Povećavanje rezolucije fotografija korištenjem duboke konvolucijske mreže

master's thesis

• Title (english): Image super-resolution using deep convolutional networks
• Author: Željko Halle
• Mentor: Domagoj Vlah (mentor)
• Committee member: Domagoj Vlah (predsjednik povjerenstva)
• Committee member: Mario Brčić (član povjerenstva)
• Committee member: Zvonko Kostanjčar (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2021-07-15, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Strojno učenje primjenjivo je u mnogim domenama. Najaktualnija vrsta strojnog učenja jest duboko učenje. Ono se može koristiti za regresijske probleme kao što je predviđanje cijene dionice, za klasifikaciju slike, za obradu prirodnog jezika itd. U ovom radu pokazano je kako se duboko učenje može primijeniti na problem super-rezolucije slike. Prezentirana su dva duboka modela koja bi se mogla iskoristiti za riješavanje tog problema. Prvi model je VDSR (engl. Very Deep Super Resolution) koji je primjer klasične duboke mreže. Nad njime je izvršen postupak pronalaska optimalnih hiperparametara metodom iscrpnog pretraživanja prostora hiperparametara. Nađeni su hiperparametri u ograničenom prostoru hiperparametara koji daju najbolje rezultate na testnom skupu podataka. Drugi razmatrani model je duboka mreža s rezidualnom arhitekturom koja se uči nešto duže od klasičnih dubokih mreža. Prezentirani su rezultati koje takva mreža ostvaruje na testnom skupu podataka i ti su rezultati prikazani na realnim slikama.

Abstract (english)

Machine learning is applicable in many domains. The most recent type of machine learning is deep learning. It can be used for regression problems such as stock price prediction, image classification, natural language processing, etc. This paper shows how deep learning can be applied to the image super-resolution problem. Two deep models that could be used to solve this problem are presented. The first model is VDSR (Very Deep Super Resolution) which is an example of a classic deep network. The procedure of finding optimal hyperparameters was performed on it by the method of exhaustive search of the space of hyperparameters known as grid search. Hyperparameters were found in the limited space of hyperparameters that give the best results on the test data set. The second model considered is a deep network with residual architecture that learns somewhat longer than classical deep networks. The results that such a network achieves on a test data set are presented and these results are shown in realistic images.

• Keywords: strojno učenje; duboko učenje; konvolucijska mreža; super-rezolucija; rezidualna arhitektura
• Keywords (english): machine learning; deep learning; convolutional network; super-resolution; residual architecture
• Language: croatian
• URN:NBN: urn:nbn:hr:168:831813
• 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 15:02:31