Improved Object Detection in UAV Images using Deep Learning

The use of unmanned aerial vehicles (UAV) for computer vision analysis is a significant trend in the current scenario. UAV technology is highly utilized for various purposes, including object detection, tracking, traffic management, environment monitoring, and agriculture sector, mainly due to the ease of data collection compared to conventional remote sensing methods through satellites. This study focuses on enhancing the YOLOv5 architecture to effectively detect small targets. The modifications made to the YOLOv5 framework specifically target the architecture, resulting in improved performance in object identification. The addition of a new feature fusion layer within the feature pyramid section of YOLOv5 plays a crucial role in achieving these improvements. To maintain resolution and prevent the loss of valuable feature information in the deeper sections of the network, a lateral connection is introduced, connecting this layer to an earlier part of the network. This addition ensures that crucial details and feature data are preserved throughout the network architecture. Additionally, data augmentation techniques such as image saturation and cropping are employed.