Abstract:Target detection from a UAV perspective often faces challenges such as small object size, severe occlusion, and significant scale variation, which result in missed and false detections. Moreover, constrained by the limited computational capacity of UAV platforms, achieving real-time detection with both high accuracy and lightweight deployment remains difficult. To address these challenges, a lightweight version of YOLOv8 is proposed. The image encoder of MobileSAM, a lightweight adaptation of the segment anything model (SAM), is integrated as the backbone network, enabling effective extraction of multi-scale features, improved detection efficiency for small targets, and enhanced generalization across diverse tasks and datasets. A lightweight detection head, termed lightweight scalable shared convolutional detection head (LSCD), is proposed. Based on shared convolution and adaptive feature scaling, LSCD reduces the parameter count and computational overhead while maintaining detection accuracy. In addition, the normalized Gaussian Wasserstein distance (NWD) loss is employed to further improve detection performance on small objects. Experimental results on the VisDrone-DET2019 dataset demonstrate that the proposed algorithm significantly enhances both precision and recall for small targets compared to the original YOLOv8s model. The mAP50 is increased by 3.2%, reaching 41.4%, while the number of parameters is reduced by 33.9%. On the DOTA v1.0 dataset, the proposed model achieves a mAP50 of 48.8%, representing an 8% improvement, indicating strong generalization capability.