Industrial Inspection
Building quality control systems? Benchmark anomaly detection, defect classification, and automated visual inspection for manufacturing.
Industrial visual inspection uses ML to detect surface defects, anomalies, and quality issues in manufacturing. MVTec AD established the benchmark standard, and anomaly detection models now achieve near-human accuracy on controlled inspection tasks while struggling with domain shift across production lines.
State of the Field (2025)
- MVTec AD benchmark is near-saturated: top methods (PatchCore, EfficientAD, SimpleNet) achieve 99%+ image-level AUROC on most categories, shifting focus to pixel-level segmentation and few-shot settings
- Memory bank approaches (PatchCore) and student-teacher distillation (EfficientAD) dominate unsupervised anomaly detection, requiring only normal samples for training
- Real-world deployment gap remains significant: models achieving 99% on MVTec AD often drop to 85-90% on actual production data due to lighting variation, camera drift, and novel defect types
- Foundation model adaptation (CLIP-based anomaly detection, WinCLIP, AnomalyCLIP) enables zero-shot and few-shot defect detection without task-specific training data
Quick Recommendations
Unsupervised anomaly detection (standard manufacturing)
PatchCore or EfficientAD
PatchCore provides best accuracy with simple memory bank architecture. EfficientAD offers 4-10x faster inference for real-time conveyor belt inspection. Both need only normal samples.
Zero-shot defect detection (new product lines)
WinCLIP or AnomalyCLIP
CLIP-based methods detect anomalies without any training images from the target domain. Critical for low-volume manufacturing where collecting defect examples is impractical.
Pixel-level defect segmentation
PaDiM or FastFlow
Probabilistic patch-level modeling provides precise defect localization maps. FastFlow adds normalizing flow for better boundary delineation on textured surfaces.
Edge deployment on inspection hardware
EfficientAD with TensorRT optimization
Sub-10ms inference on NVIDIA Jetson. Student-teacher architecture compresses well. Meets real-time requirements for high-speed production lines.
Tasks & Benchmarks
Anomaly Detection
Detecting defects and anomalies in manufacturing (MVTec AD, VisA).
Steel Defect Detection
Detecting defects in steel production: rolled-in scale, patches, pitting.
Surface Defect Detection
Detecting scratches, dents, and surface imperfections on materials.
Weld Inspection
Detecting weld defects: porosity, cracks, lack of fusion, slag inclusion.
Show all datasets and SOTA results
Anomaly Detection
Steel Defect Detection
Surface Defect Detection
Weld Inspection
Honest Takes
MVTec AD scores are meaningless for production
Every new paper claims 99.5% AUROC on MVTec. In practice, factory lighting changes, camera positions shift, and novel defect types appear. The gap between benchmark and production performance is 10-15%. Always validate on your actual production data, not academic datasets.
Unsupervised detection is the right default
Collecting labeled defect datasets is expensive and incomplete - you can never catalog every possible defect type. Train on normal samples only and flag deviations. This approach generalizes to novel defects that supervised classifiers would miss entirely.
The real bottleneck is integration, not algorithms
Most manufacturing AI failures are not model accuracy problems. They are camera placement issues, lighting inconsistency, data pipeline failures, and lack of operator trust. Spend 70% of your budget on robust imaging setup and integration, 30% on the model.
Get notified when these results update
New models drop weekly. We track them so you don't have to.