Modeli dubokog učenja za razlikovanje između umjetničkih slika i fotografija

master's thesis

• Title (english): Deep learning models for distinguishing between paintings and photographs
• Author: Ivana Stilinović
• Mentor: Alan Jović (mentor)
• Committee member: Alan Jović (predsjednik povjerenstva)
• Committee member: Željka Mihajlović (član povjerenstva)
• Committee member: Nikolina Frid (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2025-02-27, Croatia

Abstract

Ovaj rad istražuje primjenu modela dubokog učenja za razlikovanje umjetničkih slika
i fotografija. Korišteni su unaprijed naučeni modeli konvolucijskih neuronskih mreža,
uključujući VGG16, ResNet50 i EfficientNetB0, s ciljem ispitivanja njihove učinkovitosti
u binarnoj klasifikaciji. Također je analiziran mehanizam pažnje CBAM kako bi se procijenio njegov utjecaj na performance modela. Eksperimentalni rezultati pokazali su da
je VGG16 postigao najbolje rezultate (točnost 96,5%, F1-mjera... More 96,59%), dok su složeniji
modeli, poput ResNet50 i EfficientNetB0, imali poteškoće s generalizacijom. Mehanizam pažnje nije poboljšao klasifikaciju, već je povećao računalne zahtjeve. Zaključeno
je da bi dodatna optimizacija hiperparametara, povećanje skupa podataka te istraživanje
alternativnih arhitektura, poput Vision Transformera, moglo doprinijeti boljim rezultatima u ovom području. Less

Abstract (english)

This thesis investigates the application of deep learning models to distinguish between art paintings and photographs. Pre-trained CNN models, including VGG16, ResNet50
and EfficientNetB0, were used to test their performance of binary classification. The
CBAM attention mechanism was also analyzed to assess its impact on model performance. Experimental results showed that VGG16 achieved the best results (accuracy
96.5%, F1-score 96.59%), while more complex models, such as ResNet50 and... More EfficientNetB0, had difficulties with generalization. The attention mechanism did not significanlty improve the classification, but increased the computational requirements. It was
concluded that additional optimization of hyperparameters, increase of dataset, and research of alternative architectures, such as Vision Transformers, can contribute to better
results in this field.
Less

• Keywords: duboko učenje; konvolucijske neuronske mreže; klasifikacija slika; umjetničke slike; fotografije
• Keywords (english): deep learning; convolutional neural networks; image classification; paintings; photographs
• Language: croatian
• URN:NBN: urn:nbn:hr:168:716902
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Open access
• Created on: 2025-03-12 11:50:22