Enhanced SAR Ship Detection Using Multiscale Neighborhood Correlation and Self-Supervised Constraint

Ship detection in synthetic aperture radar (SAR) imagery is a cornerstone of maritime surveillance, leveraging active microwave backscattering for all-weather target acquisition. However, the coherent imaging mechanism inherently introduces speckle noise, scattering ambiguity, and sidelobe effects, causing general-purpose deep learning models to suffer from false alarms and missed detections in complex nearshore environments. To address the critical lack of domain-specific modeling in current frameworks, this article proposes an enhanced SAR ship detection approach driven by electromagnetic physical priors. In particular, acknowledging that vessels are represented as discrete clusters of dominant scattering centers, center difference convolution (CDC) is incorporated to emphasize local structural variations, thereby suppressing sidelobe-induced interference while preserving discriminative ship boundaries. Furthermore, a multiscale neighborhood correlation analysis strategy is developed to exploit the statistical divergence of neighborhood backscattering responses between ship targets and surrounding sea clutter while enhancing robustness against ambiguous local scattering patterns through multiscale contextual modeling. To mitigate the signal dilution effect during successive downsampling, a self-supervised constraint mechanism is introduced to maintain cross-layer feature consistency and compensate for the energy loss of dim small targets in deeper layers. Extensive evaluations on benchmark datasets demonstrate that the proposed method significantly improves detection performance in complex scenes, underscoring the importance of incorporating SAR-specific scattering characteristics into deep learning architectures for robust maritime target detection.