DRONE VS. BIRD CLASSIFICATION AND APPLIED COLOR TRANSFER AUGMENTATION

Abstract

Unmanned Aerial Vehicles (UAVs), commonly known as drones, have rapidly transformed the landscape of both civilian and military operations. Their usage spans a wide range of applications—from commercial photography and agriculture to tactical reconnaissance and parcel delivery. However, their growing accessibility and capabilities pose significant threats in terms of unauthorized surveillance, smuggling, and potential airborne attacks. A critical concern in automated airspace monitoring is the accurate identification of drones, particularly when differentiating them from birds—natural flying entities that share similar aerial dynamics, shapes, and appearances. Distinguishing between drones and birds in aerial images is inherently complex due to overlapping visual characteristics and varied environmental conditions, such as lighting, motion blur, or sensor noise. This thesis addresses this challenge by proposing a deep learning-based classification framework using advanced convolutional neural network architectures. The system employs ResNet-50 and ResNet-152 through transfer learning, allowing efficient feature extraction and training on limited datasets. Furthermore, color transfer techniques are applied to improve model generalization across different domains, especially when dealing with infrared imagery. Detailed experimentation was conducted using original and color-transferred datasets, comprising a balanced distribution of drone and bird imagery. The models demonstrated high classification accuracy, with ResNet-152 achieving superior performance due to its depth. This research contributes to a scalable, adaptable and defensible approach to drone detection and bird classification, laying the foundation for deployment in real-time surveillance and autonomous security systems.

Date of Award	2025
Original language	American English
Awarding Institution	HBKU College of Science and Engineering