Face Denoising Project for Context Aware Security Analysis for Computer Vision

Info

Provided models have been trained on 10.000 images 64x64, the dataset is custom made and based on CelebA with a little modification. In our project we had to find the best performing model by only looking at papers and testing it with 2 set of faces: cropped (as close to no-background as possible) and large (as much background as possible)

• Top row is the input images, the original ones that the NN has never seen before
• Second row is the noisy images, the input of our NNs
• Third row is the output images for the WIN5_LARGE model
• Fourth row is the out images for the WIN5 model

We provide a whitepaper for better understanding of the process that made this models possible.

Models and Hardware requirements

The models have been trained on a nVidia Quadro P4000, each epochs took 93-95 seconds.

• WIN5 model was trained for 75 epochs, ispired by Peng Liu, Ruogu Fang
• WIN5_BW model was trained for 25~ epochs
• WIN5_LARGE model was trained for 25 epochs
• DNCNN_BW was trained for 27~ epochs, inspired by Kai Zhang et al.
• DNCNN was trained for 25~ epochs

Example pictures and the pre-trained models are aviable in the pre-trained models folder.

How to run/install

To run the model trainer:

git clone https://github.com/Owlz/Face-Denoising.git
cd Face-Denoising-CASACV
pip install -r requirements.txt
python model_trainer_edited.py

The dataset examples are in the file dataset folder, to generate them you can use the file script.py but you have to modify it based on what you need.

Collaborators

The project was build from the ground up by our team:

• Luca Di Luccio - Owlz
• Simone Faiella - SimoneFaiella
• Jan, Muhammad Tanveer - tjan90
• Marialuisa Trerè - mery00