Using generative adverserial neural networks for creation of dog images

bachelor's thesis

Teja Boršič (Author), Sašo Karakatič (Mentor)

Abstract

Kot glavni cilj diplomske naloge smo poleg pregleda obstoječe literature preučili že razvit model generativne nasprotniške mreže StyleGAN ter podrobneje analizirali njeno implementacijo in delovanje. S pridobljenim znanjem iz literature in pregledom obstoječih rešitev smo izbrali projekt, ki smo ga prilagodili in optimizirali za delo na izbranih digitalnih slikah psov. Poudarili smo pomen uporabe ustreznih slik psov za učenje in vpliv različnih parametrov, kot so velikost serije, plasti in velikost slike, na optimizacijo zmogljivosti modela za ustvarjanje realističnih slik.

Keywords

neural networks;dogs;image generation;generative adversarial neural networks;StyleGAN;

Data

Language:	English
Year of publishing:	2021
Typology:	2.11 - Undergraduate Thesis
Organization:	UM FERI - Faculty of Electrical Engineering and Computer Science
Publisher:	[T. Boršić]
UDC:	004.8:004.93(043.2)
COBISS:	86218499
Views:	643
Downloads:	72
Average score:	0 (0 votes)
Metadata:

Other data

Secondary language:	Slovenian
Secondary title:	Using generative adverserial neural networks for creation of dog images
Secondary abstract:	As our main objective, in addition to reviewing the existing literature, we studied an already developed model of the generative adversarial network StyleGAN, analyzed its implementation and operation in more detail. With the acquired knowledge from the literature and a review of existing solutions, we adapted and optimized a chosen project to work on our selected digital images of dogs. We captured the importance of using appropriate images of dogs for learning and the impact of different parameters, such as batch size, layers, and image size, on the optimization of the performance of the model for image generation. We successfully generated multiple realistic samples of dogs.
Secondary keywords:	nevronske mreže;psi;generacija slik;generativne nasprotniške mreže;diplomske naloge;
Type (COBISS):	Bachelor thesis/paper
Thesis comment:	Univ. v Mariboru, Fak. za elektrotehniko, računalništvo in informatiko, Informatika in tehnologije komuniciranja
Pages:	IX, 34 str.
ID:	13367772