Repository for AI-based up-scaling evaluation.
We evaluated the following AI-based up-scaling methods:
- BSRGAN: https://github.com/cszn/BSRGAN
- KXNet: https://github.com/jiahong-fu/KXNet
- Real-ESRGAN: https://github.com/xinntao/Real-ESRGAN
- waifu2x: https://github.com/nihui/waifu2x-ncnn-vulkan
To download the images and pre-trained models, you need to run ./download.sh (check this script for the baseurl of the hosted data).
This requires xz-utils to be installed under Ubuntu.
upscaling: images used for the subjective test (including the 1080p reference images and up-scaled variants), the test was conducted using AVrateVoyager.evaluation: evaluation scripts (jupyter notebook required) and subjective annotations (/subjective/*ormos.csv)upscaling_featurescalculated signal and other features
The DNN experiments are modified variants of the following two repositories:
additional code for prediction (and not only training and evaluation) can be found in the corresponding repositories and must be adjusted.
Each of the DNN experiments needs specifically prepared data, where we provide scripts for the creation in the corresponding data folder.
The provided software is tested under Ubuntu 22.04 and 23.10.
- python3, python3-pip, jupyter notebook/lab
If you use this software or data in your research, please include a link to the repository and reference the following paper.
@inproceedings{goering2024aiupscaling,
author = {Steve G\"oring and Rasmus Merten and Alexander Raake},
title = {Appeal prediction for AI up-scaled Images},
booktitle={26th IEEE International Symposium on Multimedia (2024 IEEE ISM)},
year = {2024},
pages={1-8},
volume={},
month={Dec},
code={https://github.com/Telecommunication-Telemedia-Assessment/ai_upscaling},
note={to appear},
}GNU General Public License v3. See LICENSE.md file in this repository.