CFAR-DP-FW: A CFAR-Guided Dual-Polarization Fusion Framework for Large-Scene SAR Ship Detection.

Effective ship detection in synthetic aperture radar (SAR) imagery is crucial for maritime safety and surveillance. Despite the advancements in deep learning for SAR ship detection, significant challenges remain, particularly in large scenes. These challenges are twofold: the detection of extremely small ships is often hindered by inadequate feature extraction, and the presence of inshore ships is obscured by pronounced land-based interference, both of which lead to reduced detection accuracy. To address these issues, we present a novel deep learning framework that integrates constant false alarm rate (CFAR) processing with dual-polarization data, termed the CFAR-guided dual-polarization fusion framework (CFAR-DP-FW). The integration is designed to enhance the detection sensitivity for small-scale maritime targets by utilizing dual-polarization's rich feature representation, and CFAR's strength in suppressing background noise, highlighting potential targets. The proposed CFAR-DP-FW consists of three core components: the CFAR dual-polarization detector (CFAR-DPD) provides initial target indication; the CFAR field generator (CFAR-FG) constructs a probabilistic ship presence map, reducing reliance on CFAR's hard thresholds; and the CFAR guidance dual-polarization network (CFAR-GDP-Net) incorporates a novel feature extractor and enhancement module, tailored to amplify relevant features, and suppress noise. This strategic fusion within our framework markedly improves the detection of small and inshore ships. Evaluated on the Large-Scale SAR Ship Detection Dataset-v1.0 (LS-SSDD-v1.0), our framework demonstrates superior performance, surpassing twenty state-of-the-art models. It achieves a 3.28% increase in mean average precision for inshore ships over the next best-performing model, validating its efficacy in tackling the intricate challenges of large-scale SAR ship detection.