brycewang-stanford Industrial Anomaly Detection Papers Guide
Overview
Industrial anomaly detection uses machine learning to identify defects, faults, and anomalies in manufacturing and quality inspection. This curated collection covers methods from reconstruction-based (autoencoders) to memory-bank approaches (PatchCore), normalizing flows, knowledge distillation, and foundation model-based detectors. Includes benchmark datasets, evaluation metrics, and real-world deployment considerations.
Method Taxonomy
Anomaly Detection Methods
├── Reconstruction-based
│ ├── Autoencoder (AE, VAE)
│ ├── GAN-based (AnoGAN, GANomaly)
│ └── Diffusion-based (AnoDDPM)
├── Embedding-based
│ ├── Memory bank (PatchCore, PaDiM)
│ ├── Knowledge distillation (STPM, RD4AD)
│ └── Self-supervised (CutPaste, DRAEM)
├── Normalizing Flows
│ ├── FastFlow, CFLOW-AD, CS-Flow
│ └── DifferNet
├── Foundation Models
│ ├── CLIP-based (WinCLIP, AnomalyCLIP)
│ ├── SAM-based (GroundedSAM-AD)
│ └── Vision-language (AnomalyGPT)
└── 3D Anomaly Detection
├── Point cloud methods
└── Multi-modal (RGB + 3D)
Key Methods
| Method | Year | Approach | MVTec AUROC |
|---|---|---|---|
| PatchCore | 2022 | Memory bank | 99.1% |
| PaDiM | 2021 | Multivariate Gaussian | 97.9% |
| RD4AD | 2022 | Knowledge distillation | 98.5% |
| FastFlow | 2022 | Normalizing flow | 99.4% |
| SimpleNet | 2023 | Feature adaptation | 99.6% |
| WinCLIP | 2023 | CLIP zero-shot | 95.2% |
| AnomalyGPT | 2024 | Vision-language | 96.3% |
Benchmark Datasets
benchmarks = {
"MVTec AD": {
"categories": 15,
"images": 5354,
"type": "Product/texture defects",
"annotation": "Pixel-level masks",
},
"MVTec 3D-AD": {
"categories": 10,
"images": 4147,
"type": "3D point cloud + RGB",
},
"VisA": {
"categories": 12,
"images": 10821,
"type": "Complex structure anomalies",
},
"BTAD": {
"categories": 3,
"images": 2830,
"type": "Industrial body/surface",
},
"MPDD": {
"categories": 6,
"images": 1064,
"type": "Metal parts defects",
},
}
for name, info in benchmarks.items():
print(f"{name}: {info['categories']} categories, "
f"{info['images']} images — {info['type']}")
Quick Implementation
# PatchCore-style anomaly detection
from anomalib.data import MVTec
from anomalib.models import Patchcore
from anomalib.engine import Engine
# Setup dataset
datamodule = MVTec(
root="./datasets/MVTec",
category="bottle",
image_size=(256, 256),
)
# Initialize model
model = Patchcore(
backbone="wide_resnet50_2",
layers=["layer2", "layer3"],
coreset_sampling_ratio=0.1,
)
# Train and test
engine = Engine()
engine.fit(model=model, datamodule=datamodule)
results = engine.test(model=model, datamodule=datamodule)
print(f"Image AUROC: {results[0]['image_AUROC']:.3f}")
print(f"Pixel AUROC: {results[0]['pixel_AUROC']:.3f}")
Evaluation Metrics
# Standard anomaly detection metrics
from sklearn.metrics import roc_auc_score
import numpy as np
# Image-level: Is this image anomalous?
image_auroc = roc_auc_score(y_true_image, y_score_image)
# Pixel-level: Where is the anomaly?
pixel_auroc = roc_auc_score(
y_true_pixel.flatten(), y_score_pixel.flatten()
)
# PRO metric: Per-Region Overlap
# Better than pixel AUROC for small anomalies
# Weights each connected anomaly region equally
Research Frontiers
### Active Directions (2024-2025)
1. **Zero/few-shot AD** — Detect anomalies without normal training data
2. **Multi-class unified** — One model for all product categories
3. **Foundation model AD** — CLIP/SAM/LLM-based detection
4. **Logical anomalies** — Structural/contextual defects
5. **Continual learning** — Adapt to new defect types
6. **3D anomaly detection** — Point cloud and multi-modal
7. **Real-time deployment** — Edge device optimization
Use Cases
- Manufacturing QC: Automated visual inspection pipelines
- Research benchmarking: Compare new methods on standard datasets
- Survey writing: Comprehensive method taxonomy and comparison
- Course teaching: Industrial AI and computer vision curricula
- Defect analysis: Understanding failure modes and patterns