Style transfer of Aartworks using neural networks

KURBEGOVIĆ, ENIO (author), Bosnić, Zoran (mentor), Papič, Aleš (comentor)

In this thesis, we implement a neural style transfer model, which uses so-called “meta networks” to train image transformation networks. A trained image transformation network takes in two images - a content and a style image, and generates a new image, combining the content from the first with the style from the second image. We take an already existing model and train it on our own style dataset, as well as reduce the size of the content dataset, in order to see how to perform style transfer on a smaller amount of training data. Finally, we create a website, which allows users to generate their own stylized images using our trained models. At the end of the project we can say that meta networks have proven to be very efficient for operations with smaller datasets and they produce satisfactory results.

• Language: English
• Keywords: artificial intelligence, neural networks, neural style transfer, meta networks, image transformation networks, content images, style images
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2023
• PID: 20.500.12556/RUL-153006
• COBISS.SI-ID: 178678787
• Publication date in RUL: 14.12.2023

Secondary language

• Language: Slovenian
• Title: Prenos sloga umetniških del z uporabo nevronskih omrežij

Abstract

V tem članku smo implementirali model za nevronski prenos sloga, ki uporablja tako imenovana “meta omrežja” za ustvarjanje omrežij za preoblikovanje slik. Usposobljeno omrežje za preoblikovanje slik sprejme dve sliki - vsebinsko in stilsko sliko - ter ustvari novo sliko, pri čemer združi vsebino s prve slike s slogom iz druge slike. Že obstoječi model učimo na lastni bazi podatkov stilskih slik in zmanjšamo velikost baze podatkov vsebinskih slik, da bi videli, kako narediti prenos sloga na manjši količini podatkov za učenje. Ustvarili smo spletno stran, ki uporabnikom omogoča, da ustvarijo lastne stilizirane slike z uporabo naših usposobljenih modelov. Ob zaključku lahko rečemo, da so se meta omrežja izkazala za učinkovita z manjšimi nabori podatkov in dajejo zadovoljive rezultate.

• Keywords: umetna inteligenca, nevronske mreže, nevronski prenos sloga, meta mreže, mreže za preoblikovanje slik, vsebinske slike, slogovne slike