Anomaly Detection
Detecting defects and anomalies in manufacturing (MVTec AD, VisA).
Industrial anomaly detection identifies defects and out-of-distribution samples in manufacturing, using only normal (defect-free) training data. The MVTec benchmark drives research, with methods like PatchCore and EfficientAD achieving >99% AUROC. The unsupervised setting (learning only from normal data) distinguishes this from supervised defect classification.
History
MVTec Anomaly Detection dataset released — 15 object/texture categories with pixel-level ground truth
PaDiM uses pretrained features with multivariate Gaussian modeling per patch
PatchCore achieves 99.1% image-level AUROC on MVTec via coreset-based memory bank
FastFlow and CFlow apply normalizing flows to pretrained features for anomaly scoring
SimpleNet shows that simple feature-matching approaches are surprisingly effective
EfficientAD achieves real-time anomaly detection with student-teacher distillation
WinCLIP applies CLIP for zero-shot anomaly detection without any normal training images
AnomalyGPT and similar LMM-based approaches enable text-driven anomaly detection
MVTec LOCO extends to logical constraint anomalies (missing parts, wrong assembly)
Foundation models for anomaly detection show strong zero-shot industrial inspection
How Anomaly Detection Works
Feature Extraction
A pretrained backbone (WideResNet-50, EfficientNet) extracts patch-level features from the input image.
Normal Distribution Modeling
Features from defect-free training images define the 'normal' distribution — via memory banks (PatchCore), Gaussian models (PaDiM), or normalizing flows.
Anomaly Scoring
Test image features are compared against the normal distribution; high-distance patches are scored as anomalous.
Localization
Per-pixel or per-patch anomaly scores produce a heatmap highlighting defective regions.
Thresholding
A threshold on the anomaly score determines the binary decision (normal/defective) for production use.
Current Landscape
Anomaly detection for industrial inspection in 2025 is a mature field with practical deployment in manufacturing. PatchCore-family methods achieve near-perfect results on texture and object anomalies, while logical anomaly detection (missing parts, wrong assembly) is the active research frontier. The MVTec benchmark is saturated; MVTec LOCO and VisA provide harder evaluation. Zero-shot approaches using CLIP and LMMs are emerging, enabling deployment on new product lines without collecting training data. The key production challenge is managing false positive rates while maintaining high recall on true defects.
Key Challenges
Few-shot normal data — some production lines have very few defect-free reference images at startup
Logical anomalies — missing components or wrong assembly require structural understanding, not just texture anomaly detection
Changing conditions — lighting, camera angle, and material variations in production create false positives
Speed requirements — real-time inspection at production line speeds (>10 fps) constrains model complexity
MVTec saturation — >99% AUROC on MVTec means the benchmark no longer discriminates between methods
Quick Recommendations
Standard anomaly detection
PatchCore / EfficientAD
Best accuracy-speed tradeoff, production-proven on MVTec
Zero-shot detection
WinCLIP / AnomalyGPT
No normal training data needed — useful for new product lines
Logical anomaly detection
GCAD / AST (trained on MVTec LOCO)
Handles structural and assembly anomalies beyond texture defects
Real-time production
EfficientAD / FastFlow
Sub-100ms inference suitable for production line speeds
What's Next
The frontier is continual anomaly detection — systems that adapt to slowly changing production conditions without forgetting what constitutes a defect. Expect multimodal inspection (combining vision with 3D scanning and sensor data), and foundation models for industrial inspection pretrained on diverse manufacturing data.
Benchmarks & SOTA
MVTec AD
MVTec Anomaly Detection Dataset
5,354 high-resolution images across 15 object and texture categories. The gold standard for industrial anomaly detection with pixel-level annotations.
State of the Art
SimpleNet
Research
99.6
auroc
MVTec 3D-AD
MVTec 3D Anomaly Detection Dataset
4,147 high-resolution 3D point cloud scans and RGB images across 10 categories. First comprehensive 3D anomaly detection benchmark.
State of the Art
CMDR-IAD
Research
97.3
auroc
KolektorSDD2
Kolektor Surface Defect Dataset 2
3,335 images of electrical commutators with surface defects. Real industrial dataset with challenging small defects.
State of the Art
SuperSimpleNet
ViCoS Lab Ljubljana
97.4
ap
VisA
Visual Anomaly Dataset
10,821 high-resolution images across 12 objects with complex structures. More challenging than MVTec with realistic industrial scenarios.
State of the Art
SimpleNet
Research
95.5
auroc
NEU-DET
NEU Surface Defect Database
1,800 grayscale images of hot-rolled steel strip with 6 defect types: rolled-in scale, patches, crazing, pitted surface, inclusion, scratches.
State of the Art
DefectDet (ResNet)
Research
78.4
map
Severstal Steel Defect
Severstal Steel Defect Detection
12,568 steel sheet images with 4 defect classes from Kaggle competition. Real industrial data from major steel producer.
State of the Art
YOLOv8 (Weld Detection)
Ultralytics
91.2
dice
Weld Defect X-Ray
X-Ray Weld Defect Detection Dataset
Radiographic images of welds with annotations for porosity, slag inclusion, lack of fusion, cracks. Critical for pipeline and structural inspection.
State of the Art
YOLOv8 (Weld Detection)
Ultralytics
87.3
map
Related Tasks
Something wrong or missing?
Help keep Anomaly Detection benchmarks accurate. Report outdated results, missing benchmarks, or errors.