AnomalyHop: An SSL-based Image Anomaly Localization Method

Implementation for our paper. The paper can be found at AnomalyHop.

• Training in two minutes for each class
• No Neural Network / No Pre-training
• Successive Subspace Learning

1. Install Requirements / Environmen

We provide the environment setup information for virtual environment (anacodna). The environment can be easily install with (conda required):

conda env create -f environment.yml
conda activate AnomalyHop

2. Dataset Preparation

MVTec AD datasets can be downloaded from: MVTec website

It can also be downloaded by using our following commands:

cd datasets
wget ftp://guest:[email protected]/mvtec_anomaly_detection/mvtec_anomaly_detection.tar.xz
tar -xf mvtec_anomaly_detection.tar.xz
cd ..

3. Reproduce our result

An example for carpet class:

python ./src/main.py --kernel 7 6 3 2 4 --num_comp 4 4 4 4 4 --layer_of_use 1 2 3 4 5 --distance_measure glo_gaussian --hop_weights 0.2 0.2 0.4 0.5 0.1 --class_names carpet

To reproduce all results from the paper:

chmod +x run.sh
./run.sh

4. Our performance

• Anomaly Localization (ROCAUC)

MvTec AD	AnomalyHop (ours)
Carpet	0.942
Grid	0.984
Leather	0.991
Tile	0.932
Wood	0.903
Bottle	0.975
Cable	0.904
Capsule	0.965
Hazelnut	0.971
Metal nut	0.956
Pill	0.970
Screw	0.960
Toothbrush	0.982
Transistor	0.981
Zipper	0.966
All classes	0.959

5. Visualization results

• Examples from cable, capsule and wood classes.

• Examples from grid class.

If you want to save the figures, just uncomment the following line in main.py

plot_fig(test_imgs, scores_final, gt_mask_list, threshold, save_dir, class_name)

6. Citation

If you find our model is useful in your research, please consider cite our paper: AnomalyHop: An SSL-based Image Anomaly Localization Method:

@article{anomalyhop,
  title={{AnomalyHop}: An SSL-based Image Anomaly Localization Method}},
  author={Zhang, Kaitai and Wang, Bin and Wang, Wei and Sohrab, Fahad and Gabbouj, Moncef and Kuo, C.-C. Jay},
  journal={arXiv preprint arXiv:2105.03797},
  year={2021}
}
   
@article{anomalyhop,
  title={{AnomalyHop}: An SSL-based Image Anomaly Localization Method}},
  author={Zhang, Kaitai and Wang, Bin and Wang, Wei and Sohrab, Fahad and Gabbouj, Moncef and Kuo, C.-C. Jay},
  journal={IEEE Visual Communications and Image Processing (VCIP)},
  year={2021}
}

Contact person: Bin Wang, [email protected]

http://mcl.usc.edu/

Acknowledgement

The code is partially adapted from PaDiM