Deep Hyperspectral Prior: Single Image Denoising, Inpainting, Super-Resolution

Supplementary code to the paper O Sidorov, JY Hardeberg. Deep Hyperspectral Prior: Denoising, Inpainting, Super-Resolution in ICCV 2019 Workshops.

Get started

The implementation is based on original Deep Image Prior code by Dmitry Ulyanov.

The framework was modified to process hyperspectral data using 2D or 3D convolutions:

Requirements

• python = 3.6
• pytorch = 0.4
• numpy
• scipy
• matplotlib
• scikit-image
• jupyter

Prepare the data

• Input and output hyperspectral data is contained in *.mat files.
• Specify a path to the file and name of the variable to read.
  For example, if data is contained in variable image:

  file_name  = 'data/inpainting/inpainting192.mat'
  mat = scipy.io.loadmat(file_name)
  img_np = mat["image"]
  

• Use custom code or one of the *.m files located at data/%task%/ to generate *.mat file.

Run the code

• Follow one of the proposed notebook files to get the results.
  * 2D versions tend to demonstrate better accuracy.
• Try to modify parameters. Have fun.

Some results

Denoising:

Super-Resolution:

Inpainting:

Please, kindly cite the paper if you use the code!