This code is developed as a master degree course project (Machine Learning for Vision and Multimedia) at Politecnico di Torino.
Contributors:
- Aglieco Francesco.
- Comparetto Alessandra.
- Gagliardi Giuseppe.
Gan model based on:
- Generator: simplified U-Net structure.
- Discriminator: image classification CNN.
Noise used are gaussian, salt and pepper and spackle.
Losses used on training step:
- Discriminator loss: binary cross-entropy.
- Generator loss: content loss, using a VGG19 pre-trained model as feature extractor.
Images are cropped in 256x256x3 dimension, cropping window is randomly seated.
COCO dataset is used as train/test set.
BSD300 dataset is used as second test set.
Main features:
- Efficient loading of dataset on Drive folders after previous download.
- Saving and loading on Drive folders of model trained epoch by epoch (in order to deal with low resource allocated by Google Colab).
- Content Loss enforces performance results.
Details are written on pdf file located in this repo.
