STFPM_AnomalyDetection

​ A improved code of "Student-Teacher Feature Pyramid Matching for Anomaly Detection" for training, testing and predicting.

​ Original Reference Code(By PyTorch): https://github.com/gdwang08/STFPM

​ Unofficial Code (By PyTorch-Lightning): https://github.com/hcw-00/STPM_anomaly_detection

​ Paper Link: https://arxiv.org/abs/2103.04257v3

Change

​ This code is modified from hcw-00's STPM_anomaly_detection, which is based on PyTorch-Lightning.

1. It can generate "grayscale defect detection image" (ground-truth mask), which means the image is represented by only black and white. (By LuZikang)
2. Add an prediction module. (By YangBin)
3. It has changed from ResNet-18 to ResNet-152. You can change back or try to use ResNet-152 as Teacher & ResNet-18 as Student.

├ ─ test # models saved
│   ├ ─ bottle
│   └ ─ ......
├── train.py # model training
├── detecting.py # model predicting
├── tensorboard_run.py # training result visualization
├── requirements.txt
└── README.md

Dataset

​ Download Link: https://www.mvtec.com/company/research/datasets/mvtec-ad/downloads

​ Paper Link: https://ieeexplore.ieee.org/document/8954181

Hardware

​ It is merely for the reference of the following training result. Always not expected to get the same result data by given the same models.

Component	Description	Note
CPU	Intel(R) Xeon(R) Gold 6138 CPU @ 2.00GHz 2.00 GHz
GPU	None in this training	Better GPU, better training.
RAM	64.0 GB

Training

# torch==1.12.1 torchvision=0.13.1 opencv-python==4.5.2.52
pip install -r requirements.txt

python train.py --phase=train --dataset_path=...\mvtec_anomaly_detection --category=bottle --project_path=...\test

Main Parameters	Note
--num_epochs	epoch number for training
--lr	optimizer parameter, learning rate
--momentum	optimizer parameter
--weight_decay	optimizer parameter
--batch_size	Set less than 4 if ensure better performance
--project_path	the path of models saved Once a model are finished in training, It can't be re-training

Testing

python train.py --phase=test --dataset_path=...\mvtec_anomaly_detection --category=bottle --project_path=...\test --output_path=...\output

Predicting

python train.py --phase=predict --predict_path=...\mvtec_anomaly_detection --category=bottle --project_path=...\test

or

Revise detecting.py and run it.

Results

100 epochs training for each category actually.

Category	AUC-ROC(image)	AUC-ROC (pixel)
carpet	0.9662921348314606	0.9903472118834364
grid	0.9866332497911445	0.9871335697600007
leather	1.0	0.9945778707141713
tile	0.9949494949494949	0.9744931114416499
wood	0.9947	0.9642
bottle	1.0	0.9818796255452386
cable	0.8862443778110944	0.9194169093988306
capsule	0.8954926206621461	0.9461320936660839
hazelnut	0.9714285714285714	0.9881636024343283
meta_nut	0.8357771260997068	0.8974234615548793
pill	0.9318057828696126	0.9582640814865969
screw	0.8159458905513425	0.983182165681811
toothbrush	0.9333333333333333	0.9861350877979583
transistor	0.8533333333333334	0.8226574097914439
zipper	0.9726890756302521	0.9622075550349201
average	0.93590	0.95708

Citation

@inproceedings{wang2021student_teacher,
  title={Student-Teacher Feature Pyramid Matching for Anomaly Detection},
  author={Wang, Guodong and Han, Shumin and Ding, Errui and Huang, Di},
  booktitle={The British Machine Vision Conference (BMVC)},
  year={2021}
}