Note

November 2024: New Pre-trained Models are available, check the Pre-trained Models section.

Foggy-CycleGAN

This project is the implementation for my Computer Science MSc thesis in the University of Debrecen.

Dissertation: [PDF] Simulating Weather Conditions on Digital Images (Debrecen, 2020).

Table of Content

• Description
• Code
• Notebook
• Results
• Pre-trained Models

Description

Foggy-CycleGAN is a CycleGAN model trained to synthesize fog on clear images. More details in the dissertation above.

Code

The full source code is available under GPL-3.0 License in my Github repository ghaiszaher/Foggy-CycleGAN

Notebook

A Jupyter Notebook file Foggy_CycleGAN.ipynb is available in the repository.

Results

(as of June 2020)

© Ghais Zaher 2020

Pre-trained Models

As previous pre-trained models are no longer compatible with newer Keras/Tensorflow versions, I have retrained the model and made the new weights available to download.

Each of the following models was trained in Google Colab using the same dataset, the parameters for building the models and number of trained epochs are a bit different:

Model	Trained Epochs	Config
2020-06 (legacy)	145	use_transmission_map=False use_gauss_filter=False use_resize_conv=False
2024-11-17-rev1-000	522	use_transmission_map=False use_gauss_filter=False use_resize_conv=False
2024-11-17-rev2-110	100	use_transmission_map=True use_gauss_filter=True use_resize_conv=False
2024-11-17-rev3-111	103	use_transmission_map=True use_gauss_filter=True use_resize_conv=True
2024-11-17-rev4-001	39	use_transmission_map=False use_gauss_filter=False use_resize_conv=True

Results

The results of the new models are similar to the previous ones, here are some samples:

Clear	2024-11-17-rev1-000	2024-11-17-rev2-110	2024-11-17-rev3-111	2024-11-17-rev4-001