Detection of military aircraft with deep learning architectures

Recent advancements in technology have significantly enhanced the applications of computer vision and deep learning methods, including in the defense industry. The detection of military aircraft is crucial for military operations. Rapid and accurate identification of enemy aircraft improves air defense systems and enables faster strategic decision-making. Developing algorithms capable of high-accuracy and rapid object detection is a crucial need in defense technologies. Moreover, these technologies are vital for airspace security and managing civilian air traffic. Advances in computer vision and deep learning minimize human error in detecting military aircraft, facilitating the development of automated and reliable systems. Therefore, evaluating and improving the performance of deep learning-based object detection algorithms has become an important area of research for both academic and industrial applications. This thesis examines and compares modern deep learning-based object detection algorithms for identifying military aircraft. The study employs YOLOv7, YOLOv8, and RT-DETR models to detect different types of military aircraft. The dataset used consists of images of 43 different types of military aircraft captured from various angles and backgrounds. During the training and testing of the models, the effects of hyperparameters on performance were thoroughly analyzed. The experimental results show that the RT-DETR model performs more consistently with 92.7% mAP and 90.4% recall. The YOLOv8 model achieves the highest mAP at 94%, but falls behind RT-DETR with 88.1% recall. The YOLOv7 model has 90.2% mAP and 82.7% recall.