YOLOv8s-LR: an improved multiscale feature fusion network for small object detection in UAV images

Abstract

To address the problems of small object scale, dense distribution, frequent occlusion, and complex backgrounds in UAV-view images, this paper proposes an improved small object detection network based on YOLOv8s, named YOLOv8s-LR. The proposed method introduces a P2 small object detection branch into the original detection structure to enhance the representation capability of high-resolution shallow features for small objects. An LGSCDown downsampling module is designed to reduce the loss of detailed information during downsampling. An RLSCF feature fusion module is constructed to strengthen the cross-scale interaction between shallow detailed features and deep semantic features. In addition, the R-DWMPDIoU regression loss function is introduced to improve the stability of bounding box localization. Experimental results on the VisDrone2019 dataset show that YOLOv8s-LR achieves Precision, Recall, mAP50, and mAP50-95 of 56.6%, 44.5%, 46.4%, and 28.1%, respectively, which are 3.1, 4.1, 4.5, and 3.1 percentage points higher than those of the baseline YOLOv8s. The results demonstrate that the proposed method can effectively improve small object detection performance in complex UAV scenarios and achieve better overall detection performance with only a slight increase in the number of parameters.