Industrial anomoly detection with deep learning

This repo serves the source code of the Smart Manufacturing Lab project at Politecnico di Milano.

Setting up the environment

Ensure uv is installed, then:

$ uv sync
$ uv run ipython kernel install --user --env VIRTUAL_ENV $(pwd)/.venv --name=smlab

You can check JupyterLab path using:

$ uv run --with jupyter jupyter --paths

Start the JupyterLab server:

$ uv run --with jupyter jupyter lab

The web interface should automatically launch in your browser.

Models

Model	Architecture	Training
PaDiM	Frozen ResNet-18; multivariate Gaussian per spatial position	1 epoch (single-pass feature collection)
PatchCore	Frozen ResNet-18; greedy coreset memory bank (1% sampling)	1 epoch (single-pass feature collection)
EfficientAD	Lightweight PDN teacher-student + autoencoder branch	70,000 steps, Adam
RD	Frozen Wide ResNet-50-2 encoder + one-class bottleneck + mirrored student decoder	200 epochs, Adam lr=0.005

Results (metal_nut, MVTec AD)

Anomalib Engine Metrics

Model	Image AUROC	Image F1	Pixel AUROC	Pixel F1
PaDiM	0.979	0.968	0.969	0.731
PatchCore	1.000	0.989	0.986	0.825
EfficientAD	1.000	0.989	0.981	0.797
RD	1.000	0.989	0.973	0.797

Binary Classification Report (threshold = 0.5)

Model	Accuracy	Normal Prec.	Normal Rec.	Normal F1	Anomaly F1
PaDiM	0.95	0.90	0.82	0.86	0.97
PatchCore	0.98	0.95	0.95	0.95	0.99
EfficientAD	0.95	0.81	0.95	0.88	0.97
RD	0.99	0.96	1.00	0.98	0.99

Qualitative Results

License

MIT