Efficient Image Super-Resolution

This repository contains the source for our Machine Intelligence with Deep Learning (MIDL) seminar topic Efficient Image Super-Resolution from the winter term 2021/2022.

The project goal is to extend the already existing the Residual Feature Distillation Network to increase speed and accuracy while simultaneously decreasing model size. The code of the RFDN can be found here, the underlying framework of the AIM 2020 Challenge here.

Set Up

General

1. Clone the repository.

git clone [email protected]:MartinBuessemeyer/Efficient-Image-Super-Resolution.git

2. Get the data sets

   • DIV2K
   • Flickr2k
   • Set5
   • Set14
   • BSD100
   • Urban100

3. Build the enroot container. This will automatically handle all dependencies for you.

sh ./scripts/build-image-enroot.sh

Alternatively, you can execute the code locally. Make sure that you installed pytorch and the packages listed in src/requirements.txt.

How to execute

1. Run the container. The following steps should be executed inside the enroot container.

2. Adjust the src/run.sh. You can find all the available options in the src/options.py. Example configurations are listed in the src/run.sh.

3. Run the src/run.sh.

sh ./src/run.sh

4. The preferred way to view results is via WandB. Additionally, results are stored in the experiment folder.

Authors

Martin Büßemeyer, Björn Daase, and Maximilian Kleissl

License

# SPDX-License-Identifier: MIT