A Trainable Spectral-Spatial Sparse Coding Model for Hyperspectral Image Restoration

Official PyTorch implementation of the paper A Trainable Spectral-Spatial Sparse Coding Model for Hyperspectral Image Restoration (Neurips 2021).

[arxiv]

Installation

Developped with Python 3.8.8.

$ git clone https://github.com/inria-thoth/T3SC
$ cd T3SC && pip install -r requirements.txt

Training

To launch a training:

$ python main.py data={icvl,dcmall} noise={constant,uniform,correlated,stripes} [+noise-specific params]

Data should be downloaded automatically to data/ICVL or data/DCMall if it is not there already.

NOTE: For uniform and stripes noises, better results are obtained with Noise Adaptive Sparse Coding. To enable this feature, use model.beta=1 for both training and testing.

Examples

ICVL dataset with constant gaussian noise:

$ python main.py data=icvl noise=constant noise.params.sigma=50

Washington DC Mall dataset with band-dependant gaussian noise:

$ python main.py data=dcmall model.beta=1 noise=uniform noise.params.sigma_max=55

ICVL dataset with stripes noise:

$ python main.py data=icvl noise=stripes

Test

To test from a checkpoint:

$ python main.py mode=test data={icvl,dcmall} noise={constant,uniform,correlated,stripes} [+noise-specific params] model.ckpt=path/to/ckpt

Some pre-trained models can be found here.

Example

To test ICVL with constant noise:

$ python main.py mode=test data=icvl noise=constant noise.params.sigma=50 model.ckpt=path/to/icvl_constant_50.ckpt

Citation

If you find this work useful for your research, please cite:

@article{bodrito2021trainable,
  title={A Trainable Spectral-Spatial Sparse Coding Model for Hyperspectral Image Restoration},
  author={Bodrito, Theo and Zouaoui, Alexandre and Chanussot, Jocelyn and Mairal, Julien},
  journal={Adv. in Neural Information Processing Systems (NeurIPS)},
  year={2021}
}