Source-linked AI summary
Deep Industrial Image Anomaly Detection: A Survey
Jiaqi Liu, Guoyang Xie, Jinbao Wang, Shangnian Li, Chengjie Wang, Feng Zheng, Yaochu Jin
TL;DR
Industrial image anomaly detection needs a systematic understanding of methods and their applicability to manufacturing, where abnormal samples are costly and real-world conditions impose additional constraints. The paper surveys deep learning-based IAD across architectures, supervision, datasets, metrics, and manufacturing settings, then synthesizes findings and future challenges. Its review highlights architecture trade-offs, segmentation benefits, reconstruction strengths for segmentation, and practical needs including realistic datasets, multimodal inputs, faster inference, and lightweight models.
Problem
Industrial IAD requires more careful analysis of effectiveness and applicable scenarios because real-world manufacturing introduces unresolved dataset, assessment, and deployment challenges.
Method
The survey reviews IAD methods by supervision level, neural network architecture, loss functions, datasets, metrics, and industrial manufacturing settings.
Results
The survey finds that segmentation modules benefit anomaly detection, reconstruction-based methods are more suitable for segmentation tasks, and ensemble learning can improve state-of-the-art methods.
Takeaways & Limitations
The synthesis supports selecting IAD algorithms by architecture and supervision while highlighting realistic manufacturing datasets, standardized assessment, and deployment-oriented research directions.
Takeaways & Limitations
Current IAD datasets are mostly not collected from real production lines, and pretrained feature extractors can impose high computation costs that limit production deployment.
Abstract
from arXiv · showhide
The recent rapid development of deep learning has laid a milestone in industrial Image Anomaly Detection (IAD). In this paper, we provide a comprehensive review of deep learning-based image anomaly detection techniques, from the perspectives of neural network architectures, levels of supervision, loss functions, metrics and datasets. In addition, we extract the new setting from industrial manufacturing and review the current IAD approaches under our proposed our new setting. Moreover, we highlight several opening challenges for image anomaly detection. The merits and downsides of representative network architectures under varying supervision are discussed. Finally, we summarize the research findings and point out future research directions. More resources are available at https://github.com/M-3LAB/awesome-industrial-anomaly-detection.
1 Introduction
The survey reviews deep learning-based industrial image anomaly detection and examines how methods, datasets, metrics, supervision, and architectures relate to manufacturing needs. It also identifies practical challenges and future research directions for deploying IAD in industry.
- Motivation: Industrial IAD aims to identify product defects, whose severity can affect product price and determine whether products are discarded.
- Motivation: Human inspection is vulnerable to fatigue-related false positives, health problems, and high labor costs.
- Motivation: Although many deep learning IAD methods exceed 97% accuracy, real-world industrial use still presents substantial unresolved problems.
- Contributions: The survey compares neural network architectures under varying supervision and reviews algorithms within industrial manufacturing settings.
- Challenges: Industrial datasets should come from manufacturing lines, while open datasets often synthesize anomalies from normal products and may not reflect production defects.
- Contributions: The survey categorizes methods, analyzes their advantages and disadvantages, reviews datasets and metrics, and outlines future research directions.
2 Unsupervised Anomaly Detection
Unsupervised industrial anomaly detection generally trains on normal samples because abnormal data are costly to collect. Existing methods are commonly organized into feature-embedding and reconstruction-based categories, with feature embedding currently more prevalent.
- Unsupervised anomaly detection typically uses only normal samples for training and both normal and abnormal samples for testing.
- Industrial image anomaly detection methods are commonly divided into feature-embedding and reconstruction-based approaches.
- Feature-embedding methods currently account for more anomaly-detection techniques than reconstruction-based methods.
2.1 Feature Embedding based Methods
Feature-embedding methods organize industrial IAD around teacher-student, one-class classification, and related representation-learning paradigms. Their trade-offs involve pretrained-model dependence, synthetic abnormal data, feature boundaries, and localization strategies.
