Taming the Tiny: A Lightweight Hierarchical Attention Network for Small Object Detection in UAV Aerial Images

This paper proposes a lightweight object detection model called MSAD-YOLO, designed to address the challenges of small object detection in UAV aerial images, such as small scale, sparse features, complex backgrounds, and limited computational resources. Building on the YOLOv11n baseline, the model incorporates four key improvements: (1) A progressive Hierarchical Perception Dynamic Attention (HPDA) module is designed, which utilizes four variants (MS, SD, BF, CE) to achieve adaptive fusion from spatial details to semantic information, thereby enhancing the feature representation of small objects. (2) A Decoupled Downsampling Block (DDB) is introduced, which decouples spatial compression from channel expansion and significantly reduces its parameters and computational complexity. (3) The detection head is reconstructed as a shallow detection head structure, leveraging high-resolution feature maps to improve the perception of small objects. (4) The WIoU v3 loss function replaces the original CIoU loss, which optimizes the bounding box regression through a dynamic non-monotonic focusing mechanism, and further enhances the localization accuracy for small objects. Experimental results on the VisDrone2019 dataset demonstrate that MSAD-YOLO outperforms the baseline YOLOv11n, achieving a 4.9% improvement in mAP@0.5, with only 1.44M parameters. It also reaches an inference speed of 171 FPS on an NVIDIA RTX4090 GPU, striking an optimal balance between accuracy and efficiency, making it a viable solution for real-time UAV applications.