AUFFNet: an attention-based fusion network with an upsampler-free feature pyramid network for object detection in UAV aerial images

The detection of small objects in unmanned aerial vehicle (UAV) images remains a significant challenge due to the constraints imposed by UAV altitude and shooting angles. Objects in such images are often small and lack distinct features, making them vulnerable to interference from lighting variations, shadows, and background noise, which complicates detection tasks. To address these challenges, we propose an attention-based fusion network with an upsampler-free feature pyramid network (AUFFNet), designed to enhance the detection of small objects in UAV aerial imagery. The proposed framework incorporates three key innovations: the additive self-attention convolutional gated fusion module, which enhances multi-dimensional feature interactions to improve small object representation while mitigating background interference; the inter-layer feature pyramid network in UAV aerial images, an upsampler-free feature fusion architecture that avoids feature loss caused by traditional upsampling and downsampling operations through cross-scale feature fusion; and the remote sensing dense small target detection loss function, which further reduces prediction errors by addressing issues related to occlusion and scale variations. Extensive experiments on the VisDrone dataset demonstrate the effectiveness of AUFFNet. Compared with the baseline YOLOv10s model, AUFFNet achieves significant improvements in performance metrics, including precision of 57.6% (+5.9%), recall of 49.8% (+11.2%), mAP50 of 48.8% (+8.7%), and mAP50–95 of 28.7% (+4.2%). Furthermore, the model achieves these results with a reduced parameter count of 5.6M (− 2.4M), highlighting its efficiency and practicality for real-world UAV applications. These advancements underscore the potential of AUFFNet as a robust solution for small object detection in complex aerial environments.