/iRestNet

Deep unfolded proximal interior point algorithm

Primary LanguageJupyter Notebook

iRestNet

Unfolded proximal interior point algorithm for image deblurring

  • License : GNU General Public License v3.0
  • Author : Marie-Caroline Corbineau
  • Institution : Centre de Vision Numérique, CentraleSupélec, Inria, Université Paris-Saclay
  • Email : mariecaroline.corbineau@gmail.com
  • Related publication: https://arxiv.org/abs/1812.04276
  • Please use the following citation:

    Carla Bertocchi, Emilie Chouzenoux, Marie-Caroline Corbineau, Jean-Christophe Pesquet and Marco Prato, `Deep unfolding of a proximal interior point method for image restoration'. To appear in Inverse Problems, doi:10.1088/1361-6420/ab460a (2019).

Installation

  1. Install miniconda.
  2. Create an environment with python version 3.6.5. 
    $ conda create -n iRestNet_env python=3.6.5
  3. Inside this environment install the following packages. 
    $ conda activate iRestNet_env
$ conda install pytorch=0.4.0 cuda80 -c pytorch
$ pip install torchvision==0.2.1 matplotlib==3.0.2 numpy==1.16.0 jupyterlab opencv-python==4.0.0.21 scipy==1.2.0 tqdm==4.29.1 scikit-image==0.14.2
  4. Use the demo notebook to test and train iRestNet models.

Files organization

  • Datasets
    • BSD500_COCO1000_train_val
      • train: contains 200 RGB training images from the Berkeley segmentation dataset (BSD500) and 1000 RGB images from the COCO dataset
      • val: contains 100 RGB validation images from the Berkeley segmentation dataset (BSD500)
    • Groundtruth: contains groundtruth test images
      • full
        • BSD500: contains 200 groundtruth test images from the Berkeley segmentation dataset
        • Flickr30: contains 30 groundtruth test images from Flickr30
      • cropped: contains cropped test images (created by the function create_testset)
    • Testsets: contains a folder for each test configuration including associated blur kernels (kernel.mat)
  • Model_files : contains iRestNet files
    • Deg3PolySolver.py: contains the cardan class used to compute the proximity operator of the logarithmic barrier
    • iRestNet.py: contains iRestNet_class, which is the main class including train and test functions
    • model.py: includes the definition of the layers in iRestNet
    • modules.py: useful functions used in iRestNet
    • tools.py: contains some useful side functions
  • PyTorch_ssim : forward and backward functions to use SSIM as the training loss, we use the code from https://github.com/Po-Hsun-Su/pytorch-ssim
  • Trainings : contains the models used in https://arxiv.org/abs/1812.04276 for every test configuration. In the demo notebook, the results of the aforementioned models on the testsets are saved in this folder.

Demo file

demo.ipynb: shows how to test and train iRestNet

Thank you for using our code, kindly report any suggestion to the corresponding author.