLOS ONE
CC BY
이 매체는 공공·자유 라이선스로 본문을 직접 표시합니다.

Figures
Abstract
Unsupervised industrial anomaly detection is widely used because defective samples are scarce while normal data are abundant. Many recent methods follow a feature-embedding pipeline that models patch-level deep features, but they often (i) introduce cross-layer interference when multi-level features are fused and (ii) risk losing fine-grained information when reference features are aggressively reduced. This paper proposes ART-IAD, an online patch-based framework that treats multi-level CNN features as independent views and models each view with an incremental Fuzzy Adaptive Resonance Theory (Fuzzy ART) prototype memory. For each ART prototype, sufficient statistics, including the mean and covariance, are maintained, and Mahalanobis-based anomaly scores are computed, enabling anomaly detection and localization without storing all patch embeddings. On MVTec AD, ART-IAD achieves an average image-level AUROC of 99.4% and pixel-level AUROC of 98.2%, slightly outperforming the strong baseline PatchCore by +0.3% in image-level AUROC and +0.1% in pixel-level AUROC. These results indicate that ART-based online prototype modeling and multi-view separation can provide accurate detection and localization with a compact reference memory.
Citation: Zhu Y, Li P, Zhang Q, Jin G (2026) Online unsupervised industrial anomaly detection with adaptive resonance patch prototype memory and multi-level views. PLoS One 21(11): e0362092. https://doi.org/10.1371/journal.pone.0362092
Editor: Panos Liatsis, Khalifa University of Science and Technology, UNITED ARAB EMIRATES
Received: January 13, 2026; Accepted: October 20, 2026; Published: November 20, 2026
Copyright: © 2026 Zhu et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: VisA is publicly available through the official Amazon Science project and the AWS Open Data Registry under the CC BY 4.0 license. For MVTec AD, the non-commercial restriction is imposed by MVTec Software GmbH through the CC BY-NC-SA 4.0 license. As users of this third-party dataset, we cannot waive this restriction or provide the dataset under an unrestricted license. The independent institutional contact for this dataset to which dataset access requests and licensing inquiries may be sent is re-request@mvtec.com.
Funding: This work was partially supported by the Anhui Research Center of Generic Technology in Photovoltaic Industry Open Project (FYKFKT24075)(URL: https://www.fynu.edu.cn/aprcp/info/1048/1180.htm), the Natural Science Foundation of the Jiangsu Higher Education Institutions of China (24KJB520005)(URL: https://news.sciencenet.cn/htmlnews/2024/6/524989.shtm), the Jiangsu Province Education Science Planning Project (B-b/2024/01/166)(URL: https://ghb.jsies.cn/content/W8nnFAGrT5N4dHR585baeAcQ4jpC8YaR.html), the Jiangsu Province Social Education Planning Project (JSS-C-2023035)(URL: https://www.jsou.edu.cn/2023/1117/c2913a166699/page.htm), and the General Project of Philosophy and Social Science Research in Colleges and Universities in Jiangsu Province (2024SJYB0571)(URL: https://www.sju.edu.cn/kjydw/_t8/77/df/c760a96223/page.htm). There was no additional external funding received for this study. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Unsupervised industrial anomaly detection aims to identify defective samples that deviate from normal patterns when only normal data are available during training. It is of particular importance in manufacturing, where defective samples are scarce, anomalies vary from subtle surface changes (e.g., minor scratches) to severe structural defects (e.g., missing components) [1–9], and some defect types may be completely unavailable before deployment. Consequently, anomaly detection models are commonly formulated as one-class learning systems trained solely on normal samples and expected to identify anomalous samples at test time [10]. With the recent progress of Convolutional Neural Networks (CNNs), deep representations have significantly advanced anomaly detection [11–13], enabling both image-level detection and pixel-level localization.
In recent years, numerous unsupervised anomaly detection (UAD) methods have been proposed [14–28]. Reconstruction-based approaches [14,20,22,25,29,30] (e.g., AE [31–35], VAE [36–38], and GAN [18,39,40]) detect anomalies via reconstruction error or structural similarity. However, they often suffer from (i) loss of fine details during encoding and reconstruction, which may cause false alarms on normal regions, and (ii) overly strong generalization that can also reconstruct anomalous images well in some datasets, leading to missed detections. These limitations motivate methods that directly compare deep representations rather than reconstructed pixels. Diffusion-based reconstruction has also been investigated for multi-class anomaly detection; DiAD [41] uses a semantic-guided diffusion framework to reconstruct anomalous regions while preserving category-level semantics, and reports results on both MVTec AD and VisA.
Feature-embedding-based methods [26,42–44] address the above issues by extracting deep feature embeddings and measuring the distance between test features and normal references. For image-level detection, global representations are used [45–48], while for localization the image is divided into patches and each patch is embedded [1,10,49]. References are typically implemented either as a memory bank of normal embeddings [10] or as a Gaussian model [1,46]. Despite strong performance, two challenges remain critical for industrial inspection: (1) Multi-level feature fusion is commonly implemented by upsampling deep-layer feature maps and concatenating them with shallow features; this can introduce cross-layer interference and redundant responses, which may weaken the distinctiveness of low-level features that are important for subtle defect localization. (2) Many reference representations are designed for offline settings; when normal data arrive sequentially (e.g., production line updates), it becomes non-trivial to update the reference representation incrementally without storing a large number of patch embeddings.
Synthesis-based methods [2,50,51] introduce synthetic anomalies to enrich training. Image-level synthesis [50] can provide detailed textures but often lacks realism and diversity, while feature-level synthesis [2,51] is efficient yet may be insufficiently directed and can fail on near-distribution anomalies. Moreover, unified models such as UniAD [29] aim to handle multiple classes, but real-world industrial environments often require continual updates as new products are introduced. Such unified training is prone to catastrophic forgetting and increased computational complexity, which can hinder deployment in dynamic production lines. Recent work has also explored foundation-model-based anomaly synthesis. Wu et al. [52] generate realistic anomalies using GPT, SAM, and Stable Diffusion, and introduce wavelet-domain attention to improve anomaly sensitivity.
Adaptive Resonance Theory (ART) [53] is known to mitigate catastrophic forgetting, especially when previous data cannot be retained due to privacy or limited resources. Through competitive matching and vigilance-controlled learning, ART updates an existing prototype when an incoming normal pattern is sufficiently matched and creates a new prototype otherwise. This stability–plasticity mechanism preserves previously learned normal patterns while accommodating new variations, making ART particularly suitable for incrementally updating the normal reference model in evolving industrial environments. In industrial scenarios, collecting large-scale anomalous data is difficult due to high production success rates and strict privacy constraints, making online updating on normal data a practical requirement. However, the potential of ART for unsupervised industrial anomaly detection and localization remains underexplored.
To address the above gaps, we propose a novel online patch-based framework for anomaly detection and localization that (i) avoids harmful cross-layer interference by separating multi-level features into independent views and (ii) maintains a compact, incrementally updatable reference memory via ART prototypes with distribution statistics. Following prior work [45–48], we represent each image as a collection of patches and extract patch embeddings using a pre-trained CNN, but we introduce the following key innovations. The main contributions are:
- 1) Online ART prototype memory for patch embeddings. We integrate Adaptive Resonance Theory (ART) to organize patch-level features and update prototypes incrementally, enabling continual learning from sequentially arriving normal samples while mitigating catastrophic forgetting [53]. Each ART node maintains distribution statistics (mean and covariance), which allows compact reference modeling without storing all patch embeddings.
- 2) Multi-view separation to reduce cross-layer interference. Instead of directly upsampling and concatenating multi-level feature maps, we treat features from different CNN layers as distinct views and process them independently before aggregation. This reduces redundancy introduced by upsampling and preserves layer-specific semantics for more reliable anomaly localization.
- 3) Comprehensive validation on industrial benchmarks. Extensive experiments on MVTec AD and VisA demonstrate competitive anomaly detection and localization performance compared with state-of-the-art approaches. Additional efficiency and sensitivity analyses quantify the balance between memory consumption and computational efficiency of the proposed ART memory under online deployment.
Related works
Reconstruction-based approaches
Reconstruction-based methods operate on the assumption that anomalous regions in an image cannot be accurately reconstructed, as they are not present in the training data. These methods typically use the Structural Similarity Index (SSIM) loss function during training, with anomaly detection based on the pixel-level differences between input and reconstructed images. However, if anomalies share similar patterns with normal data (e.g., local edges) or if the decoder is sufficiently powerful to reconstruct certain anomalous encodings, anomalies may still be correctly reconstructed [54]. This limitation motivates approaches (including our work) that avoid relying purely on reconstruction fidelity to distinguish normal from abnormal patterns.
