LiteUAV-Det: An Efficient Lightweight Deep Learning Framework for UAV-to-UAV Small Target Detection in Complex Aerial Scenes

Abstract

Detecting small UAV targets in air-to-air scenarios is important for aerial security, autonomous surveillance, and multi-UAV defense systems. However, this task remains challenging due to severe scale variation, motion blur, and complex aerial backgrounds. Existing detectors suffer from performance degradation on extremely small targets while maintaining high computational cost. To address these issues, this paper proposes LiteUAV-Det, a lightweight and real-time UAV-to-UAV detection framework designed to achieve an optimal balance between detection accuracy and computational efficiency. The proposed method introduces a unified lightweight feature refinement strategy based on three key components. First, a Ghost-Enhanced C2f (GEC) module is designed to improve feature representation while reducing redundant computations. Second, a GEC-CSPGhostNet backbone is constructed to enhance multi-scale feature extraction with reduced parameter complexity. Third, a GEC-PANet neck is developed to improve feature fusion using dynamic upsampling and channel-guided spatial attention, enabling better preservation of small-target information. In addition, SPP-LSKA is incorporated to capture long-range spatial dependencies and enhance contextual feature modeling for improved representation of small UAV targets. Extensive experiments on the Det-Fly and DUT-Anti-UAV datasets demonstrate that LiteUAV-Det achieves competitive detection performance compared with state-of-the-art lightweight detectors. The proposed model achieves 94.7% $mAP_{50}$ with only 6.03M parameters and 12.1G FLOPs, while achieving an inference latency of 6.5 ms (154 FPS) on a GPU node under end-to-end evaluation. The results show that LiteUAV-Det provides an effective solution for real-time UAV-to-UAV small object detection in complex aerial environments.