Abstract
dc:descriptionIn recent decades, the advancement of technology in visual information recording systems has allowed the easy and massive collection of visual information, making its immediate management very important, especially in terms of identification. The almost instantaneous raw and high-dimensional visual information recorded in the form of pixels, in many cases needs to be recognized aiming at the extraction of lower-dimensional usable information depending on the purpose of each application. The recognition of visual information, in practice, consists of a set of methods for the analysis of the image pixels towards the automation of specific processes. Such processes may be related to the classification or retrieval of images, the detection of objects in images, objects labeling or subjects identification in them, finding specific parts of objects in images, etc.Oriented towards the management of visual information under such recognition procedures, this thesis research work included the development of intelligent computer vision techniques in digital image recognition problems and automatic optimization methods on convolutional neural networks, which are now comprising of effective tools for visual information recognition. Regarding the development of smart techniques, the element of locality of visual information was used as well as spectral techniques for its compression with applications in face and fingerprint recognition. Afterwards, techniques which are making use of non-parametric statistics on texture classification problems were examined.As convolutional neural network (CNN) technology became established to be a dominant technology relevant to visual information understanding, emphasis was placed on its exploitation. However, although there are high-quality CNN implementations in modern high-performance computing systems, in various areas of medium and low performance systems (e.g., smart cars, Internet of Things, medical devices, etc.) the development of special CNN architectures is required to operate under resource and energy constraints. In the context of this research work, a method was proposed for the dynamic optimization, in terms of energy consumption and execution speed, of pre-trained CNNs, with appropriate transformations in their structure. In the generated dynamic CNNs, only the kernels which are useful for processing the given input data at a time are activated and computed, offering great savings in the execution time as well as in the CNN's energy consumption. However, as the development of such networks requires specialized optimized software which must allow the dynamic changes on the behaviour of the CNN kernels, the research shifted to the field of searching for optimal ways for the development of such software, especially in heterogeneous embedded systems.
Degree
thesis:*- Grantor dc:publisher
- University of Patras
- Year dc:date
- 2023
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Pothos, Vasileios
- Πόθος, Βασίλειος
Subjects
dc:subject × 22- Υπολογιστική νοημοσύνη
- Επεξεργασία εικόνας
- Αναγνώριση προτύπων
- Βιομετρική αναγνώριση
- Μηχανική μάθηση
- Συνελικτικά νευρωνικά δίκτυα
- Βελτιστοποίηση εκτέλεσης
- Δυναμική αποκοπή
- Computational intelligence
- Image processing
- Pattern recognition
- Biometrics recognition
- Machine learning
- Convolution neural networks
- Inference optimization
- Dynamic pruning
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Τεχνητή νοημοσύνη
- Computer and Information Sciences
- Natural Sciences
- Artificial Intelligence
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/53363