FGOM-RTDETR: Far-Shore-Guided Object-Focusing Multiscale Network With Real-Time Detection Transformer for Infrared Ship Target Detection

Infrared ship detection plays a critical role in both civilian and military applications, including tracking, collision avoidance, and maritime security. However, challenges such as low image resolution and complex backgrounds hinder detection accuracy. This study proposes a novel detection algorithm, Far-shore Guided Object-focusing Multiscale Network with Real-Time Detection Transformer (FGOM-RTDETR), which integrates multi-scale local and global features. Built upon the RT-DETR framework, our method introduces a Feature Grouping Module (FGOM) to enhance multi-scale representation. FGOM consists of three key components: the Feature Reparameterization Module (FRep), the Coordinate Attention Golden Feature Pyramid Network (CAGoldFPN), and the Multi-Scale Stacked Network (MuSSNet). The FRep module addresses the loss of channel information caused by the small size of thermal infrared ship targets and the complexity of background features. The CAGoldFPN module improves multi-scale feature fusion, while MuSSNet mitigates issues of high target similarity and the susceptibility of small targets to being overlooked. Experimental results show that, compared with the baseline RT-DETR model, FGOM-RTDETR achieves notable performance gains: precision improves from 0.921 to 0.942, mAP50 rises from 0.938 to 0.958, and recall increases from 0.923 to 0.934. These results demonstrate that FGOM-RTDETR delivers superior detection performance for infrared ship targets.