Object classification subsystem as part of a UAV detection and tracking system based on computer vision methods

Abstract

A comprehensive approach to image classification based on deep learning and computer vision is proposed, including model selection, training, evaluation, and integration into application software. A comparative study of modern neural network architectures, including ResNet-50, ResNeXt-50, and EfficientNet, was conducted. ResNeXt-50 achieved the highest classification accuracy of up to 91.35%, while EfficientNet demonstrated a strong balance between accuracy and model complexity but showed lower stability on a limited custom dataset. ResNet-50 provided the most balanced results in terms of accuracy, training stability, and computational efficiency, making it the most suitable choice for practical deployment. The classification model was trained in a Python-based environment and subsequently adapted for integration into a C# application. The trained network was exported to the ONNX format, enabling inference directly within the .NET environment without requiring Python at runtime. This approach ensures efficient deployment, high performance, and seamless integration into the overall software system.