prikaz prve stranice dokumenta Stvaranje slika pomoću dubokog učenja
No public access
master's thesis
Stvaranje slika pomoću dubokog učenja
Zagreb: University of Zagreb, Faculty of Electrical Engineering and Computing, 2018. urn:nbn:hr:168:470958
Šušnjara, Juraj
University of Zagreb
Faculty of Electrical Engineering and Computing

Institutional repository: FER Repository

Cite this document

Šušnjara, J. (2018). Stvaranje slika pomoću dubokog učenja (Master's thesis). Zagreb: University of Zagreb, Faculty of Electrical Engineering and Computing. Retrieved from https://urn.nsk.hr/urn:nbn:hr:168:470958

Šušnjara, Juraj. "Stvaranje slika pomoću dubokog učenja." Master's thesis, University of Zagreb, Faculty of Electrical Engineering and Computing, 2018. https://urn.nsk.hr/urn:nbn:hr:168:470958

Šušnjara, Juraj. "Stvaranje slika pomoću dubokog učenja." Master's thesis, University of Zagreb, Faculty of Electrical Engineering and Computing, 2018. https://urn.nsk.hr/urn:nbn:hr:168:470958

Šušnjara, J. (2018). 'Stvaranje slika pomoću dubokog učenja', Master's thesis, University of Zagreb, Faculty of Electrical Engineering and Computing, accessed 02 July 2024, https://urn.nsk.hr/urn:nbn:hr:168:470958

Šušnjara J. Stvaranje slika pomoću dubokog učenja [Master's thesis]. Zagreb: University of Zagreb, Faculty of Electrical Engineering and Computing; 2018 [cited 2024 July 02] Available at: https://urn.nsk.hr/urn:nbn:hr:168:470958

J. Šušnjara, "Stvaranje slika pomoću dubokog učenja", Master's thesis, University of Zagreb, Faculty of Electrical Engineering and Computing, Zagreb, 2018. Available at: https://urn.nsk.hr/urn:nbn:hr:168:470958

Metadata
TitleStvaranje slika pomoću dubokog učenja
Title (english)Generating Images Using Deep Learning
AuthorJuraj Šušnjara
MentorMarin Šilić (mentor)
Committee memberMarin Šilić (predsjednik povjerenstva)
Committee memberSiniša Srbljić (član povjerenstva)
Committee memberGoran Delač (član povjerenstva)
GranterUniversity of Zagreb
Faculty of Electrical Engineering and Computing
Zagreb
Defense date and country2018-07-05, Croatia
Scientific / art field,
discipline and subdiscipline		TECHNICAL SCIENCES
Computing
Abstract
Generativni modeli su modeli umjetnih neuronskih mreža koji su namijenjeni za stvaranje podataka iz naučene razdiobe. Pomoću takvih modela moguće je stvarati slike. Neuronske mreže za obradu slikovnih podataka koriste konvolucijske slojeve pa su i generativni modeli za stvaranje slika sastavljeni od njih. Obrađene su dvije vrste takvih modela: varijacijski autoenkoder (VAE) i generativne suparničke mreže (GAN). VAE je jednostavnija arhitektura koja je autoenkoder, prilagođen za stvaranje novih podataka. GAN je složena arhitektura od dvije mreže: generator i diskriminator. Mreže su vrednovane nad MNIST, LFW i CIFAR-10 skupovima podataka. VAE i GAN su uspješno istrenirane za stvaranje slika iz svih skupova. VAE je uspješniji nad jednostavnim MNIST skupom dok je GAN bolji nad složenijim LFW i CIFAR-10 skupovima. Nadalje, ostvareni su uvjetni VAE i GAN. Takvi modeli stvaraju slike sa zadanim uvjetom. Ostvareni su nad MNIST skupom podataka tako da je moguće zadati znamenku čiju sliku treba stvoriti. Uvjetni generativni modeli otvaraju mogućnosti za stvaranje složenijih slika s uvjetom, kao što su slike lica sa željenim svojstvima.
Abstract (english)
Generative models are neural network model made for generating data from learned distribution. It is possible to generate images with that kind of models. Neural networks that process images use convolutional networks so generative models for images contain them as well. Variational autoencoder (VAE) and generative adversarial networks (GAN) are two types of generative models explained in this paper. VAE has simpler architecture that is autoencoder adjusted for generating new data. GAN has complex architecture that consists of two networks: generator and discriminator. Networks are evaluated on MNIST, LFW and CIFAR-10 datasets. VAE and GAN are successfully trained for generating images on all of mentioned datasets. VAE was better on simple MNIST dataset while GAN performed better on LFW and CIFAR-10 datasets. Furthermore, conditional VAE and GAN are implemented as well. Those models generate data with given condition. They are trained on MNIST dataset so it is possible to set digit which image will be generated. Conditional generative models create opportunities for generating complex images with condition, such as face images with desired attributes.
Keywords
Keywords (english)
Languagecroatian
URN:NBNurn:nbn:hr:168:470958
Study programmeTitle: Computing
Study programme type: university
Study level: graduate
Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
Type of resourceText
File originBorn digital
Access conditionsClosed access
Terms of useLicence In copyright icon
Public note
Created on2019-03-14 01:29:47