LiteUAV-Det: A lightweight network for robust small-object detection in complex aerial scenes

Recent advances in deep learning have enabled strong performance in object detection on mobile platforms, such as unmanned aerial vehicles (UAVs). However, current state-of-the-art (SOTA) models struggle to detect small, flat, and fast-moving blurred objects in complex UAV environments characterized by high altitude and low light. The primary challenge arises from the pixel-level resemblance between objects and their surroundings, which complicates detection. To overcome these limitations, we propose LiteUAV-Det, a robust lightweight model designed to enhance the accuracy and efficacy of object identification in challenging UAV environments. We introduce deep selective feature fusion (DSFF) and enhanced pyramid pooling fast (EPPF) modules to reduce model complexity while preserving superior performance. The DSFF module reduces model complexity by selectively fusing multi-scale features with minimal computational overhead, while the EPPF module enhances detection performance by efficiently aggregating rich contextual information across scales. Additionally, our proposed framework integrates an adaptive brightness enhancement module and a deblurring subnet (DBSnet) to enhance the detector’s ability to recognize visually degraded targets in challenging UAV scenarios. To support this task, we introduce UAV-SOD, a newly constructed unmanned aerial vehicle (UAV)-assisted dataset for small-object detection (SOD) under varying illumination and motion-blur conditions. Our LiteUAV-Det demonstrates the smallest parameter count (6.38M) and inference latency among SOTA methods. It achieves the smallest inference time of 2.4 ms (214 FPS) on the UAV-SOD dataset, 2.3 ms (300 FPS) on the VisDrone2019 dataset, and 4.8 ms (169 FPS) on the DOTA-v1.5 dataset. Experimental results show that the proposed model achieves competitive performance compared to other SOTA methods in complex environments, obtaining scores of 34.5%, 33.8%, and 35.3% on the UAV-SOD, VisDrone2019, and DOTA-v1.5 datasets, respectively. Additionally, the proposed model achieves improvements of 3.6% on the UAV-SOD dataset, 4.2% on the VisDrone2019 dataset, and 2.9% on the DOTA-v1.5 dataset compared with similar lightweight models, such as YOLOv9s (7.17M parameters).