RGE-YOLO: A lightweight and robust detector for infraredship detection via global-local feature alignment andempirical attention

Abstract Infrared imaging is essential for maritime surveillance due to its all-weather capabilities.However, ship detection remains challenging due to sparse target representations and complex background clutter, particularly on resource-constrained edge devices. To address these limitations, this study proposes RGE-YOLO, an ultra-lightweight and robust detector. We introduce RSCHGNet as the backbone, which utilizes structural re-parameterization to balance training complexity with inference efficiency. To mitigate cross-scale feature mis-
alignment, we construct a Global-Local Aligned Feature Pyramid Network (GLA-FPN) incorporating a high-resolution P2 layer and deformable alignment mechanisms. Furthermore,a thermal-aware ECRA detection head is designed to suppress sea clutter by focusing on in-trinsic ship thermal signatures. Experimental results on infrared datasets demonstrate that RGE-YOLO outperforms the YOLO11n baseline, improving mAP50 by 3.8% and mAP50:95 by 2.1%, while reducing parameter counts by 11.5% (to 2.3 M). This model achieves an optimal trade-off between precision and efficiency, offering a viable solution for real-time maritime monitoring.