PGNet: Progressive Feature Alignment and Gaussian Geometric-Consistency Assignment for Ship Detection in Complex...

Deep learning has advanced remote sensing object detection, with ship detection drawing attention due to its importance in maritime traffic, defense, and marine resource management. However, real-world scenarios often involve cloud and fog occlusion, sea clutter, and uneven illumination, which degrade feature representation. Moreover, existing sample assignment strategies fail to account for the high-angle sensitivity of slender ships, causing many potentially high-quality proposals to be assigned as negative samples under slight rotational deviations. To address these issues, we propose PGNet, a robust ship detection framework that jointly optimizes feature representation and sample assignment. Specifically, a progressive aligned pre-fusion feature pyramid network is designed to alleviate spatial misalignment and semantic inconsistency via adaptive multi-scale spatial alignment and progressive feature interaction. Meanwhile, a Gaussian geometric consistency assignment strategy is introduced, which jointly models overlap and center alignment to alleviate the high-angle sensitivity problem for slender ships. Experiments on the HRSC2016, SCCOS, and FGSRCS datasets demonstrate that PGNet consistently outperforms existing methods, achieving a mean average precision (mAP) of 88.8% (+1.7%) on FGSRCS, 87.52% mAP and 91.84% recall (+0.5% and +1.43%) on SCCOS, and 70.39% mAP ${}_{50:95}$ (+1.22%) on HRSC2016, indicating robustness and generalization in complex maritime environments.