LSOD-YOLOv8s: A Lightweight Small Object Detection Model Based on YOLOv8 for UAV Aerial Images

Detecting small objects in aerial drone imagery is an extremely challenging research topic. This is primarily because the target size is relatively small, the background is complex, and occlusion occurs easily, which leads traditional object detection models to struggle in achieving ideal detection results. To enhance the detection performance of small objects, this paper proposes a lightweight small object detection model for aerial images captured by drones. Based on YOLOv8, this model adds a small object detection layer, introduces the FasterNet Block and dynamic upsampling method to optimize the network structure, and designs an Inner-WIoU loss to improve the localization accuracy of small objects. Evaluations on the VisDrone2019 and UAVDT datasets illustrate that the LSOD-YOLOv8s model surpasses the original YOLOv8s in average precision at an IoU of 0.5 by 6.3% and 3.3%, respectively, while achieving a 75% reduction in model parameters. Compared to other advanced models, LSODYOLOv8s not only possesses the fewest parameters and highest average precision, but also significantly reduces false detection and miss detection rates, meeting the demands of real-time detection for UAVs.