Skip to content

Latest commit

 

History

History
228 lines (179 loc) · 10.7 KB

README.md

File metadata and controls

228 lines (179 loc) · 10.7 KB
Raw

SuperpixelGridMasks

SuperpixelGridMasks data augmentation

What is SuperpixelGridMasks?

SuperpixelGridMasks is a semantic data augmentation approach which permits to generate various complementary images from original sensor-based data of varied natures e.g. X-Ray scans, vehicular scenes, people images (see data samples). This approach allows to increase the size of your image-based training datasets towards expecting better performances in your analysis tasks. Experiments have shown that the approach can be efficient for image classification tasks. This work is currently under review. Once reviewed, source codes and complementary results will be publicly made available online.

For more details about this work:

Update December 19, 2022:

Karim Hammoudi, Adnane Cabani, Bouthaina Slika, Halim Benhabiles, Fadi Dornaika, and Mahmoud Melkemi. SuperpixelGridMasks Data Augmentation: Application to Precision Health and Other Real-world Data, Journal of Healthcare Informatics Research, 2023. https://doi.org/10.1007/s41666-022-00122-1

Update April 20, 2022 (preprint):

Karim Hammoudi, Adnane Cabani, Bouthaina Slika, Halim Benhabiles, Fadi Dornaika, and Mahmoud Melkemi. SuperpixelGridCut, SuperpixelGridMean and SuperpixelGridMix Data Augmentation, arXiv:2204.08458, 2022. https://doi.org/10.48550/arxiv.2204.08458

Augmentation examples (medical, human and environmental data)

Samples of augmented X-Ray scans (generated from the Kaggle dataset Chest X-ray Images (Pneumonia))

X-ray scans (q=1000, r=0.4, with boundaries)

X-Ray
Image 1
 X-Ray
SuperpixelGridCut
 X-Ray
SuperpixelGridMean
 X-Ray
Image 2
 X-Ray
SuperpixelGridMix

X-ray scans (q=1000, r=0.4, without boundaries)

X-Ray
Image 1
 X-Ray
SuperpixelGridCut
 X-Ray
SuperpixelGridMean
 X-Ray
Image 2
 X-Ray
SuperpixelGridMix

Samples of augmented people images (generated from a PASCAL VOC benchmark dataset)

Pedestrian images (q=100, r=0.4, with boundaries)

Person
Image 1
 Person
SuperpixelGridCut
 Person
SuperpixelGridMean
 Person
Image 2
 Person
SuperpixelGridMix

Pedestrian images (q=100, r=0.4, without boundaries)

Person
Image 1
 Person
SuperpixelGridCut
 Person
SuperpixelGridMean
 Person
Image 2
 Person
SuperpixelGridMix

Samples of augmented vehicular images (generated from a PASCAL VOC benchmark dataset)

Vehicle images (q=200, r=0.4, with boundaries)

Car
Image 1
 Car
SuperpixelGridCut
 Car
SuperpixelGridMean
 Car
Image 2
 Car
SuperpixelGridMix

Vehicle images (q=200, r=0.4, without boundaries)

Car
Image 1
 Car
SuperpixelGridCut
 Car
SuperpixelGridMean
 Car
Image 2
 Car
SuperpixelGridMix

Classification results

Augmentation Parameters Accuracy (%)
Baseline - 57.81
Horizontal flip - 62.50
Adjust brightness delta=0.1 57.81
CutOut r=0.2 65.63
CutOut r=0.4 65.63
SuperpixelGridCut (ours) (q=100, r=0.2) 64.06 (+6.25)
SuperpixelGridMean (ours) (q=1000, r=0.4) 75.00 (+17.19)
CutMix r=0.2 67.19
CutMix r=0.4 60.94
SuperpixelGridMix (ours) (q=100, r=0.2) 75.00 (+17.19)
Classification results of our proposed data augmentation methods
and diverse ones obtained by training from scratch VGG19
with a PASCAL VOC dataset and varied parametrization.


For more details and results, you can refer to the article: https://doi.org/10.48550/arxiv.2204.08458

Team

Project leaders:

Note: project leaders equally contributed to this work.

Contributors:

Bibtex references

Karim Hammoudi, Adnane Cabani, Bouthaina Slika, Halim Benhabiles, Fadi Dornaika, and Mahmoud Melkemi. SuperpixelGridMasks Data Augmentation: Application to Precision Health and Other Real-world Data, Journal of Healthcare Informatics Research, 2023. https://doi.org/10.1007/s41666-022-00122-1

@Article{HammoudiCabani2023,
author={Hammoudi, Karim
and Cabani, Adnane
and Slika, Bouthaina
and Benhabiles, Halim
and Dornaika, Fadi
and Melkemi, Mahmoud},
title={SuperpixelGridMasks Data Augmentation: Application to Precision Health and Other Real-world Data},
journal={Journal of Healthcare Informatics Research},
year={2023},
month={Jan},
day={13},
issn={2509-498X},
doi={10.1007/s41666-022-00122-1},
url={https://doi.org/10.1007/s41666-022-00122-1}
}

Karim Hammoudi, Adnane Cabani, Bouthaina Slika, Halim Benhabiles, Fadi Dornaika and Mahmoud Melkemi. SuperpixelGridCut, SuperpixelGridMean and SuperpixelGridMix Data Augmentation, arXiv:2204.08458, 2022. https://doi.org/10.48550/arxiv.2204.08458

@misc{HammoudiCabaniSuperpixelGridMasks,
  doi = {10.48550/ARXIV.2204.08458},
  url = {https://arxiv.org/abs/2204.08458},
  author = {Hammoudi, Karim and Cabani, Adnane and Slika, Bouthaina and Benhabiles, Halim and Dornaika, Fadi and Melkemi, Mahmoud},
  keywords = {Computer Vision and Pattern Recognition (cs.CV), Artificial Intelligence (cs.AI), Information Retrieval (cs.IR), Machine Learning (cs.LG), FOS: Computer and information sciences, FOS: Computer and information sciences, I.4; I.2, 65D18, 94A08},
  title = {SuperpixelGridCut, SuperpixelGridMean and SuperpixelGridMix Data Augmentation},
  publisher = {arXiv},
  year = {2022},
  copyright = {Creative Commons Attribution Non Commercial No Derivatives 4.0 International}
}