中国空间科学技术

中国空间科学技术

• 论文 •    下一篇

应用密集预测点生成的天基遥感舰船旋转检测方法

史嘉伟1,2,*，宋鹏飞1,2，高宏霞1，刘晓磊1   

  1. 1.中国空间技术研究院遥感卫星总体部，北京100094
    2.极限环境光电动态测试技术与仪器全国重点实验室，太原030051
  • 收稿日期:2026-01-30 修回日期:2026-04-13 录用日期:2026-05-18 发布日期:2026-09-23 出版日期:2026-09-23

Space-based remote sensing rotated ship detection method using dense prediction point generation

SHI Jiawei1,2,*，SONG Pengfei1,2，GAO Hongxia1，LIU Xiaolei1   

  1. 1.Institute of Remote Sensing Satellite, China Academy of Space Technology，Beijing 100094,China
    2.State Key Laboratory of Extreme Environment Optoelectronic Dynamic Measurement Technology and Instrument，Taiyuan 030051，China
  • Received:2026-01-30 Revision received:2026-04-13 Accepted:2026-05-18 Online:2026-09-23 Published:2026-09-23

PDF (PC)

3

摘要： 海量天基遥感影像中舰船目标具有朝向任意、长宽比悬殊、密集排布等复杂特征，传统旋转检测方式面临严峻的角度边界与顶点排序歧义问题，导致检测精度不足。为此，研究了一种应用密集预测点生成的天基遥感舰船旋转检测方法，通过网络预测一组与舰船位置相关的密集预测点生成精确定位信息，从而替代传统角度回归或四顶点预测方式。为进一步提升检测精度，设计边缘区域采样策略，从舰船关键结构（船艏、船身等）均匀采样225个密集点，提供高质量训练监督；提出加权-残差特征融合模块，引入跳层连接与可学习权重，实现多尺度特征高效融合；此外，采用优化的加权倒角损失函数结合分阶段定位损失策略，实现从粗到精的密集点与检测框定位。利用公开数据集HRSC2016对提出的方法进行验证，实验结果表明该方法平均精度达到92.4%，相较对比方法提升了至少0.9%，验证了其在大规模遥感影像中的检测能力，为旋转舰船检测的自动化、精准化与工程化应用提供技术基础，可将其集成应用于遥感影像自动化处理流水线。

关键词: 天基遥感, 舰船检测, 密集预测, 特征融合, 边缘采样

Abstract: Ships in massive space-based remote sensing images exhibit complex characteristics such as arbitrary orientation, large aspect ratio variation, and dense arrangement, which cause serious angle boundary and vertex sorting ambiguity in traditional rotated detection methods, leading to insufficient detection accuracy. To address this, this study developed a rotated detection method for ship targets using dense prediction point generation, which used a neural network to predict a set of dense points related to ship positions, generating precise localization and replacing traditional angle regression or four-vertex prediction. In order to further improve detection accuracy, this study designed an edge region sampling strategy to uniformly sample 225 dense points from key ship structures (bow, hull, etc.) for high-quality training supervision, proposed a weighted-residual feature fusion module with skip connections and learnable weights for efficient multi-scale feature fusion, and adopted an optimized Weighted Chamfer Loss function combined with a staged localization loss strategy for coarse-to-fine point and bounding box localization. Experimental validation on the public dataset HRSC2016 shows that the proposed method achieves 92.4% mAP, outperforming comparison methods by at least 0.9%, demonstrating its capability in detecting ships within large-scale remote sensing imagery. This approach provides a technical foundation for the automation, precision, and engineering application of rotated ship detection and can be integrated into remote sensing image processing pipelines.

Key words: space-based remote sensing, ship detection, dense prediction, feature fusion, edge sampling