Improved YOLOv8-Based Lightweight Object Detection on Drone Images

The target detection task of drones requires lightweight algorithms to fully utilize limited resources. Therefore, this paper proposes the YOLOv8-LD model. Firstly, propose the ASBiFPN neck network; Secondly, improve the detection head. Once again, introduce MPDIoU and improve the classification loss function to address the issue of imbalanced data samples. Finally, using pruning algorithms significantly reduces model volume. Improved YOLOv8-LD model in the VisDrone2019 dataset mAP@0.5 Improved by 21%. After pruning, compared with YOLOv8, the model parameters decreased by 81% and the volume decreased by 67%, mAP@0.5 increase by 3%.