SLBAF-Net: Super-Lightweight Bimodal Adaptive Fusion Network for UAV Detection in Low Recognition Environment

Unmanned aerial vehicle (UAV) detection has significant research value in the field of military and civilian applications. However, the traditional object detection algorithms commonly lack satisfying accuracy and robustness due to the intense illumination changes and extremely small size of UAVs on the remote sensor images with the sky background. This paper proposes a super-lightweight bimodal network SLBAF-Net with the adaptive fusion of visible light and infrared images for UAV detection under complex illumination and weather conditions. To handle complex illumination environments and meeting the low computing requirements of airborne computers, a super-lightweight bimodal UAV detection network inspired by YOLO's network structure is developed. In order to fuse bimodal features more effectively, the bimodal adaptive fusion module (BAFM) is proposed to perform an adaptive fusion of visible and infrared feature maps for the purpose of improving detection robustness in complex environments. To verify the superiority of our method, we build a complex dual-modal UAV dataset and conduct comprehensive comparison experiments with various state-of-art object detection networks. The experimental results show that the proposed SLBAF-Net outperforms other algorithms in terms of detection performance and robustness in harsh environments, with a precision rate of 0.909 and a recall rate of 0.912. Moreover, the SLBAF-Net can meet the real-time requirements of airborne computers, and the network size is only 5.6 MB.