HFC-UAVDet: Lightweight small object detection network for UAV with high-frequency perception and cross-layer semantic fusion.

Abstract

In aerial images, visible light images are susceptible to interference from complex backgrounds, while infrared images exhibit weak texture and low contrast due to their imaging mechanism. Furthermore, targets in both image types commonly share characteristics such as small scale and sparse features, making it challenging for lightweight detection models to strike a balance between accuracy and efficiency. To address this, we propose HFC-UAVDet-a lightweight small object detection network for UAVs that integrates high-frequency perception with cross-layer semantic fusion. First, the CHFB module employs local high-frequency response and soft gating to suppress pseudo-high-frequency noise while enhancing true object edges, adapting to both visible light detail extraction and infrared weak-edge enhancement requirements. Second, the dynamic position encoder HRDP models content correlation and relative displacement within self-attention, enabling stable localization in dense scenes while maintaining adaptability across both scenarios. Finally, the SGCF module achieves effective alignment between shallow-level details and deep-level semantics through channel shift mixing and bidirectional semantic interaction. Combined with the P2 branch for enhanced ultra-small target representation, it significantly improves detection capabilities for minute objects.Experiments evaluated on VisDrone2019, HIT-UAV, and TinyPerson datasets demonstrate that at 10.7MB parameters, mAP50 improves by 6.6%, 2.7%, and 2.2% over RT-DETR respectively.