Aristotle University Of Thessaloniki (AUTH)
Τεχνικές μηχανικής όρασης για τον χαρακτηρισμό εικόνων προσώπου
Abstract
dc:descriptionThe main field of interest of this thesis is the characterization of facial images, such as facial expression recognition, face recognition and facial pose recognition. One of the most crucial issues that every facial image analysis algorithm in computer vision encounters is the high dimensionality of the provided image data. A popular category of methods that obtain a more manageable problem is the subspace image representation algorithms, which aim to discover the latent image features by projecting the high dimensional input data to a low dimensional subspace. In this thesis two subspace learning methods are presented, that aim to enhance classes discrimination in the projection subspace. The first method modifies the cost function of the original Nonnegative Matrix Factorization algorithm by incorporating appropriate discriminant factors inspired by Clustering Discriminant Analysis, while the second operates on the random features instead of the actual high dimensional input data and derives an optimal projection matrix, such that the separating margin between the projected samples of different classes is maximized. Experimental results for face recognition and facial expression recognition on various popular datasets shown that the developed algorithms exploiting the discriminant low dimensional facial image representations achieved superior recognition performance compared against various competing algorithms. Subsequently, an algorithm for the incremental training of multiclass support vector machines is presented, which exploits a set of novel multiplicative update rules and a warm start optimization framework, in order to achieve computational efficiency, when the training data collection is sequentially enriched and dynamic adaptation of the classifier is required. Experimental results on various data collections and for facial pose recognition, verified that the developed method is faster than retraining the classifier from scratch, while the achieved classification accuracy rate is maintained at the same level.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2013
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Nikitidis, Symeon
- Νικητίδης, Συμεών
Subjects
dc:subject × 14- Μηχανική όραση
- Εκμάθηση υποχώρων
- Αναγνώριση προσώπου
- Αναγνώριση εκφράσεων προσώπου
- Αναγνώριση πόζας προσώπου
- Computer vision
- Suibspace learning
- Face recognition
- Facial expression recognition
- Facial pose recognition
- Φυσικές Επιστήμες
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Natural Sciences
- Computer and Information Sciences
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/36706