Accepted paper in IEEE Access (IEEE Press, SCIE Q1) — 2026

The proposed architecture enhances the backbone by integrating a multi-stage attention-based feature refinement strategy, composed of Context Modeling Attention (CMA), Lightweight Channel Attention (LCA), and Spatial Attention (SA). This hybrid framework strengthens the representation of salient regions and enriches channel-wise semantics through global context modeling, while simultaneously suppressing redundant background noise. By sequentially refining features along both channel and spatial dimensions, the network captures more discriminative, multi-scale information with minimal computational overhead, which is particularly effective for detecting small or occluded objects in cluttered UAV environments. To enhance model generalization, we employ a robust training pipeline that integrates advanced data augmentation with the AdamW optimizer. We also implement a cosine learning rate schedule, preceded by a warm-up phase, to stabilize the early stages of training. Experimental results on the challenging VisDrone2019 dataset demonstrate that the proposed model achieves a mAP@0.5 of 47.4% and a mAP of 29.0%, outperforming existing state-of-the-art UAV detection models. Additionally, the proposed method maintains a low computational cost of 28.566G FLOPs and achieves a real-time inference speed of 101 FPS on an RTX 4090. These findings confirm that the proposed approach provides an effective and computationally efficient solution for UAV-based object detection tasks, particularly in scenarios involving small and densely distributed objects.