Aerial small object detection via dynamic convolution and hierarchical attention fusion for UAV imagery

Small object detection in unmanned aerial vehicle (UAV) aerial imagery is challenged by small object size, sparse feature information, and complex backgrounds. This paper presents AeroVision-DET, an aerial small object detection algorithm based on dynamic convolution and hierarchical attention fusion. We construct the Adaptive Receptive Field Network (ARFNet) based on the Multi-Scale Adaptive Feature Module (MSAFM) as the backbone, which achieves adaptive receptive field adjustment through dual-path feature enhancement blocks and multi-shape adaptive convolution. We design the Semantic-Spatial Fusion Module (SSFM), which adopts a multi-scale contextual attention mechanism to achieve semantically consistent and spatially sensitive feature fusion. We further propose the Efficient Feature Encoding Layer (EFEL), which integrates polarized linear attention and a frequency-modulated feed-forward network to model long-range dependencies under linear complexity. Experiments on VisDrone2019 show that AeroVision-DET achieves 24.2% AP, a 3.4% improvement over the RT-DETR-r18 baseline, with a 3.5% gain on small-object AP and a 27.0% parameter reduction while computational complexity remains essentially unchanged. With only 14.57 M parameters and 56.9 GFLOPs, AeroVision-DET delivers 69.5 FPS real-time inference and is well-suited for deployment on resource-constrained edge devices such as embedded UAV platforms, while maintaining detection accuracy competitive with or superior to state-of-the-art (SOTA) models.