Reconstruction-based methods are subject to further categorization, which depends on the type of reconstruction network utilized, including Autoencoder-based approaches [51,55–57], Generative Adversarial Networks (GAN) [18,58], Vision Transformer (ViT)-based methods [59–61], and diffusion model-based methods [41,59,62]. Despite their effectiveness in some applications, reconstruction methods face notable limitations, such as insufficient discriminative power for anomalies that were not seen during training, leading to erroneous reconstructions. Additionally, these methods often follow an end-to-end learning paradigm, which, while intuitive, suffers from limited interpretability and requires complex network design, external constraints, and specific training strategies to enhance performance. As a result, while reconstruction-based methods are theoretically simple, they encounter significant challenges in real-world industrial applications, which further motivates our focus on a more deployable and compact anomaly modeling strategy.
Feature-embedding-based approaches
Feature-embedding-based anomaly detection methods have become a strong paradigm for industrial inspection. These methods detect anomalies by comparing test features with normal reference features, using either probabilistic models such as Gaussian Density Estimation (GDE) and Normalized Flow (NF) [1,15,16,23,43,46], or non-parametric models such as K-nearest neighbors(KNN) and kernel density estimation(KDE) [10,49,63,64]. Although effective, such reference-based designs often introduce a trade-off between detection accuracy and deployability, especially in terms of memory usage and inference-time computation. This motivates compact and incrementally updatable reference representations for patch-level anomaly detection.
Semantic Pyramid Anomaly Detection (SPADE) [49] is a notable method that uses pre-trained models for feature extraction and has shown excellent performance. However, SPADE requires storing the entire set of normal embedding vectors, leading to significant storage overhead. Additionally, when KNN is used at test time against the full training set, substantial computation is introduced, which in practice limits industrial deployment. To ease this, PaDiM [1] proposed block-wise distribution modeling, in which it is assumed that image patches at the same spatial location share one distribution, for which the mean and variance are computed. It is during the testing phase that the anomaly score for each patch is derived, utilizing the Mahalanobis distance to the mean and variance of the corresponding normal location. This method significantly reduces computational overhead, as it avoids storing all normal embeddings or searching for nearest-neighbor reference vectors. However, PaDiM relies on a spatially fixed distribution assumption, which can be sensitive to spatial misalignment. For instance, in the case of screws, images with screws placed horizontally and vertically cannot be assumed to have patches at the same position belonging to the same distribution; this sensitivity to layout variation motivates methods (including ours) that target more robust normal modeling under spatial changes.
PatchCore [10] is developed as an extension of SPADE, where a greedy coreset mechanism is utilized to retain the most representative training data, thereby significantly reducing the number of training samples required. However, PatchCore’s greedy coreset strategy improves compactness, but may discard some normal variations and thus affect coverage in deployment. It is through a Gaussian Mixture Model (GMM) that normal features are fitted in SMCC [65], a method grounded in self-updating memory and central clustering, when the backbone structure is jointly learned and the memory bank updated. Recent work also explores realistic synthetic defects for anomaly detection. Wu et al. [52] combine foundation-model-based anomaly synthesis with wavelet-domain attention and report improved results on MVTec AD and VisA. In contrast, our method models normal patch distributions online without synthetic anomalies.
Shortcut learning remains a key challenge in anomaly detection. When objectives such as reconstruction or teacher–student regression allow near-trivial mappings, models may rely on spurious cues or reproduce normal appearance statistics, causing anomaly overgeneralization: abnormal samples are also well explained and assigned low anomaly scores. Recent studies mitigate this issue by disrupting feature copying via contextual masking distillation [66], enhancing discriminative deviations with uncertainty-inspired inter-sample differences [67], or imposing multi-frequency constraints to suppress overgeneralized reconstruction [68]. More recently, frequency-decoupled anomaly detection has shown that the low-frequency bias of conventional reconstruction models may cause a spectral mismatch and the loss of anomaly-relevant high-frequency information [69]. From a complementary multi-view perspective, naive fusion may repeatedly amplify redundant view-invariant information while suppressing anomaly-relevant view-specific cues, leading to representation homogenization and degraded anomaly sensitivity [70]. Although developed mainly for network traffic anomaly detection, these insights are also relevant to industrial visual anomaly detection, where models can overfit superficial texture and color statistics, learn overly broad normal representations, or weaken subtle anomaly cues through inappropriate multi-view fusion.
Although feature-embedding-based anomaly detection has shown promise, several limitations are still observed. First, feature interference is often encountered when multi-layer features are combined, in which redundancy and noise are introduced, and it is by simplistic fusion that detection performance tends to be hindered. Second, when pre-trained models from large-scale datasets such as ImageNet are used, a domain mismatch may be induced, especially in industrial settings, where adaptation to different anomaly types can be challenging. Finally, memory-based methods are heavily dependent on accurate representations of normal features, and in terms of computational efficiency, particularly on edge devices, it is the storage and querying of large feature dictionaries that can incur high computational cost. These limitations motivate our work to improve anomaly separability while reducing memory overhead to better support practical industrial deployment. Beyond industrial vision, recent security-oriented collaborative learning studies [71] also highlight system-level requirements such as efficiency, privacy, and explainability, which are aligned with our goal of compact and deployable anomaly modeling, although our work does not consider federated training or security-threat modeling.
Method
This section presents the ART-IAD framework, in which two key innovations are introduced to address the challenges identified in Section. First, we leverage Adaptive Resonance Theory (ART) to organize patch embeddings into adaptive prototype nodes, thereby compactly modeling local normal patterns. In this way, the number of reference vectors needed to represent the dataset is significantly reduced, while critical information is retained by estimating the mean and covariance statistics of the samples projected onto each ART node. Second, to mitigate interference between features from different layers, we treat features extracted from different CNN layers as independent semantic views. These views are processed separately rather than being naively concatenated, and their outputs are then aggregated to obtain the final anomaly score. Together, these innovations enhance computational efficiency while improving sensitivity to subtle anomalies.
As shown in Fig 1, the ART-IAD framework consists of three parts: feature extraction, ART-based modeling of patch distributions, and anomaly scoring. These modules are described in turn below.
The original input image is processed by a ResNet backbone to extract feature maps from the first three convolutional layers, which serve as multi-view representations. These features are then independently modeled using three separate Fuzzy ART networks.
Feature extraction
To obtain informative visual features, we adopt a transfer learning setup that leverages pre-trained models to improve target-domain performance while requiring little labeled data. In particular, we use the ResNet architecture, which has been shown to generalize well and extract useful representations across diverse vision tasks. Pre-trained ImageNet weights are widely available, and deployment is usually straightforward, making this choice practical and effective.
We follow a patch-based strategy in which input images are divided into small, non-overlapping patches. This design is motivated by three points: first, anomalies tend to appear as small localized patterns; second, patch-wise representations are helpful for precise localization; and third, patching helps with misalignment, which is especially evident in datasets like screw that exhibit rotational variance. By comparing local patches, alignment across samples can be improved without requiring global structural correspondence.
To capture both global and local cues, embedding vectors that carry information from different semantic levels are treated as distinct views. During training, each image is split into an grid of patches. For every normal image, its patches are assigned feature embeddings taken from the pre-trained model’s feature map, so that each patch at position is associated with three embedding vectors , where t indexes the view. Following PatchCore, average pooling is applied to the feature maps to expand the receptive field and improve semantic robustness.
Modeling the distribution of image patches with ART
In anomaly detection, it is the distribution of image patches that needs to be modeled so that normal and anomalous samples can be separated with some robustness. To this end, Fuzzy Adaptive Resonance Theory (Fuzzy ART) [72] is adopted to organize patch embeddings, which are obtained from the features. Through incremental learning combined with a competitive mechanism, Fuzzy ART can be adapted dynamically to data shifts and is less prone to catastrophic forgetting, which makes it a suitable choice for feature modeling in image processing.
Overview of the Fuzzy ART algorithm.
Fuzzy ART is an unsupervised method in which input data are mapped to categories by adaptively adjusting each category’s feature representation. In practice, the model operates through a sequence of steps: matching an input vector to existing nodes, selecting a category via competition, applying the vigilance test, and updating the selected node when resonance occurs.
