ARANet: A Unified Representation Alignment Framework for SAR Ship Instance Segmentation

Deep learning has become the dominant paradigm for ship instance segmentation in synthetic aperture radar (SAR) image. However, reliable segmentation in complex maritime environments remains constrained by three representational misalignments: inconsistency between semantic and structural cues, imbalance in global-local contextual modeling, and limited adaptability to cross-dataset scale variations. To address these challenges, we propose the Adaptive Representation Alignment Network (ARANet), a unified representation alignment framework that systematically reconciles these misalignments through three complementary alignment mechanisms. Specifically, ARANet enhances semantic-structural consistency via an edge-guided boundary refinement mechanism, balances global and local contextual representations through a context-aware modeling mechanism, and improves scale adaptability using a depth-adaptive feature alignment mechanism. These mechanisms are instantiated as the Edge-Guided Boundary Refinement Module (EGBRM), the Context-Aware Module (CAM), and the Depth-Adaptive Feature Pyramid Network (DAFPN), respectively, and are integrated into a unified end-to-end architecture. Extensive experiments on the HRSID and PSeg-SSDD datasets demonstrate that ARANet achieves leading or competitive performance compared with CNN-based, transformer-based, and SAR-specific instance segmentation methods.