Mamba-OrthoNet: unified global–local feature fusion and angular encoding for accurate ship orientation detection

Abstract

Accurate ship orientation detection in high-resolution remote-sensing imagery is vital for maritime monitoring, traffic management, and search-and-rescue. Yet port scenes are crowded and ships vary in scale: global context separates adjacent ships, while fine local cues identify bow and stern, so either alone is insufficient. Angle regression also suffers from periodicity and the $$0^\circ /360^\circ $$ 0 ∘ / 360 ∘ discontinuity, causing training instability. We propose Mamba-OrthoNet, integrating a CNN branch with a Mamba state-space branch and gradually fusing them from representation alignment to residual refinement. Orthogonal Feature Fusion (OFF) maps both streams into a shared space and decomposes CNN features to extract detail, reducing redundancy for compatible fusion. Progressive residual enhancement (PRE) aggregates residuals layer by layer to iteratively refine fused features. We adopt FPBiFusion and introduce ConvNorm with CSPRep for stronger multi-scale representations. For angle modeling, dual-granularity ring encoding (DGRE) combines coarse cyclic bins with fine smoothed sub-bins for stable full-range 0 $$^\circ $$ ∘ –360 $$^\circ $$ ∘ prediction. On DOTA-ORShip, Mamba-OrthoNet achieves 96.3% mAP $$_{50}$$ 50 and 94.8% orientation accuracy, with ablation studies confirming the contributions of OFF, PRE, and DGRE under this benchmark setting.