For an M-dimensional feature vector x from an image patch, with entries , we first compute a match between the input and every ART node. For the input vector x, its similarity to node j is evaluated by the choice function, after which the best-matching node is selected for further updating.
Here, is the weight vector of node j, and is a user-defined choice parameter that controls the sensitivity of selection. The symbol ∧ denotes the fuzzy AND operation, which is used to compute the match between the input and a node.
where the norm is defined as
This operation is used to measure how similar the input vector is to a given node. By computing the match degree T(x), the relation between the input and each node can be assessed.
On the basis of the computed match degree T(x), a competitive process is carried out so that the node with the highest value is selected. In particular, node J is chosen when its value of is the largest among all candidates.
where N denotes the total number of nodes. A resonance is said to occur when the chosen category’s match satisfies the vigilance test, that is, when
If this inequality is not met, a reset is triggered.
When resonance is confirmed, it is the winning node’s weight that is updated as
where is the learning rate that controls the magnitude of the weight update. With this update, the node’s weight vector is gradually pulled toward the input features, allowing the input to be represented more faithfully. For each incoming patch embedding x, the existing ART nodes are ranked by the choice function and examined sequentially. The highest-ranked node J is updated using the rule above if it satisfies the vigilance criterion. Otherwise, it is reset and the next candidate is tested. If no existing node passes the vigilance test, a new prototype is created for x. This match–reset–search procedure is performed independently for each semantic view, enabling the prototype memory to incorporate new normal patterns incrementally.
Application of Fuzzy ART in modeling the distribution of image patches.
During training, patch-level feature vectors from normal images are fed into Fuzzy ART sequentially. For each incoming vector, the match–reset–search procedure either updates a resonant node or creates a new node when no existing node satisfies the vigilance criterion. The resulting ART nodes therefore represent adaptively formed regions of the normal embedding space. After prototype learning, the normal patch embeddings are assigned to their corresponding ART nodes to estimate the node-specific distribution statistics used for anomaly scoring.
To better describe how patches are distributed and to improve anomaly detection, a covariance matrix is computed for each node i, which is used to capture the variance of the input vectors that are associated with that node.
(1)Here, denotes the number of input vectors projected to node i, denotes the j-th feature vector assigned to that node, and is the centroid of node i. The term is added as a regularization, in which I stands for the identity matrix. This term ensures that remains full rank and invertible, improving numerical stability and making the computation more robust.
With this modeling, each ART node captures not only the prototype features of normal patches but also the variability of their distribution. As a result, more discriminative anomaly scores can be produced in the subsequent stage. The overall workflow of the algorithm is shown in Fig 1.
Anomaly scoring
Since the features extracted from different layers of the backbone are treated as independent views, the proposed method yields three trained Fuzzy ART networks, denoted as ART1, ART2, and ART3. Given a test image , we partition it into non-overlapping patches. Let represent the feature embedding of the i-th patch extracted from the t-th view, where and .
To quantify the abnormality of a patch, we compute the Mahalanobis distance between the embedding and its winner ART node in . The anomaly score of patch i in view t is computed as:
(2)Here, is the centroid of the winner ART node and is the corresponding covariance matrix estimated from the training data. Unlike Euclidean distance, which assigns equal importance to all feature dimensions and implicitly assumes an isotropic local distribution, Mahalanobis distance normalizes feature deviations according to the intra-prototype variance and accounts for correlations between feature dimensions. It therefore assigns larger scores to deviations along stable directions of a normal prototype while reducing the influence of variations that are common within that prototype, providing a locally adaptive anomaly measure.
After the Mahalanobis distance maps for all three views are computed, the maps from the second and third layers are upsampled to match the spatial size of the first view using bilinear interpolation. The final anomaly score for patch i is computed by summing the distances across all three views:
(3)This summation can be interpreted as an equal-weight late-fusion strategy over multiple semantic views. Since the three views are extracted from different CNN layers, they capture patch information at different semantic levels and receptive fields. Meanwhile, each view-specific score is computed using the Mahalanobis distance to the corresponding ART-modeled normal distribution, and therefore all scores share the same meaning, namely the degree of deviation from normal patch statistics. Based on this consistency, summing the scores provides a simple way to aggregate complementary anomaly evidence from different feature levels. Compared with early feature concatenation, this late-fusion design preserves the independence of different views and avoids mixing heterogeneous layer features before distribution modeling. Compared with learnable or tuned fusion strategies, it introduces no additional parameters and does not require anomalous validation samples, which is desirable for the unsupervised anomaly detection setting.
Finally, the patch-level anomaly scores are upsampled to the size of the original image using bilinear interpolation to generate the pixel-level anomaly map. The final score can also be obtained using a weighted combination of the three views:
(4)where denotes the weight of the t-th view. Although adjusting can improve performance in specific scenarios, we set all weights to 1 for simplicity in our experiments.
Algorithmic steps (training and inference)
The ART-IAD pipeline is composed of the following steps: (1) Multi-view feature extraction. For each image, we extract patch-wise embeddings from the first three ResNet convolutional blocks. Each layer is treated as an independent semantic view to avoid cross-layer feature interference. (2) Per-view ART modeling. For each view , all normal patch embeddings are streamed to a dedicated Fuzzy ART network . For each incoming patch embedding, the choice function is computed against all existing ART nodes, a vigilance test is applied, and the winning node is updated; if no node passes vigilance, a new node is created. (3) Node-wise distribution estimation. After the ART nodes are formed, we estimate the centroid and covariance of each node (Eq. (1)), enabling distance-based scoring with local uncertainty modeling. (4) Patch scoring at test time. Given a test image, we extract three-view patch embeddings and, in each view, compute the Mahalanobis distance (Eq. (2)) to the corresponding winning ART node. (5) Multi-view fusion and localization. The three patch-level distance maps are spatially aligned by upsampling higher-level views and then aggregated (Eq. (3)) to obtain the final patch-level anomaly map, which is further upsampled to pixel resolution for localization.
The complete scoring procedure is described in Algorithm 1.
Discussion: On the identity shortcut issue. The identity shortcut issue may cause anomaly overgeneralization in reconstruction- or distillation-based anomaly detection. ART-IAD alleviates this risk in several ways. First, it does not learn an explicit reconstruction or teacher–student regression pathway, but organizes fixed CNN patch embeddings into an incremental prototype memory, thereby avoiding a direct identity mapping. Second, the vigilance criterion in Fuzzy ART limits the generalization range of each prototype, since a patch can match an existing node only when its match degree is sufficiently high; otherwise, a new node is created. Third, node-specific covariance modeling and Mahalanobis-based scoring penalize deviations from the local normal structure. Finally, the memory is built only from normal training samples and is not updated with test data, preventing anomalous samples from being absorbed into the normal representation. Accordingly, ART-IAD does not suppress reconstruction shortcuts by imposing additional constraints on a reconstruction network; rather, it avoids the underlying reconstruction pathway and evaluates each patch according to its deviation from locally constrained normal prototype distributions. If this design operates as intended, anomalous regions should retain concentrated high responses in the predicted anomaly maps instead of being absorbed into an overly broad normal representation. We examine this expected behavior qualitatively in the Anomalous Region Localization subsection.
Algorithm 1. ART-IAD for Image Anomaly Detection and Localization
Require: Training image set , testing image set
Ensure: Trained ART models and corresponding anomaly score maps
1: # Training Phase
2: Initialize three Fuzzy ART networks: ART1, ART2, and ART3
3: for each incoming normal training image do
4: Extract patch embeddings for each view
5: for each view do
6: for each patch embedding do
7: Rank the existing nodes using the choice function
8: Apply the vigilance-based match–reset–search procedure
9: if an existing node J satisfies the vigilance criterion then
10: Update
11: else
12: Create a new node for
13: end if
14: end for
15: end for
16: end for
17: for each view do
18: for each node j in do
19: Estimate and
20: end for
21: end for
22: # Testing Phase
23: for each test image do
24: for each patch i and each view t = 1,2,3 do
25: Extract feature vector
26: Find the winner node in for
27: Compute the Mahalanobis distance w.r.t. and using Eq. (2)
28: end for
29: Fuse the three views to obtain the anomaly score using Eq. (3)
30: Upsample the anomaly scores via bilinear interpolation to generate a full-resolution anomaly map
31: end for
Complexity analysis
Let N denote the number of training images. Each image is divided into non-overlapping patches. We extract three semantic views from the first three ResNet blocks. For view , let be the embedding dimension and let be the number of Fuzzy ART nodes created after training. We denote by Cbackbone the computational cost of one forward pass of the ResNet backbone. In ART-IAD, each ART node stores a centroid vector and a full covariance matrix to describe the local distribution of normal patch embeddings assigned to that node.
