Ανίχνευση και ταξινόμηση διαστημικών απορριμμάτων σε δορυφορικές εικόνες με χρήση συνελικτικών νευρωνικών δικτύων και μετασχηματιστών όρασης
Year: 2026
Type: dissertation
Abstract: The accumulation of orbital debris represents one of the most significant threats to the long-term sustainability of space operations, making reliable detection and classification of space objects a critical component of Space Situational Awareness (SSA).This thesis investigates and compares two state-of-the-art deep learning object detection architectures -YOLOv8n, based on Convolutional Neural Networks (CNNs), and RT-DETR-L, based on Transformer attention mechanisms -in the context of automated space object detection and classification.Both models were trained using transfer learning on the Space Debris and Satellite Dataset -a dataset of 2,467 images distributed across 11 classes -under identical conditions with respect to dataset, hyperparameters, and hardware, enabling a direct and objective comparison.Evaluation was performed using Precision, Recall, [email protected], and [email protected]:0.95metrics at both overall and per-class level, alongside inference speed measurement.The results demonstrate that RT-DETR-L achieves superior overall detection accuracy in most experimental configurations, while YOLOv8n outperforms in bounding box localization accuracy and inference speed.Both models encounter difficulties in detecting small, low-contrast objects, while qualitative analysis reveals misclassification and duplicate detection phenomena not captured by quantitative metrics.Overall, the findings indicate that optimal architecture selection depends on application requirements, with RT-DETR-L being preferable in high-accuracy scenarios and YOLOv8n proving more suitable for real-time processing applications.
Author: Τηλιού, Αικατερίνη Γ.
Language: English
Subfield: Aerospace Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7210086016