Oriented SAR ship detection with scattering-inspired cross-scale fusion and localization quality estimation

Oriented ship detection in complex nearshore and harbor SAR imagery is challenged by nonhomogeneous sea clutter and land-based strong scatterers, attenuation of weak ship structures during downsampling, and mismatch between classification confidence and oriented-box localization quality. We propose PSF-LQE, a scattering-inspired intensity-domain network with four functionally complementary modules. SCS extracts absolute scattering residuals against a locally smoothed background and reconstructs features via confidence gating, suppressing sea clutter before false responses propagate to deeper layers. PTE jointly encodes local residuals, channel-wise intensity contrast, and directional responses, preserving the scattering centers, contours, and heading-dependent structures of small weak ships throughout hierarchical encoding. A bidirectional neck built on MPCA and MPCA-Down generates channel gates from concatenated adjacent-scale descriptors before upsampling and downsampling fusion, calibrating channel reliability so that deep clutter semantics and shallow noise do not contaminate each other. LQE derives a localization-quality bias from four-side discrete boundary distributions and uses it as an additive re-ranking cue on the classification scores, so that candidates whose boundary distributions are more concentrated tend to rank higher, without claiming that the scores themselves are calibrated. The network requires no polarimetric channels, scattering matrices, or decomposition features and has a compact 3.8 million parameters. Experimental results show that PSF-LQE achieves 93.8% $$mAP_{50}$$ on HRSID, corresponding to absolute gains of 19.6 percentage points over the evaluated baseline. On SRSDD, it achieves 73.6%, corresponding to gains of 14.9 percentage points. Cumulative ablations and qualitative response visualizations are consistent with the intended roles of the modules.