Training time complexity.
Feature extraction over the training set costs . For ART learning, each incoming patch embedding requires computing the choice function over all existing ART nodes and then performing the vigilance test. Since both the choice function and the vigilance test operate on -dimensional vectors, the cost for processing one patch embedding in view t is in the worst case. Therefore, for view t, the ART assignment/update cost is , and the total ART training cost across the three views is
After the ART nodes are formed, we estimate node-wise distribution statistics for each ART node. Since ART-IAD uses a full covariance matrix, computing the covariance contribution of one -dimensional feature vector requires forming an outer product, which costs . Thus, accumulating full covariance matrices for all training patch embeddings in view t costs . Across the three views, the covariance estimation cost is
In addition, for Mahalanobis-distance-based scoring, the inverse of each node-wise covariance matrix is precomputed after training. Inverting one full covariance matrix costs . Therefore, for nodes in view t, covariance inversion costs , and the total inversion cost over the three views is
Therefore, the overall training time complexity is
The first term corresponds to CNN feature extraction, the second term corresponds to ART-based node matching and updating, the third term corresponds to full covariance estimation, and the fourth term corresponds to covariance matrix inversion. When is large, the full covariance estimation and inversion terms may become non-negligible compared with the ART matching cost.
Inference time complexity.
For a test image, feature extraction costs . For each view t, winner-node selection requires evaluating the Fuzzy ART choice function over all nodes for each of the P patch embeddings. Since the choice function is computed on -dimensional vectors, winner-node selection costs
After the winner node is selected, the Mahalanobis distance is computed between the patch embedding and the distribution represented by the winner node. Because ART-IAD uses a full covariance matrix, computing
requires a matrix-vector multiplication with a inverse covariance matrix. Therefore, the Mahalanobis distance computation for one patch costs , and for all P patches in view t, it costs
Therefore, the inference time per image is
Here, is the cost of winner-node selection, while is the cost of full-covariance Mahalanobis scoring. The cost of bilinear upsampling and three-view score fusion is linear in the number of patches or pixels and is usually smaller than feature extraction, ART matching, and Mahalanobis-distance computation.
Space complexity.
For each view t, ART-IAD stores the ART node representations and the node-wise distribution statistics. Specifically, each ART node stores a centroid vector of dimension , an ART weight vector of dimension , and a full covariance matrix of size . In addition, the inverse covariance matrix can be precomputed and stored for efficient Mahalanobis-distance computation during inference. Therefore, the storage required by each node is dominated by the full covariance or inverse covariance matrix, resulting in a per-node space complexity of .
Thus, for nodes in view t, the model memory required by is
and the total model memory across the three views is
excluding the backbone parameters.
During training, if all extracted patch embeddings are explicitly stored before ART modeling and covariance estimation, the temporary feature storage requires
memory. However, this storage can be reduced by streaming features or by accumulating sufficient statistics for each node. Compared with memory-bank-based methods that store all training patch embeddings with memory complexity , ART-IAD stores a compact set of node-wise distribution models. Although full covariance matrices increase the per-node memory from to , the overall storage can still be compact when the number of ART nodes is much smaller than the total number of training patches NP.
Scalability and stability discussion
The scalability of ART-IAD is primarily determined by the number of ART nodes in each semantic view. ART dynamically allocates new nodes for diverse normal patterns, while the vigilance parameter controls the granularity of this allocation: lower vigilance produces more compact prototypes, whereas higher vigilance yields finer normal partitions. Since memory and inference cost scale with , this provides an explicit trade-off between compactness and representation fidelity. Empirically, , allowing ART-IAD to replace the full patch-level memory bank with a compact set of adaptive distributional prototypes.
Stability is enhanced by node-wise full covariance modeling. Each covariance matrix captures local feature variance and cross-dimensional correlations, enabling Mahalanobis scoring in a locally normalized space. This makes the anomaly score less sensitive to feature scaling and more sensitive to deviations from the learned normal correlation structure. The regularization term further ensures invertibility and improves conditioning when node assignments are sparse or correlated.
During inference, ART nodes and their covariance statistics are fixed and are not updated by test samples, preventing anomalous inputs from contaminating the normal reference model. For large-scale or highly diverse data, scalability can be further maintained by adjusting vigilance, pruning redundant nodes, or increasing covariance regularization.
Results
In this section, we conduct a comprehensive evaluation of the proposed ART-IAD framework on a widely used benchmark dataset for industrial anomaly detection. First, the datasets and the evaluation metrics are introduced. Next, the implementation details together with the experimental settings are described. Finally, ART-IAD is compared against several state-of-the-art methods, for which both quantitative and qualitative analyses are provided, so that its effectiveness in anomaly detection and localization can be demonstrated.
Datasets and evaluation metrics
Datasets.
The proposed method is evaluated on the MVTec Anomaly Detection (MVTec AD) dataset [31]https://www.mvtec.com/company/research/datasets/mvtec-ad, which was set up as a benchmark for unsupervised industrial anomaly detection and localization. It contains 3,629 defect-free training images and 1,725 test images, including both normal and anomalous samples. All images were captured using a 2048-pixel high-resolution RGB industrial sensor, which was equipped with bilateral telecentric lenses at 1:5 and 1:1 magnifications. The images are high-resolution RGB industrial inspection images with resolutions ranging from to . Also, pixel-level annotations are provided for every anomalous region, so that localization performance can be evaluated precisely.
MVTec AD consists of 15 categories, including 10 object categories and 5 texture categories. Each category contains multiple real-world defect types, such as scratches, dents, contamination, cracks, and structural deformations, with approximately five defect types per category on average. Representative examples are shown in Fig 2. MVTec AD has become the de facto benchmark for industrial anomaly detection, with many state-of-the-art methods reporting results on this dataset [47,73].
To further evaluate the generalization ability of ART-IAD, we also conduct experiments on the Visual Anomaly (VisA) dataset [74], which has recently been widely used as a complementary benchmark to MVTec AD. VisA contains 10,821 high-resolution industrial images from 12 object categories, including 9,621 normal images and 1,200 anomalous images. Compared with MVTec AD, VisA contains more complex object layouts and larger intra-class variations, making it suitable for evaluating cross-category generalization.
Evaluation metrics.
We adopt commonly used metrics, including Accuracy, Precision, Recall, F1-score, and AUROC, to evaluate anomaly detection and localization performance. Evaluated by these metrics are image-level classification performance as well as pixel-level localization capability.
In addition to the above, we report the Area Under the Receiver Operating Characteristic Curve (AUROC) as a threshold-independent metric, which is commonly used in anomaly detection to assess the quality of continuous anomaly scores. Specifically, two AUROC variants are considered: image-level AUROC, which measures the ability to distinguish between normal and anomalous images; and pixel-level AUROC, which evaluates the localization of defective regions within images.These metrics jointly capture both detection robustness and localization accuracy [31].
Experimental setup
We adopt Wide-ResNet502 pretrained on ImageNet as the frozen backbone of ART-IAD for both MVTec AD and VisA. For ART-IAD, images from both datasets are resized to and center-cropped to . This input setting is consistent with the standard PatchCore configuration [10]. The competing results reported below are taken directly from their original publications rather than reproduced in our experimental environment; consequently, their preprocessing procedures and feature extraction settings follow the respective source studies and may differ from those of ART-IAD. All comparisons use the dataset splits, image-level AUROC metrics, and pixel-level AUROC metrics reported in the corresponding studies, while other implementation settings follow the configurations of each method.
Intermediate feature maps are extracted from the first three blocks of the ResNet backbone (i.e., layer1, layer2, and layer3) with channel dimensions of 256, 512, and 1024, respectively. These are treated as three distinct semantic views, each modeled by an independent Fuzzy ART network (denoted ART1, ART2, and ART3). The initial prototype weights are computed from the training set, after which subsequent patch embeddings update the ART memory sequentially through the match–reset–search procedure.
