Generating Images Using Conditional Generative Adversarial Networks

Cilj ovog diplomskog rada je primijeniti generativne suparničke mreža za generiranje slika lica uz dane uvjete i time ispitati modele koji su prikladni za uspješno generiranje slika. Kroz početna poglavlja opisane su stavke dubokog učenja koje definiraju modele generativnih suparničkih mreža kao što jesu. Nakon upoznavanja terminologije i osnovnih dijelova GAN modela, opisan je primjer modela cDCGAN s pojednostavljenom arhitekturom generatora i diskriminatora, koji se koristio za generiranje slika brojeva MNIST baze podataka i za upoznavanje s radom uvjetnih GAN modela. Posljednja poglavlja opisuju modele cDCGAN i AC-DRAGAN koji su korišteni za generiranje slika lica. Ti modeli su trenirani na CelebA bazi podataka. Oba modela sadrže opis arhitekture generatora i diskriminatora, strukturiranje ulaznih podataka i rješenja koja su generirali ti modeli.The goal of this master thesis is to apply a generative adversarial networks to generate images of human faces with given conditions and thereby examine models that are suitable for successful image generation. Through the initial chapters, deep learning elements are described which define the generative adversarial networks as they are. After addressing the terminology and basic parts of the GAN model, an example of the cDCGAN model with the simplified architecture of the generator and discriminator is described, which was used for generating images of numbers of the MNIST dataset and to learn how the conditional GAN model works. The last chapters describe cDCGAN and AC-DRAGAN models which are used for generating images of human faces. These models were trained on CelebA dataset. Both models contain a description of the architecture of the generator and discriminator, the structure of input data, and images that were generated by the models