Generiranje slik pohištva z uporabo skic in globokih nevronskih mrež

VELAGIĆ, ALDIN (author), Bosnić, Zoran (mentor), Papič, Aleš (comentor)

Cilj diplomske naloge je bilo učenje generativnih nasprotniških nevronskih mrež za namen generiranja slik pohištva iz skic. Poudarek je bil na učenju cikličnih generativnih nasprotniških mrež, kjer smo preizkusili različne tipe implementacij na enem ali več razredov pohištva. Modeli so bili učeni na slikah in skicah pohištva s spleta, ki so bile razdeljene v 3 razrede, in sicer: kavči, mize in omare. Izbrali smo slike brez ozadja in jih obdelali, da so bili objekti na skicah približno enake velikosti, ter jih nastavili na velikost 400 x 400. Del skic smo izdelali tudi ročno. Sledilo je učenje generativnih nasprotniških mrež in priprava modela, ki bi generiral čim bolj realistične slike iz skic. Točnost smo ocenili na pripravljeni testni množici skic in kot meri uspešnosti uporabili Fréchetovo začetno razdaljo in mero podobnosti zaznanih slikovnih zaplat, ki sta nam pomagali pri odločitvi, kateri model generira najbolj realne slike. Fréchetova začetna razdalja za najboljše modele iznaša: kavči 100,970, omare 86,730 in mize 113,460. Mera podobnosti zaznanih slikovnih zaplat za te modele iznaša: kavči 0,479, omare 0,489 in mize 0,380. Na koncu smo izdelali spletno aplikacijo, ki omogoča lažjo uporabo izbranih modelov in generiranje slik za razrede: kavči, mize in omare.

• Language: Slovenian
• Keywords: nevronske mreže, globoko učenje, generativne nasprotniške mreže, skice, slike
• 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-144731
• COBISS.SI-ID: 147375363
• Publication date in RUL: 09.03.2023

Secondary language

• Language: English
• Title: Generating images of furniture using sketches and deep neural networks

Abstract

The aim of the thesis was to learn generative adversarial neural networks for the purpose of generating furniture images from sketches. The focus was on learning Cyclic Generative Adversarial Networks, where we tested different types of implementations on one or more furniture classes. The models were learned on images and sketches of furniture from the web, which were divided into 3 classes: sofas, tables and wardrobes. We chose pictures without background and processed them so that the objects in the sketches were approximately the same size and were placed at 400 x 400. This was followed by learning generative adversarial networks and preparing a model that would generate as realistic images as possible from the sketches. We evaluated the accuracy on a prepared test set of sketches and used the Fréchet inception distance and the similarity measure of the perceptual image patches as performance measures to help us decide which model generates the most realistic images. Fréchet's inception distance for the best models is: sofas 100.970, wardrobes 86.730 and tables 113.460. The similarity measure of the perceptual image patches for these models is: sofas 0.479, cabinets 0.489 and tables 0.380. Finally, we developed a web application to facilitate the use of the selected models and to generate images for the classes: sofas, tables and cabinets.

• Keywords: neural networks, deep learning, generative adversarial networks, sketch, image