Preslikava stila satelitskih posnetkov s pomočjo generativnih nasprotniških nevronskih mrež

magistrsko delo

Mitja Lakič (Author), Sašo Karakatič (Mentor)

Abstract

V magistrskem delu raziskujemo problematiko preslikave stila satelitskih posnetkov z uporabo generativnih nasprotniških nevronskih mrež (GAN). Najprej predstavimo osnovno strukturo nevronskih mrež, nato podrobneje opišemo generativne modele. Namen magistrskega dela je preveriti učinkovitost teh modelov pri preslikavi satelitskih posnetkov v stil zemljevida, kjer primerjamo dva različna GAN modela, in sicer Pix2Pix, ki spada med pogojne modele, in CycleGAN, ki je predstavnik cikličnih modelov. V okviru eksperimenta primerjamo pridobljene rezultate z uporabo teh modelov, kjer smo tudi preizkusili preslikavo v obratni smeri, torej iz zemljevida v stil satelitskega posnetka. Rezultati so pokazali, da je mogoče satelitske posnetke uspešno preslikati v stil zemljevida, kjer pogojni modeli na splošno zagotavljajo boljše rezultate, vendar so zelo odvisni od arhitekture omrežja. Magistrsko delo zaključimo z analizo rezultatov in odgovori na raziskovalna vprašanja.

Keywords

generativne nasprotniške mreže;globoko učenje;preslikava stila;satelitski posnetki;zemljevidi;magistrske naloge;

Data

Language:	Slovenian
Year of publishing:	2023
Typology:	2.09 - Master's Thesis
Organization:	UM FERI - Faculty of Electrical Engineering and Computer Science
Publisher:	[M. Lakič]
UDC:	004.032.26+004.85(043.2)
COBISS:	151855619
Views:	102
Downloads:	21
Average score:	0 (0 votes)
Metadata:

Other data

Secondary language:	English
Secondary title:	Transferring the style of satellite images using generative adversarial neural networks
Secondary abstract:	In the master's thesis, we investigate the problem of transferring the style of satellite images using generative adversarial neural networks (GANs). First, we present the basic structure of neural networks, and then we describe generative models in more detail. The purpose of the master's thesis is to verify the effectiveness of these models in transferring satellite images into a map style, where we compare two different GAN models, namely Pix2Pix, which belongs to conditional models, and CycleGAN, which is a representative of cyclic models. As part of the experiment, we compare the results obtained using these models, where we also tested the style transfer in the reverse direction, that is, from a map to the style of a satellite image. The results showed that satellite imagery can be successfully transferred into a map style, where conditional models generally provide better results but are highly dependent on the network architecture. The master's thesis concludes with an analysis of the results and answers to the research questions.
Secondary keywords:	generative adversarial networks;deep learning;style transfer;satellite imagery;maps;
Type (COBISS):	Master's thesis/paper
Thesis comment:	Univ. v Mariboru, Fak. za elektrotehniko, računalništvo in informatiko, Informatika in tehnologije komuniciranja
Pages:	1 spletni vir (1 datoteka PDF (X, 92 f.))
ID:	18057106