Back to search

Aristotle University Of Thessaloniki (AUTH)

Τεχνικές ομαδοποίησης και ταξινόμησης με εφαρμογές στην ανθρωποκεντρική ανάλυση εικόνων και εικονοσειρών

Abstract

dc:description

In this doctoral thesis, we present the results of the research conducted in the fields of data clustering and classification with applications in anthropocentric analysis of images and video. Our main contribution to the field of classification consists of the introduction of an image searching data structure based on binary search trees for quick shape matching. According to the properties of a balanced binary search tree, we can insert, search and delete shapes in logarithmic time. In each node of the tree, a weak classifier decides which path it should direct the search. In its initial form, the tree is constructed incrementally. Starting from an empty tree, the data structure is filled in by inserting the shapes into the tree. The classifiers in each node are trained using a simple and quick training algorithm. The evolution of these tree takes advantage of the entire training set, aiming to improve the speed and performance of the proposed data structure. The construction of the tree and the training of the nodes are implemented with more sophisticated algorithms. The training algorithm of the classifiers maximizes the a posteriori probability that an input shape is directed to the correct subtree based on the shape templates that are contained in each subtree. The tree construction algorithm organizes the shape templates in such a way, that the training of each classifier is as easy as possible, while maintaining the balance of the tree. Consequently, we present our contributions to the field of similarity matrix based data clustering. Our first and most important contribution in this field concerns the trimming of the kernel matrix used by the kernel k-means algorithm. The proposed trimming algorithm attempts to estimate the cardinality of the cluster that each sample belongs to and trims the corresponding matrix row based on this estimation. This results in the reduction of the kernel matrix size and the improvement of the performance of the kernel k-means algorithm. Also described is the distributed implementation of the proposed algorithm. Furthermore, we show how the computation of the kernel matrix can be performed very fast on modern CPUs using specialized software. Finally, we present a stochastic graph clustering method based on the discreet cosine transform.

Degree

thesis:*
Grantor dc:publisher
Aristotle University Of Thessaloniki (AUTH)
Year dc:date
2015

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Tsapanos, Nikolaos
  • Τσάπανος, Νικόλαος

Subjects

dc:subject × 12

Rights

Language dc:language
gre

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/38506

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

Tsapanos, Nikolaos; Τσάπανος, Νικόλαος. Τεχνικές ομαδοποίησης και ταξινόμησης με εφαρμογές στην ανθρωποκεντρική ανάλυση εικόνων και εικονοσειρών. Aristotle University Of Thessaloniki (AUTH), 2015. http://hdl.handle.net/10442/hedi/38506