prikaz prve stranice dokumenta Generativne suparničke mreže i primjene
  Download
PDF 8.35 MB
master's thesis
Generativne suparničke mreže i primjene
Zagreb: University of Zagreb, Faculty of Science, 2022. urn:nbn:hr:217:156267
Peroš, Mate
University of Zagreb
Faculty of Science
Department of Mathematics

Institutional repository: Repository of the Faculty of Science

Cite this document

Peroš, M. (2022). Generativne suparničke mreže i primjene (Master's thesis). Zagreb: University of Zagreb, Faculty of Science. Retrieved from https://urn.nsk.hr/urn:nbn:hr:217:156267

Peroš, Mate. "Generativne suparničke mreže i primjene." Master's thesis, University of Zagreb, Faculty of Science, 2022. https://urn.nsk.hr/urn:nbn:hr:217:156267

Peroš, Mate. "Generativne suparničke mreže i primjene." Master's thesis, University of Zagreb, Faculty of Science, 2022. https://urn.nsk.hr/urn:nbn:hr:217:156267

Peroš, M. (2022). 'Generativne suparničke mreže i primjene', Master's thesis, University of Zagreb, Faculty of Science, accessed 03 November 2024, https://urn.nsk.hr/urn:nbn:hr:217:156267

Peroš M. Generativne suparničke mreže i primjene [Master's thesis]. Zagreb: University of Zagreb, Faculty of Science; 2022 [cited 2024 November 03] Available at: https://urn.nsk.hr/urn:nbn:hr:217:156267

M. Peroš, "Generativne suparničke mreže i primjene", Master's thesis, University of Zagreb, Faculty of Science, Zagreb, 2022. Available at: https://urn.nsk.hr/urn:nbn:hr:217:156267

Metadata
TitleGenerativne suparničke mreže i primjene
Title (english)Generative rival networks and applications
AuthorMate Peroš
MentorZvonimir Bujanović (mentor)
Committee memberZvonimir Bujanović (predsjednik povjerenstva)
Committee memberGoran Muić (član povjerenstva)
Committee memberVedran Čačić (član povjerenstva)
Committee memberJosip Tambača (član povjerenstva)
GranterUniversity of Zagreb
Faculty of Science
(Department of Mathematics)
Zagreb
Defense date and country2022-03-03, Croatia
Scientific / art field,
discipline and subdiscipline		NATURAL SCIENCES
Mathematics
Abstract
Povećanje razlučivosti česta je tema istraživanja u području računalnog vida. Problem povećanja razlučivosti je proces dobivanja slike visoke razlučivosti iz odgovarajuće slike niske razlučivosti i pronalazi veliku primjenu u satelitskom snimanju, obradi medicinskih prikaza, analizi fotografija teksta te biometrijskom prepoznavanju. U ovom radu dali smo pregled teorije generativnih suparničkih mreža, modela koji se sastoje od dvije suprotstavljene mreže, generatorske i diskriminatorske, gdje generatorska ima zadatak naučiti distribuciju ulaznih podataka s ciljem generiranja novih, uvjerljivih podataka, a diskriminatorska nastoji postići što veći uspjeh u raspoznavanju stvarnih podataka od generiranih. Također, promotrili smo dvije funkcije pogreške, prvu temeljenu na binarnoj unakrsnoj entropiji, a drugu na Wasserstein udaljenosti i kažnjavanju odstupanja gradijenta. Konačno, predložili smo dva modela za rješavanje problema povećanja razlučivosti temeljena na generativnih suparničkim mrežama, prvog s funkcijom pogreške temeljenom na binarnoj unakrsnoj entropiji, a drugog s funkcijom pogreške temeljenoj na Wasserstein udaljenosti i kažnjavanju odstupanja gradijenta. Takvi modeli pokazali su se kao kvalitetna zamjena za duboke konvolucijske modele i postigli su fine detalje na slikama povećane razlučivosti bez potrebe za velikim brojem konvolucijskih slojeva.
Abstract (english)
Single image super-resolution is a common research topic in the field of computer vision. The problem of single image super-resolution is the process of obtaining a high-resolution image from a corresponding low-resolution image and finds great application in satellite imaging, medical imaging, text-to-photo analysis and biometric recognition. In this thesis, we provide an overview of the theory of generative adversarial networks which are models of two adversarial networks, generator and discriminator, where the generator has the task of learning the distribution of input data to generate new, compelling data, and the discriminator seeks to achieve success in recognizing real data from generated. We also considered two error functions, the first based on binary cross entropy and the second on Wasserstein distance and gradient penalty. Finally, we proposed two models to solve the problem of single image super-resolution based on generative adversarial networks, the first with a loss function based on binary cross entropy, and the second with a loss function based on Wasserstein distance and gradient penalty. Such models have proven to be a quality replacement for deep convolutional models and have achieved fine detail in high-resolution images without the need for a large number of convolutional layers.
Keywords
Keywords (english)
Languagecroatian
URN:NBNurn:nbn:hr:217:156267
Study programmeTitle: Computer Science and Mathematics
Study programme type: university
Study level: graduate
Academic / professional title: magistar/magistra računarstva i matematike (magistar/magistra računarstva i matematike)
Type of resourceText
File originBorn digital
Access conditionsOpen access
Terms of useLicence In copyright icon
Created on2022-03-30 10:09:29