Poboljšanje procjene duboke neuronske mreže kroz variranje slikovnog ulaza

master's thesis

• Title (english): Improving of Deep Neural Network Estimation Through Variation of Input Images
• Author: Roko Franetović
• Mentor: Marko Subašić (mentor)
• Committee member: Marko Subašić (predsjednik povjerenstva)
• Committee member: Juraj Petrović (član povjerenstva)
• Committee member: Krešimir Križanović (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2022-07-12, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Generiranje umjetnih podataka i njihovo korištenje unutar područja
strojnog učenja česta je tehnika koja se koristi za testiranje i
poboljšanje performansi modela. U mnogim situacijama generiranje
umjetnih podataka je znatno lakši postupak naspram generacije stvarnih
podataka, pogotovo u području računalnog vida gdje su podaci vrlo
često slike. U ovom radu smo proučili raspoloživa programska rješenja
za generiranje fotorealističnih slika 3D scena. Također smo proučili
duboke neuronske mreže za procjenu parametara na temelju slika s
posebnim naglaskom na konvolucijske mreže. Ispitali smo mogućnosti
povezivanja odabrane arhitekture neuronske mreže i programskog
rješenja za generiranje slika iz 3D scena. Istražili smo mogućnost
povezivanja odabranog algoritma treniranja duboke mreže i 3D
scene. Implementirali smo interaktivno mijenjanje teksture 3D modela s
ciljem poboljšanja rezultata procjene duboke neuronske mreže.

Abstract (english)

Generation of artificial data and its use within the area of machine
learning is a common technique used for testing and improving
performance of our models. In many situations, generating artificial
data is a much easier task compared to the generation of real data,
especially in the area of computer vision where our data is often
images. In this work, we have studied the available software solutions
for generation of photorealistic images of 3D scenes. We also studied
deep neural networks for parameter estimation based on images with
special emphasis on convolutional neural networks. We examined methods
for connecting our neural network architecture and the software
solutions for generating images from 3D scenes. We also explored
methods of connecting the chosen deep network training algorithm and
the 3D scene. We implemented an interactive texture modification of
the 3D model with the aim of improving the results of the deep neural
network.

• Keywords: računalni vid; duboke neuronske mreže; konvolucijske neuronske mreže; računalna grafika; Blender; generiranje umjetnih podataka
• Keywords (english): computer vision; deep neural networks; convolutional neural networks; computer graphics; Blender; artificial data generation
• Language: croatian
• URN:NBN: urn:nbn:hr:168:587141
• 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: Closed access
• Created on: 2023-01-13 12:14:38