National Technical University of Athens (NTUA)
Προηγμένοι αλγόριθμοι ψηφιακής επεξεργασίας εικόνας και ταυτοποίησης γραφέα: μελέτη του λάθους πεπερασμένης ακρίβειας σε σημαντικούς αλγορίθμους και ευσταθειοποίηση αυτών
Abstract
dc:descriptionIn the present thesis, a novel approach for the robust and efficient computation of the Zernike moments / coefficients is introduced, together with a new general methodology for the evaluation of the numerical error due to the finite world length. Moreover, a computational improvement of an existing method for pattern recognition is presented. More specifically: Initially, in the present work, it is established a new approach for the exact evaluation, accumulation and tracking of the finite precision error associated with the operation of subtraction in an arbitrary computing machine. The necessity of this study has been arised from the fact that quite many important and frequently employed algorithms seriously suffer from this type of numerical error, including various algorithms performing Zernike moments computation. Here, a new methodology is introduced, which allows for the exact tracking and prediction of the finite precision error generated in one or many successive operation of subtractions. This methodology is practically independent of the order of the quantities involved in the operations of subtraction in the considered algorithm. Furthermore, it is established that the numerical error due to the finite word length does not depend on the exact value of the employed word length. Next, a novel method is proposed for correcting the finite precision error appearing in the well - known Bernoulli and Baker’s chaotic maps. As a next step, a set of novel recursive relations for the evaluation of the Zernike moments is established. These moments incorporate important information concerning an image and, in particular, the one associated with the contour and details of the objects existing in the image. For these reasons, the Zernike coefficients are widely used in biometrics, pattern recognition, image segmentation and / or compression, etc. However, the representation of an image via Zernike moments includes various considerable numerical errors, such as the geometric one and the integration error; these numerical errors highly deteriorate the quality of the reproduced image by means of the Zernike coefficients. Employing polar pixels lifts most of the aforementioned numerical problems. Hence, towards this direction, in the present thesis a new ensemble of iterative relations is introduced and proved for the efficient computation of the Zernike coefficients, on the basis of polar pixels. Application of these recursive formulae in the evaluation of the Zernike moments offers excellent quality of the corresponding introduced images, since they lift the geometric and integration errors. In addition, the presented method is independent of the rotation of the image and quite efficient, while it practically does not suffer from finite precision error. Finally, in the last section of the present work, it is established a new computational improvement of an already existing method for the identification of the writer of important historical documents. The associated, developed software system has been applied to a first study of the important open question, concerning the writer of the documents “Saganaki tis trelas” and “Fysikis apanthisma”. The open issue is the eventuality that the great Rigas Feraios has written both documents.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2021
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Chalatsis, Konstantinos
- Χαλάτσης, Κωνσταντίνος
Subjects
dc:subject × 18- Αλγόριθμοι υπολογισμού των συντελεστών Zernike moments
- Πολικό εικονοστοιχείο
- Γένεση λάθους πεπερασμένης ακρίβειας
- Συσσώρευση λάθους πεπερασμένης ακρίβειας
- Ακριβής προσδιορισμός λάθους πεπερασμένης ακρίβειας στη πράξη της αφαίρεσης
- Αυτόματη αναγνώριση γραφέα
- Algorithms for the computation of Zernike moments
- Polar pixel
- Finite precision error generation
- Finite precision error accumulation
- Exact tracking of finite precision error in the operation of subtraction
- Automatic writer identification
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Υπολογιστές, Υλικό (hardware) και Αρχιτεκτονική
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Computer science, Hardware and Architecture
Rights
dc:rights- Statement dc:rights
-
- BY_NC_ND
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/49183