National Technical University of Athens (NTUA)
Τεχνικές μηχανικής μάθησης για διαχείριση γνώσης σε συστήματα ανάκτησης εικόνων
Abstract
dc:descriptionThe rapid increase in the number of available digital images during recent years has established the problem of image retrieval as one of the most interesting and popular problems in the field of computer applications. The current thesis addresses the problem of image retrieval by employing machine learning techniques. In particular, we mainly consider content-based image retrieval, the most prominent characteristic of which is the existence of the semantic gap between the digital content of the images and the semantic criteria of the users. In this framework, we focus on the technique of relevance feedback. Our goal is the design and implementation of new methodologies, aiming at the incorporation of the information which comes from the evaluation provided by the user on the current results, in order to sequentially improve them. In what concerns image representation, we were based mainly on Gaussian Mixtures, for the description of the distribution of the locally extracted feature vectors. The most important issue regarding Gaussian Mixtures in this context is the use of a distance measure that can be computed efficiently. For this reason, we proceeded to a thorough examination on popular and recently proposed distance measures between Gaussian Mixtures. Afterwards, two new relevance feedback methods were implemented, which use Gaussian Mixtures for the representation of the images and distance measures that can be computed in a fast way. The first method is based on a probabilistic technique to update the model of the user’s query according to the user’s examples. On the contrary, the second method is based on classifiers to distinguish between the images that are relevant to the user’s query and those that are irrelevant. Finally, in what concerns the case of representing each image using a multi-dimensional vector, the possibility of employing feature selection methods was investigated. The aim of this approach is the increase of the classifier ability to distinguish between the relevant and irrelevant images. At every stage of the research and for each proposed method, the results of appropriately designed and performed experiments and comparisons are provided. From these experiments, the large possibilities of improvement of the image retrieval results by using relevance feedback are demonstrated. Furthermore, the success of the proposed methods and the advantages of using Gaussian Mixtures for image modeling are proved.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2009
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Μαρακάκης, Απόστολος
- Marakakis, Apostolos
Subjects
dc:subject × 20- Μηχανική μάθηση
- Ανάκτηση εικόνων
- Συσχετιστική ανατροφοδότηση
- Γκαουσιανά μείγματα
- Ταξινομητές
- Επιλογή Χαρακτηριστικών
- Σημασιολογικό κενό
- Μηχανές διανυσματικής υποστήριξης
- Machine learning
- Image retrieval
- Relevance feedback
- Gaussian Mixture Models
- Classifiers
- Feature Selection
- Semantic gap
- Support vector machines
- Επιστήμες Μηχανικού και Τεχνολογία
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Engineering and Technology
- Electrical Engineering, Electronic Engineering, Information Engineering
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/17755