Globoka nevronska mreža za ostrenje slik z glajenjem

Gornik, Maja (author), Kristan, Matej (mentor)

V diplomskem delu se ukvarjamo s problemom ostrenja zamegljenih slik. Za reševanje tega problema predlagamo nov pristop k ostrenju slik, ki temelji na globoki konvolucijski nevronski mreži z dodanimi filtri za glajenje slik (UnsharpNet). Predstavimo več različic osnovnega modela, ki se med seboj razlikujejo glede na velikost in pozicijo uporabljenih filtrov. Rezultate modelov nato ocenimo kvantitativno in kvalitativno na podatkovni zbirki GOPRO ter analiziramo vpliv velikosti filtrov in njihove pozicije v arhitekturi mreže na dobljen rezultat. Eksperimentalni rezultati kažejo, da so dobljenimi modeli praviloma bolj uspešni kot osnovni model, ki ne vsebuje dodatnih filtrov. Rezultat osnovnega modela uspemo izboljšati za 1.1% glede na oceno PSNR in 0.68% glede na SSIM. Najboljšo različico modela UnsharpNet nato primerjamo še s sorodnimi deli. Rezultati kažejo, da kljub enostavnosti naša metoda dosega rezultate, ki so primerljivi z najnovejšimi metodami iz tega področja.

• Language: Slovenian
• Keywords: računalniški vid, konvolucijska nevronska mreža, ostrenje slik
• Work type: Bachelor thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2019
• PID: 20.500.12556/RUL-109874
• COBISS.SI-ID: 1538344387
• Publication date in RUL: 09.09.2019

Secondary language

• Language: English
• Title: A deep neural network for deblurring by blurring

Abstract

In this thesis we address the problem of image deblurring. We propose a new approach that is based on a deep convolutional neural network with added filters for image smoothing (UnsharpNet). We present several different models that differ from each other in size and positioning of added filters. We evaluate developed models both quantitatively and qualitatively on the GOPRO dataset and also analyze the impact of filter sizes and positions in the model architecture on achieved results. Experiments show that presented methods generally achieve better results than our baseline method, which does not have any additional filters. We improved the results of our baseline method by 1.1% in terms of PSNR and 0.68% in terms of SSIM. We then compare our best model with related work and show that despite simplicity our method achieves results that are comparable to current state of the art image deblurring methods.

• Keywords: computer vision, convolutional neural network, image deblurring