RGE-YOLO: a lightweight and robust detector for infrared ship detection via global–local feature alignment and empirical attention

Abstract Infrared imaging is essential for maritime surveillance due to its all-weather capabilities. However, ship detection and accurate spatial localization remains challenging due to three inherent limitations: severe resource constraints on edge deployment devices, sparse feature representations of tiny targets, and complex background sea clutter. To systematically address these challenges, this study proposes RGE-YOLO, a lightweight and robust detector . First, to overcome deployment constraints, we introduce RSCHGNet as the backbone, which utilizes structural re-parameterization to achieve an optimal balance between training feature capacity and inference efficiency. Second, to mitigate cross-scale feature misalignment and the loss of tiny targets during downsampling, we construct a global–local aligned feature pyramid network incorporating a high-resolution P2 layer and deformable alignment mechanisms. Furthermore, to effectively suppress sea clutter, an empirical-attention-aware efficient channel-region attention detection head is designed to focus the network strictly on intrinsic ship salient structural features. Experimental results on infrared datasets demonstrate that RGE-YOLO significantly outperforms the YOLO11n baseline, improving mAP 50 by 3.8% and mAP 50 : 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 highly viable solution for real-time maritime monitoring and reliable visual sensing.