Feature Pyramid Optimization-Based Small Object Detection in UAV Imagery

Abstract

Unmanned Aerial Vehicles (UAVs) are increasingly used in key applications such as surveillance, search and rescue. However, due to the small scale of objects, cluttered backgrounds, and inherent information loss, accurate object detection from a high-altitude perspective is still a major challenge. Although two-stage detectors can provide high accuracy, their high computational costs are not suitable for real-time edge deployment. In contrast, the most advanced single-stage detectors, such as YOLOv10, often fail to capture small objects because the deep pyramid structure loses key spatial details in the downsampling process. In order to overcome these limitations, this paper proposes a systematic optimization framework for small object detection in high-altitude imagery. We first introduce FemtoDet-P2, which combines the MambaOut backbone network with a high-resolution P2 detection head. As a high-precision baseline, this architecture verifies that strengthening early feature extraction is crucial to solving the problem of feature vanishing, even if structural redundancy is introduced. Based on this, we propose LSCNet to balance the trade-off between detection accuracy and computational efficiency. LSCNet eliminates the redundancy in the baseline model and enhances multi-level feature fusion by constructing a shallow feature cascade and introducing a lightweight attention mechanism. This lightweight design ensures that advanced deep learning models can be efficiently deployed on resource-constrained UAV platforms without compromising their ability to identify small objects in complex environments.