Focusing Propagation Pyramid Network for Small Object Detection in Unmanned Aerial Vehicle Images

Unmanned Aerial Vehicle (UAV) images have dis-tinctive characteristics such as significant scale variance, back-ground interference, and fuzzy objects, which brings many challenges for object detection algorithms. The traditional Fea-ture Pyramid Network (FPN) cannot effectively address the aforementioned issues. Therefore, this paper proposes a novel Focusing Propagation Pyramid Network (FPPN). FPPN utilizes a customized feature focusing module and a focus-and-propagate mechanism. The customized feature focusing module can receive inputs from three scales of the backbone network and employs Inception-Style modules in its internal structure to focus on rich cross-scale information. The focus-and-propagate mecha-nism spreads feature information to different detection scales, enabling each scale to possess detailed contextual information, which is advantageous for subsequent object detection and classification tasks. Experiments demonstrate that FPPN has good performance on UA V images. In the VisDrone2023 dataset, the mean Average Precision (mAP) of YOLOv8s with FPPN is better than YOLOv8s by 7.9% and the mAP of RT-DETR has been improved by deploying FPPN, increasing by 5.8%.