- Teacher-Student Architecture: Teacher-student methods depend on pretrained models such as ResNet, VGG, and EfficientNet, making teacher-model selection crucial.
- Teacher-Student Architecture: Teacher-student models transfer normal-feature extraction from a fixed pretrained teacher to a student and compare their features during inference.Normal samples produce similar teacher and student features, whereas abnormal samples produce more dissimilar features.
- One-Class Classification: Synthetic abnormal samples can severely compromise one-class performance when their quality is poor, while most methods rely on SVDD or cross-entropy losses.
- One-Class Classification: One-class classification learns a hypersphere around normal features and classifies test samples by their relative position, often using artificially generated anomalies.
- One-Class Classification: CutPaste augments one-class training by cutting and pasting normal-image regions, while MemSeg stores normal features and introduces foreground anomalies from external datasets.
2.1.3 Distribution Map
Distribution-map methods map pretrained normal-image features to a target distribution, commonly a multivariate Gaussian, and detect anomalies through deviations from that distribution. Normalizing flows dominate this paradigm but require suitable mapping objectives.
- Distribution Map: Distribution-map performance depends on selecting a suitable mapping objective and mapping method.
- Distribution Map: Distribution-map methods map features into desired distributions rather than directly finding feature boundaries as one-class classification methods do.
- Distribution Map: A common pipeline extracts normal features with a pretrained network, maps them to a multivariate Gaussian, and detects anomalies by distributional deviation.
- Distribution Map: Normalizing flows construct complex distributions through a series of invertible mappings and predominate among distribution-map methods.
- Distribution Map: DifferNet introduced normalizing flows for industrial image anomaly detection, while later variants add cross-convolution, positional encoding, or global-local distribution modeling.
2.1.4 Memory Bank
Memory-bank methods compare test features with stored normal features and typically require little training, whereas reconstruction-based methods rebuild images or features to expose anomalies. The paradigms trade pretrained-feature strength against reconstruction flexibility and semantic capability.
- Memory Bank: Memory-bank methods often require no loss function for training, can be constructed quickly, and are described as currently most effective for IAD tasks.
- Memory Bank: Memory-bank inference compares test-image features with stored normal features, using spatial distance to estimate abnormality.
- Memory Bank: PatchCore coreset-subsamples its memory bank to reduce inference cost while maintaining performance, then uses nearest-neighbor feature distances for detection.
- Reconstruction Based Methods: Reconstruction-based methods self-train encoders and decoders, reducing reliance on pretrained models but weakening image-classification capability because high-level semantic features are not extracted.
- Reconstruction Based Methods: Training reconstruction networks from scratch without robust pretrained models results in inferior performance compared with image-level feature embedding.
- Reconstruction Based Methods: DRAEM synthesizes abnormal images, reconstructs them as normal, and feeds original and reconstructed images to a segmentation network for anomaly-region prediction.
2.1.6 Generative Adversarial Networks
Generative approaches reconstruct or synthesize images to support industrial anomaly detection, with GANs offering scene-specific advantages but less stability than autoencoders. Transformers and diffusion models extend reconstruction by modeling global context or large anomaly regions.
- GAN reconstruction is less stable than AE reconstruction, although its discriminant network performs better in some scenes.
- SCADN masks image portions, reconstructs them with GANs, and detects anomalies by comparing inputs with reconstructions.
- Transformers model global information and can reconstruct image patches or pretrained features for anomaly localization.
- AnoDDPM applies diffusion models to industrial anomaly detection and captures large anomaly regions with simplex noise without requiring large datasets.
3 Supervised Anomaly Detection
Supervised and semi-supervised anomaly detection methods use abnormal samples, synthetic defects, or learning strategies designed for severe class imbalance. The section also covers fully and weakly supervised detection and segmentation models.
- Abnormal data is difficult to collect, motivating training with small numbers of abnormal samples and many normal samples.
- Semi-supervised methods address data imbalance by exploiting loss changes or synthesizing abnormal samples alongside normal data.
- Some studies overlook the unbalanced distribution of normal and abnormal samples and rely mainly on abnormal samples for supervised training.
