Deep Learning with Generative Adversarial Networks

OData support
Supervisor:
Kahloot Khalid M.M.
Department of Automation and Applied Informatics

Generative Adversarial Networks (GAN) is the state of the art generative model, especially for realistic image generation. Most research in GANs focuses on improving the quality of the generated images. GANs are able to learn the latent distributions in the images while given large data examples. Apparently, there is not much to do when the dataset is small that consequently results in low-resolution images.

In this thesis, we propose a novel technique to address small dataset impediment for the GANs. The proposed technique is called smart augmentation, which carries out pixel manipulation on the original dataset, in addition, filters out the badly distorted images. A deep convolution neural network classifier is trained to judge the goodness of the augmentation. The smart augmentation technique is able to extend the size of the datasets at least two to four times, which, in turn, has a huge impact on GANs image generation. We have applied the smart augmentation for MNIST, Fashion-MNIST, Bob Ross paintings, WikiArt landscape paintings, and Dogs vs Cats datasets with using DCGAN and ProGAN. The results of smart augmentation with ProGAN have shown significant improvement in realistic high-quality image generation.

Downloads

Please sign in to download the files of this thesis.