Treffer: Generiranje slika uvjetnim generativnim suparničkim mrežama ; Generating Images Using Conditional Generative Adversarial Networks

Title:
Generiranje slika uvjetnim generativnim suparničkim mrežama ; Generating Images Using Conditional Generative Adversarial Networks
Contributors:
Hrkać, Tomislav
Publisher Information:
Sveučilište u Zagrebu. Fakultet elektrotehnike i računarstva.
University of Zagreb. Faculty of Electrical Engineering and Computing.
Publication Year:
2025
Collection:
Croatian Digital Theses Repository (National and University Library in Zagreb)
Document Type:
Dissertation bachelor thesis
File Description:
application/pdf
Language:
Croatian
Rights:
http://rightsstatements.org/vocab/InC/1.0/ ; info:eu-repo/semantics/openAccess
Accession Number:
edsbas.D2510001
Database:
BASE

Weitere Informationen

Naslov: Generiranje slika uvjetnim generativnim suparničkim mrežama U ovom završnom radu implementiran je CGAN model u programskom jeziku Python, koristeći biblioteku PyTorch. Model je dizajniran za generiranje slika rukom pisanih znamenki uvjetovane klasom znamenke. Opisane su arhitekture generatora i diskriminatora, uključujući konvolucijske i transponirane konvolucijske slojeve te uvjetna batch normalizacija. Prikazan je proces treniranja modela, odabir funkcije gubitka, optimizatora i hiperparametara. Rezultati rada uključuju kvalitativnu analizu generiranih slika i praćenje tijeka funkcije gubitka tijekom treniranja. Diskutirana su ograničenja implementiranog sustava i moguća poboljšanja. ; Title: Image generation using Conditional Generative Adversarial Networks This bachelor's thesis implements a CGAN model using the Python programming language and the PyTorch library. The model is designed to generate images of handwritten digits conditioned by the class of the digit. The architectures of the generator and the discriminator are described, including convolutional and transposed convolutional layers as well as conditional batch normalization. Also shown are the training process and the selection of the used loss function, optimizer and hyperparameters. The results are evaluated based on qualitative analysis of the generated images and loss function behavior. The limitations of the implemented system and possible further improvements are discussed.