Back to search

University of Western Macedonia

Σχεδιασμός και υλοποίηση εφαρμογών σε ετερογενή υπολογιστικά συστήματα

Abstract

dc:description

In recent years, there has been a significant growth in the design and implementation of applications that differ from the traditional architecture of personal computers, servers, and laptops.These architectures correspond to heterogeneous computing systems (HCS), which integrate and combine different types of processing units in order to accelerate demanding applications with high computational complexity, while also offering improved energy efficiency and more efficient utilization of available computing resources. Applications characterized by high computational complexity constitute a primary domain of deep learning (DL), which has brought significant advancements and introduced new capabilities in artificial intelligence (AI) research. DL has emerged as one of the most dynamic and effective approaches for solving complex problems by leveraging the power of neural networks to learn from massive datasets and generate highly accurate predictions. Particular emphasis has been placed on applications in Computer Vision (CV), which have greatly benefited from advances in DL techniques. Breakthroughs in this area have led the research community to adopt deeper and more complex models, enhancing the ability of systems to recognize and interpret visual data with high precision. These applications include, among others, pattern recognition, image analysis, and video understanding, contributing substantial improvements in domains such as medical diagnosis, autonomous driving, and security monitoring. This dissertation focuses on the design, implementation, and acceleration of high-complexity DL models, with emphasis on Convolutional Neural Networks (CNNs), on heterogeneous architectures.To achieve these goals, we propose a methodology for the development of accelerated and specialized architectures suitable for applications on HCS. The research aims to design and optimize architectures that fully exploit the capabilities of advanced HCS, such as Field-Programmable Gate Array – System on Chip (FPGA-SoC) and System on Module (SoM) devices. The main objective is to enhance computational efficiency through neural model acceleration, while ensuring the adaptability of the systems to a broad range of applications. By leveraging parallel processing, programmability, and architectural specialization offered by FPGA-SoCs and SoMs, such systems can meet the increasing demands of DL applications. This strategy enables high-performance solutions with reduced environmental footprint, making HCS ideal for scenarios requiring a balance between computational power, accuracy, and energy efficiency. The evaluation of the proposed architectures, as implemented through the methodology we propose, is conducted on real-world datasets, focusing on two critical applications: medical diagnosis and the agricultural sector. Both of these domains are characterized by high processing demands and the need for accurate and timely data analysis. In medical diagnostics, the rapid processing of large-scale datasets such as X-ray images enables healthcare professionals to detect diseases early and make improved clinical decisions. In agriculture, image analysis of crops supports efficient resource management and early disease prevention. The results of the present dissertation demonstrate the efficiency of the proposed methodology and the developed architectures, highlighting their flexibility to adapt to the diverse requirements of various applications. The exploitation of HCS allows for flexible allocation of computational resources based on data complexity and volume, ensuring high throughput, reduced latency, and energy-efficient processing. The proposed architectures significantly contribute to the realization of high-performance solutions in domains such as healthcare, agriculture, and other industrial applications, ultimately aiming to achieve a well-balanced trade-off among accuracy, speed, and energy footprint.

Degree

thesis:*
Grantor dc:publisher
University of Western Macedonia
Year dc:date
2025

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Sanida, Theodora
  • Σανιδά, Θεοδώρα

Subjects

dc:subject × 18

Rights

Language dc:language
gre

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/59978

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

Sanida, Theodora; Σανιδά, Θεοδώρα. Σχεδιασμός και υλοποίηση εφαρμογών σε ετερογενή υπολογιστικά συστήματα. University of Western Macedonia, 2025. http://hdl.handle.net/10442/hedi/59978