Datasets:

Modalities:
Image
ArXiv:
Tags:
image
image
YAML Metadata Warning: empty or missing yaml metadata in repo card (https://huggingface.co/docs/hub/datasets-cards)

Description

Introduction of new dataset for unsupervised fabric defect detection This dataset aims to provide a color dataset with real industrial fabric defect gathered in a visiting machine with several industrial cameras. It has been designed with the same nomenclature as MVTEC AD dataset (https://www.mvtec.com/company/research/datasets/mvtec-ad) for unsupervised anomaly detection.

Type Total Train(Good) Test(Good) Test(Defective) Sample
type1cam1 386 272 28 86
type2cam2 257 199 19 39
type3cam1 689 588 54 47
type4cam2 229 199 19 11
type5cam2 298 199 19 80
type6cam2 291 199 19 73
type7cam2 917 711 89 117
type8cam1 868 711 89 68
type9cam2 856 721 86 49
type10cam2 871 717 90 64

Download

The dataset can be downloaded in google drive with this link : LINK

Utilisation

This dataset is designed for unsupervised anomaly detection task but can also be used for domain-generalization approach. The nomenclature is designed as :

  • category/
    • train/
      • good/
        • img1.png
        • ...
    • test/
      • anomaly/
        • img1.png
        • ...
      • good/
        • img1.png
        • ...

As in any unsupervised training, train data are defect-free. Defective samples are only in the test set.

Exemples

Exemple of defect segmentation obtained with our knowledge distillation-based method

Documentation

List of articles related to the subject of textile defect detection

Auteurs

1 University of Technology of Troyes, France

Citation

If you use this dataset, please cite

@inproceedings{Thomine_2023_Knowledge,
    author    = {Thomine, Simon and Snoussi, Hichem},
    title     = {Distillation-based fabric anomaly detection},
    booktitle = {Textile Research Journal},
    month     = {August},
    year      = {2023}
}

Licence

This project is under the MIT license MIT.

license: mit

Downloads last month
0
Edit dataset card