Detecting defects at scale.
From MVTec AD to weld X-rays, this is the register of models powering Industry 4.0 quality control. One-class learning, pixel-level localisation, zero-defect production lines.
The shape of the task is narrow: a fixed set of datasets, AUROC as the scoreboard, three families of approaches. Progress now shows up in tail metrics — PRO, pixel AUROC, and inference speed on factory hardware.
What a number actually says.
Area under the ROC curve
The gold standard for this task. Measures the trade-off between true-positive and false-positive rate across all thresholds. 99%+ is production-ready; below 97% and the operator is in the loop.
Per-pixel localisation
Same curve, evaluated at pixel resolution. Separates models that can draw the defect from models that only flag the image. Drops sharply on small or thin defects.
Per-region overlap
Averages IoU across connected anomaly regions. More sensitive than pixel AUROC to small defects — a single missed crack hurts PRO much more than it hurts the pixel score.
Architectures, ranked.
| # | Model | MVTec AD | VisA | Architecture |
|---|---|---|---|---|
| 01 | SimpleNet Research | 99.6% | 95.5% | Feature Adapter + Discriminator |
| 02 | EfficientAD Research | 99.1% | 94.8% | Student-Teacher + Autoencoder |
| 03 | PatchCore Amazon | 99.1% | 92.1% | Memory Bank + k-NN |
| 04 | FastFlow Research | 99.4% | — | Normalizing Flow + CNN |
| 05 | Reverse Distillation Research | 98.5% | — | Student-Teacher Reverse |
| 06 | CFLOW-AD Research | 98.3% | — | Conditional Normalizing Flow |
| 07 | DRAEM Research | 98.0% | — | Reconstructive + Discriminative |
| 08 | PaDiM Research | 97.9% | — | Multivariate Gaussian + Pretrained CNN |
Higher AUROC is better. MVTec AD image-level score drives the ranking; VisA used as a tiebreaker.
The families that dominate.
- 01
Memory Bank
Store embeddings of normal samples, flag nearest-neighbour distance. No training after feature extraction.
PatchCore · PaDiM · SPADE - 02
Normalising Flows
Learn the distribution of normal features; anomalies have low likelihood. Precise localisation.
FastFlow · CFLOW-AD · CS-Flow - 03
Student–Teacher
Student network mimics teacher on normal data; discrepancy is the anomaly score. 600+ FPS on factory hardware.
EfficientAD · Reverse Distillation
Pipeline integrity. Structural safety.
Surface inspection and X-ray radiography are where anomaly detection earns its keep: defects invisible to a human inspector, at line speed, with an auditable score per image.
- · Porosity
- · Slag inclusion
- · Lack of fusion
- · Cracks
- · Rolled-in scale
- · Pitted surface
YOLOv8 fine-tuned · 30+ FPS
Kaggle competition winning solution
6 defect types · hot-rolled steel
The register.
MVTec AD
20195,354 high-resolution images across 15 object and texture categories. The gold standard for industrial anomaly detection with pixel-level annotations.
MVTec 3D-AD
20214,147 high-resolution 3D point cloud scans and RGB images across 10 categories. First comprehensive 3D anomaly detection benchmark.
VisA
202210,821 high-resolution images across 12 objects with complex structures. More challenging than MVTec with realistic industrial scenarios.
KolektorSDD2
20213,335 images of electrical commutators with surface defects. Real industrial dataset with challenging small defects.
Weld Defect X-Ray
2021Radiographic images of welds with annotations for porosity, slag inclusion, lack of fusion, cracks. Critical for pipeline and structural inspection.
Severstal Steel Defect
201912,568 steel sheet images with 4 defect classes from Kaggle competition. Real industrial data from major steel producer.
Run your own benchmarks.
We’re looking for GPU compute partners to run comprehensive anomaly-detection benchmarks on industrial datasets — including datasets we can’t publish publicly.