An Enhanced YOLOv5 Model for Small Object Detection in UAV Scenarios

In the domain of object detection, identifying small objects presents a significant challenge. While numerous approaches have been proposed to enhance performance in this area, such as incorporating multiple attention blocks or modifying feature fusion network structures, these methods often incur substantial computational costs, rendering real-time detection systems impractical. To address these limitations, we introduce an enhanced YOLOv5 model, termed HSD-YOLOv5. This model integrates several key innovations: firstly, it incorporates an additional prediction head specifically tailored for small objects, which provides a higher-resolution feature map to improve prediction accuracy. Secondly, it includes a SE_LSKblock module positioned between the backbone and neck, aimed at optimizing the channel information of the feature map. Additionally, we employ a Distance-IoU-based Non-Maximum Suppression (DloU-NMS) technique to refine the detection process by taking into account both distance and overlap between targets. Our results demonstrate that HSD-YOLOv5 has improved mAPO.5 by 14.8% and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$mAP_{0.50.95}$</tex> by 11.8% on the VisDrone2019 dataset.