During inference, each test image is divided into patches. When the feature embedding for each patch is extracted from the corresponding view, it is fed into the trained ART network, where the Mahalanobis distance to its winning prototype node is computed. Three distance maps (one per view) are produced and upsampled to a common resolution () using bilinear interpolation. The final anomaly map is generated by summing the distance maps pixel-wise. For reproducibility, the detailed hyperparameter settings, feature dimensions, implementation details, and hardware/software environment are summarized in Table 1.
Experimental results and analysis
Performance comparison.
We evaluate ART-IAD across all 15 categories of MVTec AD, reporting Accuracy, Precision, Recall, and F1-Score (Table 2). The results demonstrate consistently strong performance across categories, confirming the robustness and generalizability of our method.
As shown in Table 2, ART-IAD achieves stable performance across all 15 categories, with the accuracy of most categories exceeding 95%. This indicates that the proposed prototype-based modeling can effectively distinguish normal and anomalous samples under different object and texture appearances. In particular, categories such as Screw, Leather, Capsule, Carpet, and Grid show very high scores for the normal class, suggesting that the learned ART prototypes can well characterize normal patterns. Relatively lower scores for the anomalous class are observed in categories such as Grid, Screw, and Leather, which is mainly because the defective regions in these categories are small and sparse, making threshold-based anomaly classification more sensitive to minor false positives or false negatives. Nevertheless, the overall accuracy remains high, demonstrating the stable detection capability of ART-IAD across diverse industrial categories.
For comparison with state-of-the-art methods, we benchmark ART-IAD against both reconstruction-based approaches, including AE-SSIM [57], RIAD [54], DR/EM [55], CutPaste [75], and PNPT [25], as well as feature-embedding-based approaches, including PaDiM [1], RevDist [42], PatchCore [10], and OMAC [26]. Results are summarized in Table 3.
Table 3 reports the image-level AUROC and pixel-level AUROC on the MVTec AD benchmark. Overall, ART-IAD achieves the best average performance among all compared methods, obtaining 99.4% image-level AUROC and 98.2% pixel-level AUROC. Compared with reconstruction-based methods, ART-IAD shows a clear advantage in image-level anomaly detection. Specifically, it improves the overall image-level AUROC over the strongest reconstruction-based baseline PNPT from 98.3% to 99.4%, demonstrating that feature-prototype modeling is more effective than reconstruction error for distinguishing subtle industrial anomalies. In terms of pixel-level localization, ART-IAD also achieves the highest overall average pixel-level AUROC of 98.2%, slightly outperforming PNPT and PatchCore, both of which achieve 98.1%. ART-IAD achieves image-level and pixel-level AUROC values that are 0.3 and 0.1 percentage points higher, respectively, than the values reported for PatchCore.
A more detailed comparison further shows that ART-IAD performs consistently well on both texture and object categories. For texture categories, ART-IAD achieves the highest average image-level AUROC of 99.7%, indicating strong capability in detecting appearance changes such as defects on carpet, leather, tile, and wood. Although its average pixel-level AUROC on textures is 97.6%, which is comparable to the best competing methods, the image-level detection performance remains the strongest. For object categories, ART-IAD achieves the best average performance in both metrics, with 99.3% image-level AUROC and 98.6% pixel-level AUROC. This indicates that the proposed ART-based prototype learning strategy not only captures discriminative normal patterns for image-level anomaly detection but also preserves spatially meaningful information for accurate anomaly localization.
Although ART-IAD shows strong results on most categories, the AUROC on Capsule and Screw is still slightly behind some competing methods. A closer look suggests that many normal samples there already include tiny imperfections or some background artifacts, which is likely to raise the false positive rate. As shown in Figs 3 and 4, fine-grained anomalous regions can still be localized accurately by our method, indicating that ART-IAD remains robust under challenging imaging conditions. If data acquisition is improved, for example with higher resolution, more uniform lighting, or stricter quality control, this kind of background interference could be further reduced.
To further evaluate the generalization ability of ART-IAD, we also conduct comparisons on the VisA dataset, as reported in Table 4. Compared with existing feature-embedding-based methods, ART-IAD achieves an average pixel-level AUROC of 98.6%, which is 7.3 and 0.5 percentage points higher than the values reported for FMAS+WDAM and OMAC, respectively. Since DiAD [41] only reports dataset-level results on VisA rather than per-category performance, we additionally compare its reported AUROC-cls and AUROC-seg scores. ART-IAD achieves 94.3% image-level AUROC and 98.6% pixel-level AUROC, exceeding DiAD by 7.5 and 2.6 percentage points. This indicates that the proposed ART-based prototype learning strategy not only maintains strong image-level discrimination but also provides more accurate anomaly localization on complex industrial objects. Although ART-IAD does not achieve the highest average image-level AUROC on VisA, it remains competitive and provides the best pixel-level localization performance.
A category-wise analysis further shows that ART-IAD achieves the best or tied-best pixel-level performance on most VisA categories, including Capsules, Cashew, Fryum, Macaroni1, Macaroni2, PCB1, PCB2, PCB3, and Pipe Fryum. These results demonstrate that ART-IAD can effectively preserve fine-grained spatial anomaly cues and robustly localize small or irregular defects. For image-level anomaly detection, ART-IAD obtains a competitive average AUROC of 94.3%, achieving the best results on Chewinggum and PCB4 and remaining close to the best method on categories such as Cashew, PCB1, PCB3, and Pipe Fryum. The relatively lower image-level AUROC on Capsules and Macaroni2 suggests that image-level discrimination on VisA is more sensitive to intra-class appearance variations and subtle global differences. Nevertheless, the superior pixel-level performance confirms the effectiveness of ART-IAD for industrial anomaly localization.
Runtime and configuration analysis.
In addition to detection and localization accuracy, we further report the per-category configuration and runtime of ART-IAD in Table 5. The total runtime is decomposed into feature extraction, ART training, and inference. Since the backbone is frozen, feature extraction remains relatively stable across categories, with an average cost of 22.6 seconds. The main computational cost comes from ART prototype construction and Mahalanobis-distance-based inference, whose runtime depends on the number of training and testing images as well as the number of generated prototypes.
As shown in Table 5, ART-IAD requires 459.6 seconds on average for one category. Texture categories require a higher average runtime than object categories, mainly because some texture classes, such as Carpet, generate a larger number of prototypes under a high vigilance setting. In contrast, categories such as Toothbrush require much less time due to their smaller data size. The results show that ART-IAD remains practical for industrial anomaly detection while preserving the advantage of prototype-based online learning. Moreover, the vigilance parameter provides a controllable trade-off between prototype granularity and computational cost: a larger generally leads to finer categories and potentially higher ART training time, whereas a smaller encourages more compact prototype representation.
Comparison with KMeans clustering.
To further verify the effectiveness of Fuzzy ART as the prototype learning module, we compare it with a KMeans baseline under the same three-layer multi-view fusion setting. For KMeans, the number of clusters is fixed according to the spatial resolutions of the three views, namely 3136, 784, and 196 clusters for L1, L2, and L3, respectively, resulting in 4116 clusters in total. In contrast, Fuzzy ART adaptively determines the number of active statistical prototypes used for node-wise scoring through the vigilance parameter .
The comparison is reported in Table 6. Overall, ART achieves better or comparable performance than KMeans across image-level AUROC, pixel-level AUROC, and PRO-AUC. On average, ART improves image-level AUROC from 99.1% to 99.4%, and PRO-AUC from 92.9% to 93.2%, while maintaining the same average pixel-level AUROC of 98.2%. This indicates that the adaptive prototype formation of ART is more suitable for modeling category-specific normal patterns than using a fixed number of clusters.
Another important observation is that ART often uses fewer prototypes than KMeans. The average number of prototypes is reduced from 4116 to 3841, while achieving slightly better overall performance. This suggests that ART can remove redundant cluster centers and allocate prototypes according to the complexity of the normal feature distribution. For example, Grid activates 1897 statistical prototypes for anomaly scoring, much fewer than KMeans, but still achieves better image-level and pixel-level AUROC. These results support the use of Fuzzy ART as an adaptive and online prototype learner in ART-IAD.
Online efficiency analysis.
To quantify the deployment cost of explicit online ART updating, we report an online efficiency analysis in Table 7. Unlike the runtime analysis for the accuracy experiments in Table 5 and the comparison of active prototypes in Table 6, this supplementary analysis tracks node growth during online training, where each normal patch either updates a matched ART node or creates a new one when the vigilance criterion is not satisfied. The table reports node growth, memory consumption, ART update time, and inference latency. For comparison with representative anomaly detection methods based on feature embeddings, we also include a reference model that stores all normal patch embeddings from the same three feature levels, following the memory bank designs represented by SPADE and PatchCore.
