A Lightweight YOLO-Based Model in Small-Object Detection for AAV Optical Sensors

Detecting small objects within extensive remote sensing images (RSIs) captured by autonomous aerial vehicle (AAV) optical sensors is crucial for a variety of military and civilian purposes, but it poses considerable challenges. In order to address the difficulties associated with the numerous parameters and the high computational requirements of existing object detection models, as well as the inadequate performance in identifying small objects in RSIs captured by AAV optical sensors, this article proposes a lightweight model for small-object detection. First, Focal scalable intersection over union (SIoU) loss derived by integrating the concept of Focal loss into the traditional SIoU loss. Second, this study uses channel attention (CA) and GhostNet to reconstruct the backbone. A layer specifically designed for detecting small objects is incorporated into YOLOv5s using a depth residual network neck. This study, furthermore, establishes a small fire dataset and evaluates the YOLOv5s model on the small fire, AI-TOD dataset and DOTA dataset. Experimental results show that the proposed model achieves the lightweight goal while maintaining high detection performance for small objects.