- Fully supervised methods adapt object detection and semantic segmentation architectures to detect and segment industrial defects.
4 Industrial Manufacturing Setting
Industrial manufacturing settings emphasize data-efficient, robust, synthetic, and three-dimensional anomaly detection. These directions reduce labeling burdens or add information, but current datasets and methods still limit direct deployment in production.
- Few-Shot Anomaly Detection: Few-shot anomaly detection can reduce industrial data collection and annotation costs, while investigating which data is most valuable.
- Few-Shot Anomaly Detection: Zero-shot anomaly detection uses large-model generalization without training, and MAEDAY and WinCLIP demonstrate it as a promising direction.
- Noisy Anomaly Detection: Noisy-learning methods target labeling errors and false detection, but most have not been verified on real industrial image datasets.
- 3D Anomaly Detection: 3D anomaly detection provides spatial information beyond RGB, yet current methods mainly use RGB-D images while real manufacturing datasets consist of point clouds.
- Anomaly Synthesis: Anomaly synthesis can augment limited data and reduce collection and labeling costs by generating more abnormal training samples.
5 Datasets and Metrics
Industrial anomaly detection depends on suitable datasets and metrics, but many datasets are not collected from production lines and standard metrics may underweight tiny defects. The survey reviews dataset comparisons, metric conventions, and representative-method visualizations.
- Datasets: Industrial photographs are difficult to obtain, which hampers progress in industrial image anomaly detection.
- Datasets: Most industrial anomaly datasets are not generated on real production lines, limiting their alignment with manufacturing demands.
- Metrics: Metric tables distinguish whether higher or lower values indicate better performance and describe each metric’s relevance to anomaly detection.
- Metrics: Common metrics such as F1, AU-ROC, and AU-PR may not correspond well to IAD performance because anomalies are tiny and require greater weighting.
6 Total Performance Analysis
The survey finds that strong image-level anomaly classification does not guarantee strong pixel-level segmentation. Its analysis indicates different architecture choices, metrics, and supervision strategies matter across tasks.
- Similar image-classification performance can conceal significant differences in pixel-level segmentation.Figure 7 supports the statistical results reported for MVTec AD.
- Memory bank-based approaches are most effective for image-level anomaly detection but inadequate for pixel-level anomalies.
- Ensemble learning can dramatically improve state-of-the-art anomaly detection performance.
- Segmentation modules benefit both classification and segmentation tasks, even when trained with artificially generated anomalies.MemSeg and DRAEM outperform methods without segmentation modules; artificial supervision is usually inferior to real supervision.
- AU-PR is more valuable than AU-ROC for segmentation tasks, with reconstruction-based methods outperforming others on pixel AU-PR.
7 Future Directions
Future work targets industrial deployment constraints, including insufficient modalities, streaming inference demands, feature drift, computation costs, and limited supervision. The survey also calls for unified industrial and medical anomaly-detection frameworks.
- Multi-modal IAD datasets should combine RGB images with information such as X-ray and ultrasound for assembly-line anomaly detection.
- IAD methods need faster inference for sequentially arriving production-line samples, motivating architectures that trade off speed and accuracy.
- Industrial-image pre-trained models are needed because ImageNet pre-training can produce feature drift.
- Lightweight but efficient anomaly-detection models are needed because computation costs limit production-end deployment.
- A unified industrial and medical anomaly-detection framework requires better baselines and benchmarks despite similarities in data and experimental settings.
8 Conclusions
The paper presents a literature review of industrial image anomaly detection organized around supervision, neural architectures, datasets, metrics, and manufacturing settings. It also analyzes architecture effects and identifies future research directions.
- The survey reviews industrial image anomaly detection across supervision levels, neural architectures, dataset properties, and evaluation metrics.
- It characterizes a promising industrial-manufacturing setting and reviews current IAD algorithms under that setting.
- The paper investigates which neural-network architecture designs can considerably improve anomaly-detection performance.
- The survey highlights future research directions for image anomaly detection.