DRS-YOLO: an improved YOLOv8-based small object detection model for UAV aerial imagery

Abstract In unmanned aerial vehicle (UAV) imagery, the high proportion of small objects and limited computational resources pose significant challenges for object detection, making it difficult for conventional methods to balance accuracy and efficiency. To enhance small object detection performance, this paper proposes an improved YOLOv8-based model, named DRS-YOLO. The model incorporates spatial depth convolution to improve feature retention during downsampling for better perception of small objects. A Downsampling Compensation and Dual-path Fusion module is introduced, integrating the path aggregation feature pyramid network structure, a hybrid downsampling strategy via the DownSimper component, and adaptive upsampling using the DySample mechanism, enabling efficient cross-scale information fusion. Additionally, the paper proposes a refined feature extraction module, RepDNeckELAN4, which builds on the cross stage partial architecture by integrating Reparameterized Convolution and the efficient layer aggregation network, and further introduces multi-scale dilated convolution paths to enhance local feature extraction and improve detection accuracy under complex backgrounds. In the detection head, a 160 × 160 resolution branch is added to strengthen the recognition of tiny objects, while the 20 × 20 branch is pruned to reduce computational overhead and improve inference efficiency. Experimental results on the VisDrone2019 dataset show that, compared with the baseline YOLOv8s model, the proposed DRS-YOLO achieves significant improvements, with mAP@0.5 increased by 15.4% and mAP@0.95 increased by 10.8%, demonstrating its effectiveness in improving small object detection in UAV imagery.