CDF-Net: Cross-Scale Direction-Aware Focusing Network for Robust SAR Ship Detection with Horizontal Bounding Boxes

Ship detection with horizontal bounding box in synthetic aperture radar (SAR) images remains challenging due to multi-scale imbalance, arbitrary ship orientations, severe sea clutter, and high false alarm rates in dense inshore scenes. We propose a Cross-scale Direction-aware Focusing Network (CDF-Net) with three components. The Hierarchical Prior Module (HPM) uses cascaded branches with increasing kernel sizes to extract multi-scale features before the backbone. The Multi-scale Directional Strip Convolution (MDSC) combines horizontal and vertical depthwise strip convolutions at three kernel sizes, using learned weights to fuse their outputs. The Iterative Focus Cascade Encoder (IFCE) aligns features from three pyramid levels, fuses them at the intermediate resolution, and redistributes them to the finer and coarser levels in two successive stages. CDF-Net achieves AP values of 0.742, 0.718, and 0.434 on HRSID, SSDD, and IPSD, respectively, the highest among the methods compared under the reported protocol. On HRSID, the gain over the D-FINE baseline is 4.8 percentage points.