This is a system capable of anomaly detection for two distinct products from the MVTec Anomaly Detection dataset, screw and metal nut.
The whole pipeline is implemented via pipeline.py.
The machine learning model is based on EfficientAD. https://arxiv.org/abs/2303.14535. The code is built based on https://github.com/nelson1425/EfficientAD.
By using Lightweight Student–Teacher + Autoencoder architecture for anomalies detection and Patch description networks (PDN) for feature extraction. The model enables a fast handling of anomalies with low error rate, making it a perfect choise for abnomaly detection in manufacturing industry.
-
Types and possible locations of defects are unknown --> Use Student–Teacher structure so that it perform well even trained only on normal images.
-
Industrial settings requires strict runtime limits --> Reduce computational cost by drastically reducing depth for feature extractor, performing down-sampling early, and using light-weight S-T structure.
-
Violations of logical constraints regarding the position, size, arrangement, etc. --> Use Autoencoder to detect logical anomalies.
| Product | Model | AU-ROC |
|---|---|---|
| screw | EfficientAD-S | 96.8 |
| screw | EfficientAD-M | 97.4 |
| metal nut | EfficientAD-S | 98.7 |
| metal nut | EfficientAD-M | 99.3 |
| Model | GPU | Latency |
|---|---|---|
| EfficientAD-S | Quadro RTX 6000 | 4.0 ms |
| EfficientAD-M | Quadro RTX 6000 | 5.9 ms |
Python==3.10
numpy==1.18.5
torch==1.13.0
torchvision==0.14.0
scikit-learn==1.2.2
tifffile==2021.7.30
tqdm==4.56.0
Pillow==7.0.0
scipy==1.7.1
tabulate==0.8.7
Download dataset (if you already have downloaded then set path to dataset (--mvtec_ad_path) when calling efficientad.py).
mkdir mvtec_anomaly_detection
cd mvtec_anomaly_detection
wget https://www.mydrive.ch/shares/38536/3830184030e49fe74747669442f0f282/download/420938113-1629952094/mvtec_anomaly_detection.tar.xz
tar -xvf mvtec_anomaly_detection.tar.xz
cd ..
Training and inference for screw and metal nut:
python mvtec_ad_training/efficientad_2objects.py --dataset mvtec_ad
Training with EfficientAD-M for two kinds of images:
python mvtec_ad_training/efficientad_2objects.py --model_size medium --weights models/teacher_medium.pth --dataset mvtec_ad
Evaluation with Mvtec evaluation code:
python mvtec_ad_evaluation/evaluate_experiment.py --anomaly_maps_dir './output/4/anomaly_maps/mvtec_ad/' --output_dir './output/4/metrics/mvtec_ad/' --evaluated_objects screw
Anomaly detection pipeline for one image:
python pipeline.py --sample_path './mvtec_anomaly_detection/metal_nut/test/scratch/000.png'