Research on Small Target Detection Algorithm for UAV Aerial Images Based on YOLO

To address high missed detection and false alarm rates caused by small targets, drastic scale variations, and complex background interference in UAV aerial images, this paper proposes DST-YOLO, an improved algorithm based on YOLOv11. First, Space-to-Depth Convolution (SPDConv) is introduced in the Backbone to replace original downsampling layers, losslessly converting spatial information into the depth dimension to preserve fine-grained details of tiny targets. Second, a Semantic and Detail Injection (SDI) module is embedded in the Neck. By utilizing attention mechanisms, it integrates shallow details with deep semantics to bridge the semantic gap and enhance discrimination in complex backgrounds. Finally, a P2-ASFFHead module is constructed by incorporating a high-resolution P2 feature layer and an Adaptive Spatial Feature Fusion (ASFF) mechanism, enabling the model to dynamically prioritize sparse small target features. Experimental results on the VisDrone2019 dataset demonstrate that DST-YOLO increases mAP50 and mAP50:95 by 8.5% and 5.6%, respectively, compared to the baseline YOLOv11n. The model significantly improves detection performance in challenging scenarios such as vertical angles, night scenes, and long-distance views, better meeting the requirements of UAV-based detection tasks.