Deep-Learning-Super-Resolution was used as a basis for this project. It includes different loss functions and models and was made as a project for university by me and my friend Mihailo Pacaric.
This code is an implementation of the SRGAN Paper.
I trained a recreated version of their SRResNet and SRGAN. Also included is a trained modified version of their generation network which was made for 2 x upscaling.
The tensorflow model source code can be found at dlsr/models.
The perceptual loss that was used can be found at dlsr/losses/srgan.py
The dataset I used was DIV2K and the code for loading the dataset can be found at dlsr/data/div2k.py, it is a modified version of code found here.
| Low Resolution | SRResNet 4x | SRGAN 4x |
|---|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
 |
| Low Resolution | SRResNet 4x | SRGAN 4x |
|---|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
 |
Note: Python 3.10.5 was used