KOÇ UNIVERSITY
GRADUATE SCHOOL OF SCIENCES & ENGINEERING
ELECTRICAL AND ELECTRONICS ENGINEERING
MS THESIS DEFENSE BY OGÜN KIRMEMİŞ
Title: Learned Image Restoration and Super-Resolution
Speaker: Ogün Kırmemiş
Time: August 29, 2018, 14:00
Place: ENG B29
Koç University
Rumeli Feneri Yolu
Sariyer, Istanbul
Thesis Committee Members:
Prof. A. Murat Tekalp (Advisor, Koç University)
Assoc. Prof. Engin Erzin (Koç University)
Prof. Bilge Günsel (İstanbul Technical University)
Abstract:
It is known that deep learning is having a huge impact on every aspect of signal processing. Everyday a new architectures employing deep learning are proposed substituting human-engineered algorithms and methods. Although it is not easy to analyze a neural network to understand how it operates on inputs by looking at hidden variables, we can still analyze them by looking at how they interact with certain kinds of inputs. In this thesis, we propose two measures to assess the complexity of images so that we can predict the performance of the neural network in the context of image processing. We show that these measures can be used for explaining the behavior of the neural network on test sets. Additionally, we propose a post-processing network which enhances visual quality of BPG compressed images. We show that although we trained the network with single QP, the network is able to enhance images encoded with various QPs.