Generative Adversarial Network (GAN) extensions and applications in image processing and generation

When human brains process input from the senses they are able to effortlessly imagine new instances and scenarios from only a small amount of input experience. Generative Adversarial Networks (GANs) manages to some extent to achieve this imagination ability for datasets. We explore this technique and its applicability in image processing and generation. This has applications in areas like medicine, physics, and artificial intelligence.

For example we implement the pix2pix algorithm for transforming satellite images into maps. This algorithm is application independent and can handle vastly different problems without significant tweaking. This shows that it is possible to create generalized methods to do complicated domain to domain transformations.

The main weight of this thesis will nonetheless fall on the extensive literature study of GAN variants. Here we cover different loss functions for GANs, like the Wasserstein metric, functional gradients for fine tuning GANs, ways of controlling the generated output, like conditional GAN, CycleGAN, and InfoGAN, and finally a Bayesian extension of GAN that provides uncertainty and inference to GANs.

Full text

• research
• thesis