Runtime comparison between DPCNet and YOLO11n.

Small object detection in unmanned aerial vehicle imagery is challenged by tiny target scales, dense layouts, and cluttered backgrounds that blur fine details and destabilize multiscale representations. We present DPCNet, a single-stage detector that combines dual-path cross perception with deep and shallow feature interaction and a decoupled detection head. The Dual-Path Cross Perception block separates a detail stream and a semantic stream and performs gated bidirectional fusion, preserving edges while enriching context. The Deep and Shallow Feature Interaction block aligns features across levels through dynamic up-sampling and down-sampling and similarity-guided masking, which strengthens cross-scale consistency. The Dual-Path Decoupled Detection Head keeps classification and regression separate yet enables lightweight cross-branch channel and spatial guidance, and bounding-box regression adopts a geometry-sensitive Shape-IoU loss. Experiments on VisDrone2019 and HIT-UAV show consistent gains over the YOLO11n baseline: DPCNet improves mAP@0.5 by 2.0% and 5.1%, respectively, with higher precision and recall, especially for small, dense, low-light, and occluded targets. Despite modest computational overhead from cross-path interactions, the parameter count is reduced by about 45%, indicating a compact and robust solution for small object detection in challenging UAV scenarios.