Here, “Init.”, “Final”, and “New” denote the initial, final, and newly created ART nodes in the supplementary online-growth setting, while Table 6 reports active statistical prototypes used in the main scoring pipeline. Compared with storing all normal patch embeddings, ART-IAD represents the reference set using a compact set of adaptive ART nodes. The reduced setting further lowers memory and latency while preserving competitive detection accuracy, showing the trade-off between representation compactness and online update cost.
Multi-view fusion ablation.
We further conduct a multi-view ablation study to evaluate the contribution of different ResNet feature layers. Specifically, L1, L2, and L3 denote the features extracted from layer1, layer2, and layer3, respectively, while Full denotes the fusion of all three views. The pixel-level AUROC results are shown in Table 8.
The results demonstrate that different layers contribute differently to anomaly localization. L1 features preserve fine spatial details and are beneficial for detecting small or sharp local defects, such as those in Grid, Tile, Metal Nut, and Screw. L2 features achieve the best average performance on texture categories, reaching 97.8% P-AUROC, indicating that mid-level representations provide a good balance between local detail and semantic robustness. L3 features are useful for categories requiring higher-level semantic information, such as Transistor, where L3 achieves the best P-AUROC of 98.1%.
Although a single view can be optimal for certain categories, the fused model achieves the best overall average P-AUROC. In particular, the full model obtains the highest average P-AUROC on object categories, reaching 98.6%. This confirms that the three views are complementary: low-level features capture fine local texture variations, middle-level features provide stable defect patterns, and high-level features encode more semantic object-level information. Therefore, late fusion of ART distance maps is an effective strategy for robust anomaly localization.
Anomalous region localization.
To qualitatively evaluate anomaly localization, we compare the predicted anomaly maps with the ground-truth annotations in Figs 5–7. Each row shows the input image, ground-truth mask, predicted heatmap, predicted binary mask, and segmentation result. Higher anomaly scores are represented by warmer colors in the heatmaps.
ART-IAD generally produces compact high-response regions that align well with the annotated defects. It accurately localizes structural anomalies, such as cracks, broken regions, and missing components, while remaining sensitive to local irregularities in repetitive textures. For categories containing small and visually subtle defects, including Capsule, Pill, and Screw, mild responses may appear around the target regions, but the principal defects remain clearly activated. These observations are consistent with the average pixel-level AUROC of 98.2% on MVTec AD reported in Table 3.
These localization results also provide qualitative evidence regarding the reconstruction-shortcut problem. Unlike reconstruction-based models, which may reproduce anomalous structures and consequently yield weak residual responses, ART-IAD directly measures the deviation of each test patch from a locally constrained normal prototype distribution. The vigilance criterion limits the coverage of each prototype, while prototype-specific Mahalanobis scoring preserves responses to deviations from local normal statistics. Consequently, the concentrated activations over defective regions in Figs 5–7 are consistent with the intended avoidance of reconstruction-based anomaly overgeneralization. Since ART-IAD contains no explicit reconstruction branch, we interpret these results as qualitative support for this design rather than evidence that all forms of shortcut learning are eliminated.
Impact of the vigilance parameter .
As the key hyperparameter of the Fuzzy ART network, the vigilance parameter is what largely decides how fine the categories are formed, which in turn affects the clustering behavior and anomaly detection performance. To see its effect, experiments are first conducted on the Screw category in MVTec AD, where is varied in the range [0.1, 1.0]. The results are shown in Fig 8.
It is when is set in the range 0.1 to 0.7 that stable, high performance is achieved by the model, in terms of AUROC at both image and pixel level. In most cases, optimal or near-optimal detection is reached within this interval. But once goes beyond 0.7, performance is seen to drop quite a lot. This seems to indicate that an overly high vigilance causes too much category fragmentation, which may fragment the learned representation and degrade generalization.
When the generalizability of across different scenarios was examined, a more comprehensive study was carried out on the MVTec AD dataset. The AUROC trends of each category, as changes, are summarized in Fig 9. Although the best does vary by category, it is still when falls in 0.3 to 0.7 that most categories reach their best or near-best results. This indicates that ART-IAD is not overly sensitive to within a broad stable range, although category-specific values are used in our final experiments to obtain the best performance.
Six performance metrics are reported, including Accuracy, Precision, Recall, F1-score, pixel-level AUROC (P-AUROC), and image-level AUROC (I-AUROC).
Based on these observations, is selected within a stable range for each category in the final experiments, as reported in Table 5. Most selected values fall within the empirically stable interval, while a small number of categories benefit from slightly higher or lower vigilance due to their different intra-class variations and defect scales.
Beyond detection accuracy, we further analyze how affects online prototype growth and memory usage. This supplementary analysis is not used to select the accuracy-oriented values in the main experiments; instead, it uses stricter vigilance values to make online node growth explicit and to quantify the resulting memory and runtime cost. As shown in Table 9, a larger imposes a stricter matching criterion, leading to more newly created ART nodes and higher prototype memory. In contrast, a smaller encourages more compact representations by allowing more patches to update existing nodes. The results are consistent with the trend observed in Figs 8 and 9: moderate values of provide a favorable trade-off between accuracy and memory cost, whereas overly high vigilance may over-fragment normal patterns and degrade image-level detection performance.
We also consider the parameter used for prototype-level statistical estimation. ART nodes activated by very few samples may produce unstable covariance estimates. Therefore, a minimum sample threshold is used when estimating the distribution of each node. A smaller threshold allows more nodes to participate in anomaly scoring but may introduce noisy statistics, whereas a larger threshold improves statistical stability but may exclude rarely activated nodes. In our experiments, we use a small threshold to preserve online adaptability while avoiding unstable statistics from nodes activated by only one sample.
Streaming and continual learning evaluation.
A key advantage of Fuzzy ART is its online and incremental learning capability. To evaluate whether ART-IAD can retain previously learned normal patterns when new classes arrive sequentially, we conduct a streaming learning experiment on MVTec AD. Classes are grouped into three sequences, including Texture, Metal, and Mixed groups. For each group, classes are trained one by one using the same ART model. After each new class is learned, winmaps for all seen classes are rebuilt with the current ART weights, and the final streaming performance is compared with independently trained ART models.
The results are shown in Table 10. The forgetting degree is measured by , defined as the difference between independent P-AUROC and final streaming P-AUROC. A smaller indicates better retention. Overall, the average is only 0.2 for the first classes and 1.2 for later classes, suggesting that ART-IAD does not suffer from severe catastrophic forgetting. In several cases, such as Leather and Bottle, the streaming setting even slightly improves the P-AUROC, indicating that shared prototypes may provide beneficial regularization when related normal patterns are learned sequentially.
We further analyze the sensitivity of ART-IAD to class arrival order in Table 11. For each group, three classes are trained sequentially under both original and reversed orders. denotes the P-AUROC drop of the k-th class, computed as the difference between independent and final streaming P-AUROC; thus, positive values indicate forgetting, while negative values indicate slight improvement due to shared ART weights. Across all evaluated sequences, the absolute performance change remains below 0.05, demonstrating that ART-IAD is robust to different class arrival orders and does not suffer from catastrophic forgetting. Although the Mixed group shows a relatively larger degradation for Transistor in the original order, the drop remains moderate. These results suggest that ART-IAD is well suited for streaming industrial inspection scenarios, where new product classes can be incorporated without retraining the entire model from scratch.
Discussion
The results on MVTec AD demonstrate that ART-IAD achieves strong and consistent anomaly detection and localization performance. As shown in Table 3, it obtains 99.4% image-level AUROC and 98.2% pixel-level AUROC on average, outperforming PatchCore by 0.3% and 0.1%, respectively. This suggests that online ART-based prototype modeling is a viable alternative to fixed memory-bank nearest-neighbor matching for industrial anomaly detection.
The ablation and comparison studies further validate the proposed design. Table 8 shows that multi-level CNN features provide complementary localization cues, while Table 6 confirms that Fuzzy ART adaptively allocates prototypes and achieves comparable or better performance than fixed-cluster KMeans with fewer prototypes on average. The runtime results in Table 5 indicate that the framework remains practical for offline model construction. Moreover, the streaming experiments in Table 10 and Table 11 show that ART-IAD can incorporate new classes sequentially without severe catastrophic forgetting, which is beneficial for evolving industrial inspection scenarios.
