A Coordinated Feature Learning Framework for Oriented SAR Ship Detection

Abstract

Synthetic aperture radar (SAR) supports all-weather and day-and-night maritime surveillance, yet oriented ship detection remains difficult in scenes affected by sea clutter, dense port scatterers, speckle noise, large scale variation, and ambiguous ship orientation. During hierarchical feature extraction, repeated downsampling can weaken shallow spatial details and intermediate structural responses. In addition, fixed multi-scale propagation paths may not adaptively determine which feature levels and spatial regions are most useful for a specific target pyramid level. The periodicity and boundary discontinuity of angle parameters can also reduce the stability of oriented-box regression. To address these issues, we develop a coordinated oriented SAR ship detection framework composed of an Attention Residual Stage Network (ARS-Network), a Target-Aware Semantic–Spatial Adaptive Feature Pyramid Network (TSSA-FPN), and Consistency-Regularized Angle Modeling (CR-ACM). ARS-Network progressively retrieves complementary information from preceding stages through attention-weighted and gated residual connections. TSSA-FPN independently constructs each target pyramid level by combining source-stage semantic relevance, target-guided spatial selection, and gated residual injection. CR-ACM regularizes the magnitude and cross-frequency phase consistency of a periodic angle representation to improve angular continuity. Under a common experimental protocol, the framework obtains AP/AP50/AP75 values of 0.5500/0.9350/0.6110 on SSDD and 0.4910/0.9250/0.4890 on RSDD. The complete model contains 38.83 M parameters and reaches 55.9 FPS on SSDD and 40.0 FPS on RSDD. These results indicate consistent improvements over the adopted baseline while retaining practical inference efficiency.