.png){kind=icon}
.png){kind=icon}
Return
MHamba: Morphological State Space Module and heterogeneous feature learning for ship detection
.png){kind=icon}
PRE
AI
W
S
G
M
Y
Z
M
Z
H
DOI：10.1016/j.displa.2026.103440.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Ship detection from remote sensing imagery faces two critical challenges: inadequate capture of elongated ship morphologies in complex nearshore environments and severe scale imbalance when small and large vessels coexist. Existing CNN-ViT hybrid methods struggle with quadratic complexity in modeling longrange dependencies and rely on fixed multi-scale aggregation strategies that lack adaptability. To address these limitations, we propose MHamba, a Marine ship Structure-Aware Vision Mamba Network that achieves rotation-robust morphological feature learning and heterogeneous multi-scale fusion. Our key contribution is the Morphological State Space Module (MSSM), which employs Zero-in Convolution for efficient local shape refinement and introduces directional consistency regularization over structure-aware scanning routes, transforming discrete multi-directional perception into continuous manifold representations. For scale-adaptive fusion, the Heterogeneous Feature Fusion Module (HFFM) dynamically reweights feature contributions through heterogeneous kernel selection and consistency-aware learnable fusion. Extensive experiments demonstrate that MHamba achieves state-of-the-art performance on HRSC2016 (91.1% AP50, 72.8% AP75, 61.8% AP50 & ratio;95) while maintaining better generalization on Seaships7000 and ShipRSImageNet benchmarks.
Keywords:
Remote sensing
State space model
Mamba
Ship detection
Journal
IF:
3.4
Papers:
2.1K
Citations:
3.2K
Organization
Cited Papers