ART-IAD still has several limitations. Categories such as Capsule and Screw remain challenging because tiny imperfections or background artifacts in normal samples may disturb the anomaly score distribution. In addition, the current equal-weight fusion cannot adaptively suppress unreliable views for specific categories. Future work will investigate adaptive view weighting and more robust score calibration while preserving the online prototype-memory property of ART-IAD.
Conclusion
We presented ART-IAD, an online unsupervised framework for industrial anomaly detection and localization. By modeling multi-level patch embeddings with separate Fuzzy ART prototype memories and estimating per-prototype distribution statistics, ART-IAD enables compact normal-reference modeling without storing all normal patch features.
On MVTec AD, ART-IAD achieves 99.4% image-level AUROC and 98.2% pixel-level AUROC on average, outperforming PatchCore by 0.3 and 0.1 percentage points, respectively. Additional results on VisA further validate its generalization ability for anomaly localization. These results demonstrate the effectiveness of ART-based adaptive prototype learning for industrial anomaly detection, while the added efficiency and sensitivity analyses quantify its online updating behavior and the balance between memory consumption and computational efficiency.
Future work will focus on adaptive multi-view fusion and robust score calibration while preserving the simplicity and incremental nature of ART-IAD.
References
- 1.
Defard T, Setkov A, Loesch A, Audigier R. Padim: A patch distribution modeling framework for anomaly detection and localization. International Conference on Pattern Recognition. Springer; 2021. p. 475–89.
- 2.
Liu Z, Zhou Y, Xu Y, Wang Z. Simplenet: A simple network for image anomaly detection and localization. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; 2023. p. 20402–11.
- 3.
Ugwu CI, Casarin S, Lanz O. Fractals as pre-training datasets for anomaly detection and localization. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; 2024. p. 163–72.
- 4.
Tu Y, Zhang B, Liu L, Li Y, Zhang J, Wang Y, et al. Self-supervised feature adaptation for 3d industrial anomaly detection. European Conference on Computer Vision. Springer; 2025. p. 75–91.
- 5. Liu J, Xie G, Wang J, Li S, Wang C, Zheng F, et al. Deep industrial image anomaly detection: A survey. Mach Intell Res. 2024;21(1):104–35.
- 6.
Chen Q, Luo H, Lv C, Zhang Z. A unified anomaly synthesis strategy with gradient ascent for industrial anomaly detection and localization. European Conference on Computer Vision. Springer; 2025. p. 37–54.
- 7. Xie G, Wang J, Liu J, Lyu J, Liu Y, Wang C, et al. IM-IAD: Industrial image anomaly detection benchmark in manufacturing. IEEE Trans Cybern. 2024;54(5):2720–33. pmid:38381632
- 8. Iqbal A, Amin R, Alsubaei FS, Alzahrani A. Anomaly detection in multivariate time series data using deep ensemble models. PLoS One. 2024;19(6):e0303890. pmid:38843255
- 9. Yang T, Jiang X, Li W, Liu P, Wang J, Hao W, et al. Cloud-edge collaborative data anomaly detection in industrial sensor networks. PLoS One. 2025;20(6):e0324543. pmid:40498796
- 10.
Roth K, Pemula L, Zepeda J, Schölkopf B, Brox T, Gehler P. Towards total recall in industrial anomaly detection. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; 2022. p. 14318–28.
- 11. Zhang Q, Zhu Y, Cordeiro FR, Chen Q. PSSCL: A progressive sample selection framework with contrastive loss designed for noisy labels. Pattern Recognit. 2025;161:111284.
- 12. Zhang Q, Zhu Y, Yang M, Jin G, Zhu Y, Chen Q. Cross-to-merge training with class balance strategy for learning with noisy labels. Expert Syst Appl. 2024;249:123846.
- 13. Li W, Lai T, Liu G, Fan H, Li Z. Novelty detection of leukocyte image via mean-shifted feature and directly optimized subspace. IEEE Trans Instrum Meas. 2023;72:1–12.
- 14. Lu R, Wu Y, Tian L, Wang D, Chen B, Liu X, et al. Hierarchical vector quantized transformer for multi-class unsupervised anomaly detection. Adv Neural Inf Process Syst. 2023;36:8487–500.
- 15.
Yu J, Zheng Y, Wang X, Li W, Wu Y, Zhao R, et al. Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows. arXiv:211107677 [Preprint]. 2021.
- 16.
Gudovskiy D, Ishizaka S, Kozuka K. Cflow-ad: Real-time unsupervised anomaly detection with localization via conditional normalizing flows. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision; 2022. p. 98–107.
- 17.
Hou J, Zhang Y, Zhong Q, Xie D, Pu S, Zhou H. Divide-and-assemble: Learning block-wise memory for unsupervised anomaly detection. Proceedings of the IEEE/CVF International Conference on Computer Vision; 2021. p. 8791–800.
- 18.
Yan X, Zhang H, Xu X, Hu X, Heng PA. Learning semantic context from normal samples for unsupervised anomaly detection. Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35; 2021. p. 3110–8.
- 19. Jiang X, Liu J, Wang J, Nie Q, Wu K, Liu Y, et al. Softpatch: Unsupervised anomaly detection with noisy data. Adv Neural Inf Process Syst. 2022;35:15433–45.
- 20. Tao X, Gong X, Zhang X, Yan S, Adak C. Deep learning for unsupervised anomaly localization in industrial images: A survey. IEEE Trans Instrum Meas. 2022;71:1–21.
- 21.
Ristea NC, Madan N, Ionescu RT, Nasrollahi K, Khan FS, Moeslund TB, et al. Self-supervised predictive convolutional attentive block for anomaly detection. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; 2022. p. 13576–86.
- 22.
Liu J, Wu K, Nie Q, Chen Y, Gao BB, Liu Y, et al. Unsupervised continual anomaly detection with contrastively-learned prompt. Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38; 2024. p. 3639–47.
- 23.
Zheng Y, Wang X, Deng R, Bao T, Zhao R, Wu L. Focus your distribution: Coarse-to-fine non-contrastive learning for anomaly detection and localization. 2022 IEEE International Conference on Multimedia and Expo (ICME). IEEE; 2022. p. 1–6.
- 24.
Jiang X, Chen Y, Nie Q, Liu J, Liu Y, Wang C, et al. Toward multi-class anomaly detection: Exploring class-aware unified model against inter-class interference. arXiv:240314213 [Preprint]. 2024.
- 25. Yao H, Cao Y, Luo W, Zhang W, Yu W, Shen W. Prior normality prompt transformer for multiclass industrial image anomaly detection. IEEE Trans Ind Inf. 2024;20(10):11866–76.
- 26. Tan P, Wong WK. Unsupervised anomaly detection and localization with one model for all category. Knowl-Based Syst. 2024;289:111533.
- 27. Li X, Peng J, Li W, Song Z, Du X. Generative adversarial local density-based unsupervised anomaly detection. PLoS One. 2025;20(1):e0315721. pmid:39854383
- 28. Zhao G, Wang L, Wang X. A Mallows-like criterion for anomaly detection with random forest implementation. PLoS One. 2025;20(6):e0323333. pmid:40478801
- 29. You Z, Cui L, Shen Y, Yang K, Lu X, Zheng Y, et al. A unified model for multi-class anomaly detection. Adv Neural Inf Process Syst. 2022;35:4571–84.
- 30. Luo W, Yao H, Yu W. Template-based Feature Aggregation Network for industrial anomaly detection. Eng Appl Artif Intell. 2024;131:107810.
- 31.
Bergmann P, Fauser M, Sattlegger D, Steger C. MVTec AD–A comprehensive real-world dataset for unsupervised anomaly detection. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; 2019. p. 9592–600.
- 32.
Bergmann P, Fauser M, Sattlegger D, Steger C. Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; 2020. p. 4183–92.
- 33.
Gong D, Liu L, Le V, Saha B, Mansour MR, Venkatesh S, et al. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. Proceedings of the IEEE/CVF International Conference on Computer Vision; 2019. p. 1705–14.
- 34.
Somepalli G, Wu Y, Balaji Y, Vinzamuri B, Feizi S. Unsupervised anomaly detection with adversarial mirrored autoencoders. Uncertainty in Artificial Intelligence. PMLR; 2021. p. 365–75.
