Super-rezolucija slika strojnim učenjem

master's thesis

• Title (english): Image super-resolution using machine learning
• Author: Marko Lazarić
• Mentor: Marko Čupić (mentor)
• Committee member: Marko Čupić (predsjednik povjerenstva)
• Committee member: Alan Jović (član povjerenstva)
• Committee member: Željka Mihajlović (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2022-06-28, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Super-rezolucija na temelju jedne slike je zadatak gdje je cilj dobiti sliku visoke rezolucije na temelju jedne slike niske rezolucije.
To je iznimno loše postavljen problem s fokusom na ljudsku percepciju jer svaki piksel može poprimiti veliki broj vrijednosti i zato što ljudi nisu jednako osjetljivi na svaku promjenu u slici.
Pristupi temeljeni na strojnom učenju su ostvarili dobre rezultate na tom problemu, ali zahtijevaju skupu računalnu opremu za pokretanje i treniranje.

U ovom radu, predložen je pristup temeljen na konvolucijskoj neuronskoj mreži koji ostvaruje dobru točnost i brzinu izvođenja.
Model je baziran na FSRCNN modelu s dodatnim modifikacijama kao što su samoansambliranje i kombinacija s interpolacijom za komponente prostora boja na koje ljudi nisu jako osjetljivi.
Trenirani model je evaluiran na popularnim skupovima podataka za super-rezoluciju i uspoređen s pristupima temeljenim na interpolaciji i drugim popularnim modelima strojnog učenja.
Jednostavna web aplikacija je implementirana koja omogućava korisniku da jednostavno usporedi različite pristupe super-rezoluciji na proizvoljnim slikama vizualno i numerički.

Abstract (english)

Single-image super-resolution is the task of transforming a single low-resolution image into a high-resolution one.
It is a fundamentally ill-posed and human-oriented problem due to the number of solutions for each pixel and the fact that people are not equally sensitive to all changes in the image.
Machine learning based approaches to the problem have achieved great results in the problem, but require expensive hardware to run and train.

In this thesis, a convolutional neural network based approach to the problem is explored to find a model with good accuracy and speed.
It is based on the FSRCNN model with additional modifications of self-ensembling and combination with interpolation for the color space components which people are not as sensitive to.
The trained model is evaluated on popular super-resolution datasets and compared with interpolation based approaches and other popular machine learning models.
Finally, a simple web application is implemented to visually and numerically evaluated the different super-resolution approaches on arbitrary images.

• Keywords: super-rezolucija; strojno učenje; duboko učenje; konvolucijske neuronske mreže
• Keywords (english): super-resolution; machine learning; deep learning; convolutional neural network
• Language: english
• URN:NBN: urn:nbn:hr:168:255873
• 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 15:23:48