Back to search

National and Kapodistrian University of Athens

Model-based and data-driven approaches meet redundancy in signal processing

Abstract

dc:description

The present dissertation is divided in two parts. In the first one, we address two inverse problems, namely compressed sensing (CS) and speech denoising, through the lens of the analysis sparsity model. To that end, we leverage ideas from time-frequency analysis and introduce a new redundant analysis operator associated to a Gabor frame, which efficiently sparsifies the signals of interest. Subsequently, we employ an iterative method for solving both problems and perform a numerical comparison of our analysis operator to state-of-the-art Gabor analysis operators, on both synthetic and real-world data. Our experimental results indicate improved performance when our proposed framework is employed to solve both inverse problems, since it outperforms all other Gabor transforms, consistently for all datasets. In the second part of this dissertation, we resolve the CS problem by employing the newly-introduced field of deep unfolding, which stems from the interpretation of classic iterative algorithms as deep neural networks. In order to do so, we ``unfold'' the iterations of two state-of-the-art algorithms, traditionally employed to solve the analysis-sparsity-based CS problem, into layers of neural networks. The latter jointly learn corresponding decoders and redundant sparsifying analysis operators. Our main novelty is attributed not only to the development of these two unfolding networks, but also to the estimation of the generalization error produced by each one of them. Regarding the latter, we take advantage of chaining techniques, in order to upper bound the Rademacher complexity of each network's hypothesis class. Then, we leverage this estimate so as to deliver the desired generalization error bounds. Finally, we conduct several experiments -- on synthetic and real-world datasets -- validating our derived theory and compare our proposed unfolding networks to a state-of-the-art one. As illustrated by the experimental results, both our unfolding networks outperform the baseline, consistently for all datasets.

Degree

thesis:*
Grantor dc:publisher
National and Kapodistrian University of Athens
Year dc:date
2023

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Kouni, Vassiliki
  • Κούνη, Βασιλική

Subjects

dc:subject × 32

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/54805

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Kouni, Vassiliki; Κούνη, Βασιλική. Model-based and data-driven approaches meet redundancy in signal processing. National and Kapodistrian University of Athens, 2023. http://hdl.handle.net/10442/hedi/54805