Challenges in Infrared Small-Target Detection: A Benchmark of YOLO Models on UAV and Bird Infrared Imagery

Detecting small objects in infrared (IR) images, such as birds and Unmanned Aerial Vehicles (UAVs), presents significant challenges due to their reduced size, limited pixel representation, and lack of distinct features. These challenges are further complicated in real-world environments where complex backgrounds, such as skies, clouds, and vegetation, combined with low signal-to-noise ratio (SNR) and low signal-to-clutter ratio (SCR), obscure the targets. Existing detection models often struggle to accurately detect small targets due to low contrast and indistinct edges. This highlights the need for detection techniques tailored to small object detection in infrared imagery. In this study, we evaluate the performance of various You Only Look Once (YOLO) models, focusing on their applicability and limitations for detecting small infrared objects. The novelty of this work lies in systematically analyzing the impact of preprocessing techniques and fine-tuning on detection performance. Fine-tuned YOLO models are evaluated for small target detection, with their strengths and weaknesses identified under diverse conditions. Insights are provided for adapting these models to enhance real-time surveillance systems in critical applications.