diplomsko delo
Urban Tanko (Author), Danijel Skočaj (Mentor)

Abstract

Pri generiranju slik se vse več uporabljajo metode GAN. Ena od slabosti je dolgotrajnost njihovega učenja. V diplomski nalogi jo poskusimo odpraviti z uporabo modelov za translacijo med slikami, s katerimi želimo izboljšati kvaliteto generiranih slik. To storimo tako, da zberemo podatkovno množico in na njej naučimo model za generiranje slik StyleGAN. Generirane slike nato poženemo skozi naslednje modele za translacijo med slikami: SR-GAN, Pix2pix, CycleGAN, Pix2pixHD, U-GAT-IT in DeblurGAN. Za vsakega od modelov opišemo generirane slike in jih ocenimo z metriko FID ter človeško oceno, pridobljeno z uporabo ankete. Pridobljene rezultate tudi primerjamo med seboj.

Keywords

strojno učenje;umetna inteligenca;nevronske mreže;generativne nasprotniške mreže;translacija med slikami;podatkovna množica;ekstrakcija podatkov;računalništvo in informatika;univerzitetni študij;diplomske naloge;

Data

Language: Slovenian
Year of publishing:
Typology: 2.11 - Undergraduate Thesis
Organization: UL FRI - Faculty of Computer and Information Science
Publisher: [U. Tanko]
UDC: 004.85(043.2)
COBISS: 1538565827 Link will open in a new window
Views: 937
Downloads: 225
Average score: 0 (0 votes)
Metadata: JSON JSON-RDF JSON-LD TURTLE N-TRIPLES XML RDFA MICRODATA DC-XML DC-RDF RDF

Other data

Secondary language: English
Secondary title: Improving the quality of generated images using image-to-image translation models
Secondary abstract: The application of GAN methods for the purpose of image synthesis has grown considerably. One of their weaknesses is long training time. In this thesis we try to eliminate it by using image-to-image translation models to improve generated image quality. We first gather our dataset and train an image synthesis model StyleGAN. We then feed the generated images into various image-to-image translation models: SR-GAN, Pix2pix, CycleGAN, Pix2pixHD, U-GAT-IT in DeblurGAN. For each of the models we describe the visual properties of generated images. We also calculate the FID scores and human scores, obtained with a survey. At the end we compare the results of the models.
Secondary keywords: machine learning;artificial intelligence;neural networks;generative adversarial networks;image-to-image translation;dataset;data scraping;computer and information science;diploma;
Type (COBISS): Bachelor thesis/paper
Study programme: 1000468
Embargo end date (OpenAIRE): 1970-01-01
Thesis comment: Univ. v Ljubljani, Fak. za računalništvo in informatiko
Pages: 83 str.
ID: 11502266