DSTD-YOLOv8: Enhanced YOLOv8 with Dense Small Target Detection Model for UAV Aerial Images

To address the challenges of insufficient detection accuracy for small objects in drone aerial imagery due to feature loss, and the existing YOLOv8 model's difficulty balancing lightweightness and detection performance in complex scenarios, this paper proposes a dense small object detection enhancement model, DSTD-YOLOv8, based on YOLOv8n. First, a multibranch feature fusion C2fSCDownC module is designed. This significantly reduces model parameters and computational overhead while preserving shallow, high-resolution features through residual connections, preventing feature degradation of small objects. Second, a new P2- layer small object detection head is added, which is deeply fused with the Backbone output features to address the semantic deficiencies of shallow features and the blurred details of deep features. Finally, the MPDIoU loss function replaces the traditional CIoU loss, incorporating bounding box vertex distance and input image width and height factors to optimize bounding box regression accuracy for non-overlapping scenes and objects of varying scales. Experiments on the VisDrone2019 dataset show that DSTD-YOLOv8 achieves 46.71% Precision, 37.04% Recall, and 36.29% mAP50. The mAP50-95 reaches 21.46%, which is 3.00%, 3.69%, 3.41% and 2.50% higher than YOLOv8 respectively. It is also superior to mainstream models such as YOLOv5, YOLOv10, YOLOv11 and YOLOv13. This proves that the model effectively improves the accuracy of small target detection in complex scenes and fully meets the requirements of dense small target detection in drone aerial images.