Code for the paper Adversarial Regularizer in Inverse Problems.
Inverse Problems are traditionally solved using purely model-based methods, as in variational regularization methods. We propose a new framework for applying data-driven approaches to inverse problems, using a neural network as regularization functional.
The network is trained as a critic as in WGANs, learning to discriminate between the distribution of ground truth images and the distribution of unregularized reconstruction.
Once trained, we use the learned regularization functional to solve inverse problems by minizing the associated variational functional. In the context of computed tomography we employ early stopping to obtain the best results.
Reconstructions obtained on the LIDC/IDRI dataset
Download the LIDC_IDRI and/or the BSDS500 dataset and split the raw data into two folders named Training_Data and Evaluation_Data inside your DATA_PATH folder.
The package requirements are numpy, tensorflow, scipy, scikit-image, matplotlib, pydicom, astra-toolbox, and odl. All requirements can be installed directly into the conda environment advReg via
$ conda env create -f advReg.yml
$ source activate advRegThe tensorboard loggings can be found in SAVES_PATH.
To add custom data sets, forward operators for different inverse problems or network architectures, write your custom implementation of the corresponding abstract classes in ClassFiles. Finally, set the get_Data_pip, get_model or get_network in your experiments accordingly.