DSGAN

A PyTorch implementation of the DSGAN model, which allows to create datasets for Super-Resolution. The model is trained in an unsupervised fashion on the HR images. It then learns to generate the appropriate LR images, which have the same corruptions as the original HR images.

Requirements

• install Pytorch according to their documentation
• install other dependencies:

conda install -y numpy pyyaml tqdm pillow
pip install tensorboardX

• download LPIPS and save it at the same level as this repository. Alternatively, change the path in line 7 of loss.py.
• For visualisation use tensorboard

Usage

Train

Add the path to your dataset to paths.yml as follows:

dataset:
  artifacts:
    hr:
      train: 'path/to/HR/train/data'
      valid: 'path/to/HR/validation/data'
    lr:
      train: 'OPTIONAL/path/to/LR/train/data'
      valid: 'OPTIONAL/path/to/LR/validation/data'    

The model can then be trained with:

python train.py --dataset dataset --artifacts artifacts

Thereby, dataset and artifacts correspond to the first and second line of the paths.yml file example, respectively. Use python train.py -h to view optional arguments.

Create dataset

You can use the trained parameters (saved in the folder checkpoints) to create a dataset from your original images. To do so, you need to first add your desired dataset location to paths.yml as follows:

datasets:
  dataset: 'path/to/desired/location'

The dataset can then be created with:

python create_dataset.py --dataset dataset --checkpoint checkpoint

Two datasets will be generated, one for the SDSR case and one for the TDSR case. Use python create_dataset.py -h to view optional arguments.

Pre-Trained Models

We provide the pre-trained models only for the SDSR case. However, for the TDSR case we used the same models to introduce the artificial corruptions.

	DSGAN
DF2K Gaussian	SDSR
DF2K JPEG	SDSR
DPED	SDSR
AIM 2019	SDSR

Setup for Reproducing the Results

Datasets

DF2K

• Download the DIV2K dataset (bicubic x4)
• Download the Flickr2K dataset (bicubic x4)
• Combine the HR and LR images of DIV2K and Flickr2K to create DF2K
• Adapt the paths in paths.yml if necessary
• Add Artificial corruptions as described below

DPED

• Download the DPED dataset
• Adapt the paths in paths.yml if necessary

Add Artificial Corruptions

Before you run the code, make sure that you added/edited the respective entries in the paths.yml file. The corrupted datasets can then be generated with:

python add_corruptions.py --dataset dataset 
                          --artifacts artifacts
                          --noise_std_dev None
                          --jpg_quality None
                          --resolution hr

To add corruptions, change the parameter for noise_std_dev or jpg_quality. We use a standard deviation of 8 for our experiments with sensor noise and a quality of 30 for our experiments with JPEG artifacts. Use python add_corruptions.py -h to view optional arguments.