Algoritmi za super-resolucijo v 3D grafiki

Pristavnik Vrešnjak, Matic (author), Klemenc, Bojan (mentor)

Zaradi naraščanja resolucije zaslonov je postala uporaba algoritmov za super-resolucijo bistveno bolj pogosta. Z uporabo algoritmov dosežemo hitrejšeizrisovanje slik. Prav tako algoritmi za super-resolucijo omogočijo razširjanje slike iz manjše resolucije v večjo. Algoritmi lahko delujejo samo z eno ali več slikami. Danes se večinoma uporabljajo nevronske mreže.Ker so se za uporabo v 3D grafiki pojavili novi algoritmi za super-resolucijo in ker o njih ni objavljeno veliko performančnih podatkov, sem opravil analizo algoritmov in jih primerjal. Slike za analizo sem pridobil z Unity. Za analizo sem uporabil naslednje metode: PSNR, SSIM, MSE in BRISQUE. Bolj podrobno analizo sem uporabil razlike slik, ki jih algoritem izračuna glede na referenčno sliko v visoki resoluciji. Za vsak posamezni algoritem sem opisal njegovo delovanje in prednosti ter slabosti. Za zaključek sem primerjal rezultate med posameznimi algoritmi in jih poskušal kategorizirati.

• Language: Slovenian
• Keywords: super-resolucija, 3D grafika, upodabljanje, nevronske mreže
• 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-149490
• COBISS.SI-ID: 163736835
• Publication date in RUL: 07.09.2023

Secondary language

• Language: English
• Title: Algorithms for super-resolution in 3D graphics

Abstract

Due to the increasing resolution of displays, the use of algorithms for super-resolution has become significantly more common. By employing these algorithms, faster image rendering can be achieved. Additionally, super-resolution algorithms enable the upscaling of images from lower to higher resolutions. These algorithms can operate on a single image or multiple images. Today, neural networks are predominantly used for this purpose. As new algorithms for super-resolution have emerged in 3D graphics, and there is limited performance data about them, I conducted an analysis and compared them. I obtained images for analysis using Unity. For the analysis, I employed the following methods: PSNR, SSIM, MSE and BRISQUE. To examine details, I calculated difference images using the baseline reference full resolution image. For each individual algorithm, I described its functioning, strengths, and weaknesses. In conclusion, I compared the results among the various algorithms and attempted to categorize them.

• Keywords: super-resolution, 3D graphics, rendering, neural networks