Wavelet-Decoupled Representation with Rotation-Domain Prior Gating for PolSAR Ship Detection

Polarimetric synthetic aperture radar (PolSAR) provides multichannel scattering responses, offering richer cues for ship detection. However, existing data-driven methods primarily focus on spatial-domain feature modeling and underutilize physically motivated priors embedded in polarimetric scattering mechanisms. To address this, we propose a wavelet-based rotation-prior gated network (WRPGN). First, via equivalent polarization-basis rotation, we construct a rotation-domain polarimetric prior map. This map exploits polarimetric information to enhance signal-to-noise ratio, yielding reliable spatial confidence for subsequent gating. Second, we introduce a high-frequency prior-gated (HFPG) block into the feature pyramid. It decomposes features into low-frequency structural components and high-frequency details, and maps the rotation-domain prior to learnable gating weights that selectively modulate the details. These modulated details are then residually fused with the structural components to reconstruct multiscale representations that are both physically interpretable and highly discriminative. Experiments on the FPSD dataset demonstrate the effectiveness and competitive performance of the proposed method.