- 35.
Lee WY, Wang YCF. Learning disentangled feature representations for anomaly detection. 2020 IEEE International Conference on Image Processing (ICIP). IEEE; 2020. p. 2156–60.
- 36.
Pol AA, Berger V, Germain C, Cerminara G, Pierini M. Anomaly detection with conditional variational autoencoders. 2019 18th IEEE international conference on machine learning and applications (ICMLA). IEEE; 2019. p. 1651–7.
- 37.
Liu W, Li R, Zheng M, Karanam S, Wu Z, Bhanu B, et al. Towards visually explaining variational autoencoders. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; 2020. p. 8642–51.
- 38.
Venkataramanan S, Peng KC, Singh RV, Mahalanobis A. Attention guided anomaly localization in images. European Conference on Computer Vision. Springer; 2020. p. 485–503.
- 39.
Perera P, Nallapati R, Xiang B. Ocgan: One-class novelty detection using gans with constrained latent representations. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; 2019. p. 2898–906.
- 40. Schlegl T, Seeböck P, Waldstein SM, Langs G, Schmidt-Erfurth U. f-AnoGAN: Fast unsupervised anomaly detection with generative adversarial networks. Med Image Anal. 2019;54:30–44. pmid:30831356
- 41.
He H, Zhang J, Chen H, Chen X, Li Z, Chen X, et al. A diffusion-based framework for multi-class anomaly detection. Proceedings of the AAAI conference on artificial intelligence. vol. 38; 2024. p. 8472–80.
- 42.
Deng H, Li X. Anomaly detection via reverse distillation from one-class embedding. Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition; 2022. p. 9737–46.
- 43.
Rudolph M, Wehrbein T, Rosenhahn B, Wandt B. Fully convolutional cross-scale-flows for image-based defect detection. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision; 2022. p. 1088–97.
- 44. Lyu S, Mo D, keung Wong W. REB: Reducing biases in representation for industrial anomaly detection. Knowl-Based Syst. 2024;290:111563.
- 45.
Bergman L, Hoshen Y. Classification-based anomaly detection for general data. arXiv:200502359 [Preprint]. 2020.
- 46.
Rippel O, Mertens P, Merhof D. Modeling the distribution of normal data in pre-trained deep features for anomaly detection. 2020 25th International Conference on Pattern Recognition (ICPR). IEEE; 2021. p. 6726–33.
- 47.
Mishra P, Piciarelli C, Foresti GL. Image anomaly detection by aggregating deep pyramidal representations. International Conference on Pattern Recognition. Springer; 2021. p. 705–18.
- 48.
Reiss T, Cohen N, Bergman L, Hoshen Y. Panda: Adapting pretrained features for anomaly detection and segmentation. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; 2021. p. 2806–14.
- 49.
Cohen N, Hoshen Y. Sub-image anomaly detection with deep pyramid correspondences. arXiv:200502357 [Preprint]. 2020.
- 50.
Tien TD, Nguyen AT, Tran NH, Huy TD, Duong S, Nguyen CDT, et al. Revisiting reverse distillation for anomaly detection. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; 2023. p. 24511–20.
- 51.
Zavrtanik V, Kristan M, Skočaj D. Dsr–a dual subspace re-projection network for surface anomaly detection. European Conference on Computer Vision. Springer; 2022. p. 539–54.
- 52.
Wu W, Lu Z, Lu Z, He Z, Sun X, Wang Z, et al. Improving anomaly detection with foundation-model synthesis and wavelet-domain attention. Available from: SSRN 5525650
- 53. Wunsch Ii DC. Admiring the Great Mountain: A celebration special issue in honor of Stephen Grossberg’s 80th birthday. Neural Netw. 2019;120:1–4. pmid:31587821
- 54. Zavrtanik V, Kristan M, Skočaj D. Reconstruction by inpainting for visual anomaly detection. Pattern Recognit. 2021;112:107706.
- 55.
Zavrtanik V, Kristan M, Skočaj D. Draem-A discriminatively trained reconstruction embedding for surface anomaly detection. Proceedings of the IEEE/CVF International Conference on Computer Vision; 2021. p. 8330–9.
- 56.
Schlüter HM, Tan J, Hou B, Kainz B. Natural synthetic anomalies for self-supervised anomaly detection and localization. European Conference on Computer Vision. Springer; 2022. p. 474–89.
- 57.
Bergmann P, Löwe S, Fauser M, Sattlegger D, Steger C. Improving unsupervised defect segmentation by applying structural similarity to autoencoders. arXiv:180702011 [Preprint]. 2018.
- 58. Liang Y, Zhang J, Zhao S, Wu R, Liu Y, Pan S. Omni-frequency channel-selection representations for unsupervised anomaly detection. IEEE Trans Image Process. 2023;32:4327–40. pmid:37486834
- 59.
Mousakhan A, Brox T, Tayyub J. Anomaly detection with conditioned denoising diffusion models. arXiv:230515956 [Preprint]. 2023.
- 60.
Pirnay J, Chai K. Inpainting transformer for anomaly detection. International Conference on Image Analysis and Processing. Springer; 2022. p. 394–406.
- 61. Jiang J, Zhu J, Bilal M, Cui Y, Kumar N, Dou R, et al. Masked swin transformer unet for industrial anomaly detection. IEEE Trans Ind Inf. 2023;19(2):2200–9.
- 62. Zhang F, Chen Z. IDDM: An incremental dual-network detection model for in-situ inspection of large-scale complex product. J Ind Inf Integr. 2023;33:100463.
- 63.
Bergman L, Cohen N, Hoshen Y. Deep nearest neighbor anomaly detection. arXiv:200210445 [Preprint]. 2020.
- 64.
Sohn K, Li CL, Yoon J, Jin M, Pfister T. Learning and evaluating representations for deep one-class classification. arXiv:201102578 [Preprint]. 2020.
- 65. Liu Y, Gao X, Wen JZ, Luo H. Unsupervised image anomaly detection and localization in industry based on self-updated memory and center clustering. IEEE Trans Instrum Meas. 2023;72:1–10.
- 66. Lian X, Zheng Y, Liu Y, Zhou F, Peng C, Gao X. Contextual masking distillation for network traffic anomaly detection. IEEE Trans Inf Forensic Secur. 2026;21:1273–86.
- 67.
Lian X, Cao C, Liu Y, Xu X, Zheng Y, Zhou F. Facing Anomalies Head-On: Network traffic anomaly detection via uncertainty-inspired inter-sample differences. Proceedings of the ACM on Web Conference 2025; 2025. p. 3908–17.
- 68. Lian X, Zheng Y, Dang Z, Peng C, Gao X. Semi-supervised anomaly traffic detection via multi-frequency reconstruction. Pattern Recognit. 2025;161:111215.
- 69.
Lian X, Cao C, Zhong T, Wang Y, Chen K, Zhou F. Decompose to understand, fuse to detect: Frequency-decoupled anomaly detection for encrypted network traffic. IEEE INFOCOM 2026-IEEE Conference on Computer Communications. IEEE; 2026. p. 1–10.
- 70.
Lian X, Cao C, Zhong T, Zhou F. Disentangling multi-view scanning in mamba for network traffic anomaly detection. Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2; 2026. p. 2991–3002.
- 71. Khan IA, Razzak I, Pi D, Zia U, Kamal S, Hussain Y. A novel collaborative SRU network with dynamic behaviour aggregation, reduced communication overhead and explainable features. IEEE J Biomed Health Inform. 2024;28(6):3228–35. pmid:38198252
- 72. Carpenter GA, Grossberg S, Rosen DB. Fuzzy ART: Fast stable learning and categorization of analog patterns by an adaptive resonance system. Neural Netw. 1991;4(6):759–71.
- 73. Jiang Z, Zhang Y, Wang Y, Li J, Gao X. FR-PatchCore: An industrial anomaly detection method for improving generalization. Sensors (Basel). 2024;24(5):1368. pmid:38474903
- 74.
Zou Y, Jeong J, Pemula L, Zhang D, Dabeer O. Spot-the-difference self-supervised pre-training for anomaly detection and segmentation. European Conference on Computer Vision. Springer; 2022. p. 392–408.
- 75.
Li CL, Sohn K, Yoon J, Pfister T. Cutpaste: Self-supervised learning for anomaly detection and localization. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; 2021. p. 9664-–74.

관련 뉴스