Unsupervised Anomaly Detection

1. Installation

• Requirement: Python 3.10

git clone https://github.com/lamnguyenvu98/unsupervised-anomaly-detection.git

cd unsupervised-anomaly-detection

python -m pip install .

2. Training

2.1 Dataset structure

2.1.1 Structure for DFR and Padim

├── transistor (root)
│   ├── train
│   |    └── good
│   |        ├── image1.png
│   |        ├── image2.png
│   |        ├── ...
│   |        └── imageN.png
│   └── test
│        ├── good
│        |   ├── image1.png
│        |   ├── image2.png
│        |   ├── ...
│        |   └── imageN.png
│        └── anomaly
│            ├── image1.png
│            ├── image2.png
│            ├── ...
│            └── imageN.png

2.1.2 Structure for RegAD

├── mvtec-ad (root)
|       ├── transistor
│       |    ├── train
│       |    |    └── good
│       |    |        ├── image1.png
│       |    |        ├── image2.png
│       |    |        ├── ...
│       |    |        └── imageN.png
│       |    └── test
│       |        ├── good
│       |        |    ├── image1.png
│       |        |    ├── image2.png
│       |        |    ├── ...
│       |        |    └── imageN.png
│       |        └── anomaly
│       |            ├── image1.png
│       |            ├── image2.png
│       |            ├── ...
│       |            └── imageN.png
|       ├── bottle
│       |    ├── train
│       |    |    └── good
│       |    |        ├── image1.png
│       |    |        ├── image2.png
│       |    |        ├── ...
│       |    |        └── imageN.png
│       |    └── test
│       |        ├── good
│       |        |    ├── image1.png
│       |        |    ├── image2.png
│       |        |    ├── ...
│       |        |    └── imageN.png
│       |        └── anomaly
│       |            ├── image1.png
│       |            ├── image2.png
│       |            ├── ...
│       |            └── imageN.png
|       ├── ...

2.2 Modify config file (configs/*.yaml)

There are some parameters need to be clarified:

• DATA_DIR: root directory of dataset

• SAVE_DIR: directory where to save model checkpoints

• CHECKPOINT_PATH: path to model checkpoint for inference or resuming training process

2.2.1 Padim

• REDUCE_FEATURES: number of features reduce to

2.2.2 DFR

• NUM_LAYERS: number of backbone layers to get feature maps and aggregate them

2.2.3 RegAD

• BACKBONE: name of backbone for feature extraction. Options: "resnet18", "resnet34", "resnet50", "resnet101", "resnet152".

• STN_MODE: Default is "rotation_scale". Other options: "affine", "translation", "rotation", "scale", "shear", "translation_scale", "rotation_translation", "rotation_translation_scale".

• N_SHOT: number of support set (4, 8, 16,...)

• N_TEST: number of rounds to evaluate model

• TRAIN_DATA_DIR: root directory of dataset which contain multiple classes (screw, bottle, transistor,...)

• TEST_DATA_DIR: directory of an object for model evaluation, this object shouldn't be included in TRAIN_DATA_DIR. If it's inside TRAIN_DATA_DIR, set IGNORE_CLASS to name of that object.

• IGNORE_CLASS: not include data of this class during training (if the folder of this class is inside root folder)

Example:

TRAIN_DATA_DIR: "/content/mvtec-ad"
TEST_DATA_DIR: "/content/mvtec-ad/transistor"
IGNORE_CLASS: "transistor"

transistor class was used for model evaluation. But it was also inside TRAIN_DATA_DIR. So that, setting IGNORE_CLASS to "transistor" helped model ignored this class during training.

2.3 Train model

2.3.1 RegAD

python train/train_regad.py --config configs/regad_config.yaml

2.3.2 Padim

python train/train_padim.py --config configs/padim_config.yaml

2.3.3 DFR

python train/train_dfr.py --config configs/dfr_config.yaml

Acknowledgement

Original repo RegAD: https://github.com/MediaBrain-SJTU/RegAD

@inproceedings{huang2022regad,
  title={Registration based Few-Shot Anomaly Detection}
  author={Huang, Chaoqin and Guan, Haoyan and Jiang, Aofan and Zhang, Ya and Spratlin, Michael and Wang, Yanfeng},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2022}
}

Original repo DFR: https://github.com/YoungGod/DFR

@misc{yang2020dfr,
      title={DFR: Deep Feature Reconstruction for Unsupervised Anomaly Segmentation}, 
      author={Jie Yang and Yong Shi and Zhiquan Qi},
      year={2020},
      eprint={2012.07122},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Original repo Padim: https://github.com/xiahaifeng1995/PaDiM-Anomaly-Detection-Localization-master