Super-Resolution with GANs

Based on the paper Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network.
Related work: https://github.com/leftthomas/SRGAN.

Datasets

Setup

Install pytorch as described on their website
pip install -r requirements.txt
Execute the setup script to download the VOC2012 and DIV2k data and partition them into train test and validation sets

Usage

Load images to data/val and data/train folder
Run python srgan/train.py --upscale_factor 4 All options can be found with train --help
Find outcome pictures in results/val folder
Find statics (PSNR, SSIM,...) in logs/statistics
The weights will be saved in log/epochs/
For testing load files to the data/test folder and run '''test.py''' with the trained weights as parameter
After several runs run the cleanupscript to remove half of all validation pictures and weights
Run python srgan/use_webcam.py to use the network for generating sr pictures from webcam pictures. Press space for generation and closing the transformed image window. Escape for closing the program