BAF-Net: Efficient Small-Target Detection in UAV Aerial Images via Hierarchical Feature Fusion and Bidirectional Attention

Object detection serves as a cornerstone task in computer vision with significant applications in UAV remote sensing. Addressing prevalent challenges of insufficient feature representation for small objects and complex background interference in aerial imagery, this paper proposes an enhanced architecture based on YOLOv11n. We innovatively design a C3k2_MAFEP module that hierarchically decomposes and dynamically fuses features through a triple-convolution pyramid structure to enhance multi-scale perception. Additionally, a high-resolution P2 detection head extends the top layer of the feature pyramid, substantially improving structural capture capability for minute targets. The framework further incorporates a Bidirectional Attention Fusion (BAF) module with gated attention mechanisms to optimize cross-scale feature interaction. Experimental evaluations on the VisDrone2019 dataset demonstrate that the improved model achieves \(34.0\%\) in mAP@0.5 and \(19.9\%\) in mAP@0.5:0.95 while maintaining controlled parameter count and computational overhead, representing \(7.5\%\) and \(5.0\%\) improvements over the baseline YOLOv11n, and \(2.4\%\) and \(1.5\%\) enhancements compared to YOLOv11s. This study provides an efficient solution for UAV detection systems.