Skip to content

Repository files navigation

TAB: Text-Align Anomaly Backbone Model for Industrial Inspection Tasks

arXivViewCount

Installation:

git clone git@github.com:Howeng98/TAB.git
cd TAB
conda create -n TAB
conda activate TAB
pip install -r requirements.txt

Pre-training

You can kindly modify the pre-training dataset, backbone structure, and hyperparameters in the config.py

python main.py

Anomaly Backbone Checkpoint Download

DescriptionWeights Download
MVTecAD ResNet18Download Link
IndustrialDataset ResNet18Download Link

Industrial Dataset

We propose a well-organized dataset for industrial visual inspection named Industrial Dataset. This dataset was constructed by selectively aggregating open-source datasets from global websites and official public conference workshops as references. The Industrial Dataset encompasses 30 categories, including 15 object categories and 15 texture categories, amounting to a total of 17,393 images. Notably, the Industrial dataset exclusively comprises normal images.

Dataset Structure

 Industrial Dataset
├─ Washer
│ ├─ 0001.png
│ ├─ 0002.png
│ ...
│ └─ 3241.png
│ ├─ Metal Tube
│ ├─ 001.png
│ ├─ 002.png
│ ...
│ └─ 122.png
├─ Hazelnut
...

Dataset Sources

IdxClass NameNumber of normal samplesSourceNote
1Bent Metal Stripe185linkMPDD
2Board40linkDAGM
3Bolt128linkMPDD
4Bottle746link
5Brick Tile80linkUIUC
6Capsule749link
7Chip1211linkPCB-AoI Public
8Cork216linkKTH-TIPS2
9Encapsulates Carpet728link-
10Hazelnut50linkADFI
11Hazelnut Toy34link
12Jagged Carpet80linkKylberg
13Leather600linkLeather Defect Cls
14Metal Flat54linkMPDD
15Metal ornament289linkMPDD
16Metal Pole110linkMPDD
17Metal Tube122linkMPDD
18Plaid20linkUIUC
19Rhombus Carpet80linkKylberg
20Screw59link-
21Solar Board2624linkELPV
22Sponge40linkKTH-TIPS2
23Streak Board1041linkDAGM
24Stripe Carpet336linkKylberg
25Stripe Wood1000linkDAGM
26Tile230link-
27Toy Wheel3200link-
28Upholstery20linkUIUC
29Washer3241link
30Wave Carpet80linkKylberg

NTHU CVLAB

For more details, please visit CVLAB.

License

MIT Licence

References

Citation

If you found TAB useful in your research or applications, please kindly cite using the following BibTeX:

@article{lee2023tab,
title={TAB: Text-Align Anomaly Backbone Model for Industrial Inspection Tasks},
author={Lee, Ho-Weng and Lai, Shang-Hong},
journal={arXiv preprint arXiv:2312.09480},
year={2023}
}

About

Official Pytorch Implementation for paper: TAB: Text-Align Anomaly Backbone Model for Industrial Inspection Tasks

Resources

Stars

9 stars

Watchers

1 watching

Forks

Releases

Packages

Used by

Contributors

Languages