Department of Informatics and Telecommunications
Πληροφορική

2017-05-18

2017

Xenaki Spyridoula-Irida

Σέργιος Θεοδωρίδης - Καθηγητής - Πληροφορικής και Τηελεπικοινωνιών - ΕΚΠΑ,
Κωνσταντίνος Κουτρούμπας - Κύριος Ερευνητής - ΙΑΑΔΕΤ - ΕΑΑ,
Αθανάσιος Ροντογιάννης - Διευθυντής Ερευνών - ΙΑΑΔΕΤ - ΕΑΑ,
Κωνσταντίνος Μπερμπερίδης - Καθηγητής - Μηχανικών Η/Υ & Πληροφορικής - Παν. Πάτρας,
Δημήτριος Μαρούλης - Καθηγητής - Πληροφορικής και Τηελεπικοινωνιών - ΕΚΠΑ,
Δημήτριος Γουνόπουλος - Καθηγητής - Πληροφορικής και Τηελεπικοινωνιών - ΕΚΠΑ,
Ηλίας Μανωλάκος - Καθηγητής - Πληροφορικής και Τηελεπικοινωνιών - ΕΚΠΑ,

Advances in Possibilistic Clustering with Application to Hyperspectral Image Processing

English

Advances in Possibilistic Clustering with Application to Hyperspectral Image Processing

Clustering is a well established data analysis methodology that has been extensively used in various fields of applications during the last decades. The main focus of the present thesis is on a well-known cost-function optimization-based family of clustering algorithms, called Possibilistic C-Means (PCM) algorithms. Specifically, the shortcomings of PCM algorithms are exposed and novel batch and online PCM schemes are proposed to cope with them. These schemes rely on (i) the adaptation of certain parameters which remain fixed during the execution of the original PCMs and (ii) the adoption of sparsity. The incorporation of these two characteristics renders the proposed schemes: (a) capable, in principle, to reveal the true number of physical clusters formed by the data, (b) capable to uncover the underlying clustering structure even in demanding cases, where the physical clusters are closely located to each other and/or have significant differences in their variances and/or densities, and (c) immune to the presence of noise and outliers. Moreover, theoretical results concerning the convergence of the proposed algorithms, also applicable to the classical PCMs, are provided. The potential of the proposed methods is demonstrated via extensive experimentation on both synthetic and real data sets. In addition, they have been successfully applied on the challenging problem of clustering in HyperSpectral Images (HSIs). Finally, a feature selection technique suitable for HSIs has also been developed.

Digital Signal and Image Processing

Computer science

possibilistic clustering, parameter adaptation, sparsity, cluster elimination, convergence analysis, online clustering, feature selection, hyperspectral image processing

Yes

4

Yes

96

202

https://pergamos.lib.uoa.gr/uoa/dl/object/1518963/file.pdf

https://pergamos.lib.uoa.gr/uoa/dl/object/1518963