Reinforcing Out-of-Distribution SAR Ship Detection: A Guidance-Constraint Hybrid Approach

Out-of-distribution (OOD) generalization remains a critical challenge in synthetic aperture radar (SAR) ship detection, where training and test data often originate from different distributions. Conventional approaches, such as domain generalization and domain adaptation, rely heavily on data diversity, limiting their efficacy in real-world scenarios with limited training data. To address this, we propose a guidance-constraint hybrid framework that integrates structured knowledge into data-driven detection pipeline. Candidate regions are first identified using a classical local constant false alarm rate detector, serving as guidance. Confidence estimation is then refined via second-order Gaussian filtration, acting as a constraint to suppress noise and enhance spatial coherence. The data-driven branch employs Enhanced ShuffleNet v2 (ESv2), a lightweight architecture designed to improve semantic correlation and attention to relevant regions. Experimental results demonstrate that our hybrid model achieves robust performance under severe distributional shifts, with an absolute improvement of approximately 20% in mAP@0.5 on OOD scenarios compared to conventional data-driven methods. These findings highlight the potential of informed machine learning to complement statistical approaches and advance strict OOD generalization in SAR ship detection.