计算机与现代化 ›› 2024, Vol. 0 ›› Issue (03): 29-33.doi: 10.3969/j.issn.1006-2475.2024.03.005

• 人工智能 • 上一篇    下一篇

基于多任务学习的近岸舰船检测方法

  

  1. (1.中国科学院空天信息创新研究院,北京 100190; 2.中国科学院大学电子电气与通信工程学院,北京 100049;
    3.中国科学院网络信息体系技术重点实验室,北京 100190)
  • 出版日期:2024-03-28 发布日期:2024-04-28
  • 作者简介:刘馨嫔(1999—),女,湖南邵阳人,硕士研究生,研究方向:信号与信息处理,E-mail: xinpinliu@126.com; 通信作者:王洪(1975—),男,副研究员,硕士生导师,博士,研究方向:地理空间大数据挖掘,E-mail: wanghong@aircas.ac.cn; 赵良瑾(1992—),男,助理研究员,研究方向:遥感图像检测识别,E-mail: zhaolj004896@aircas.ac.cn。
  • 基金资助:
    国家自然科学基金资助项目(62201550)

Inshore Warship Detection Method Based on Multi-task Learning

  1. (1. Aerospace Information Research Institute, Chinese Academy of Science, Beijing 100190, China;
    2. School of Electronic, Electrical and Communication Engineering, University of Chinese Academy of Sciences, Beijing 100049, China;
    3. Key Laboratory of Network Information System Technology, Chinese Academy of Sciences, Beijing 100190, China)
  • Online:2024-03-28 Published:2024-04-28

摘要:
摘要:在遥感光学图像近岸舰船目标检测任务中,针对近岸复杂场景中存在形状近似目标的虚警问题,提出一种基于多任务学习的近岸舰船目标检测方法。该方法通过构建海陆分割任务与舰船检测任务并行双路框架,将传统的任务串行处理流程优化为并行处理方式,设计联合损失函数进行双路优化训练约束,提升模型训练的稳定性,通过双分支融合模块剔除陆地掩膜中的检测结果,实现陆地虚警滤除。采用谷歌地球遥感图像制作的数据集进行实验,将本文提出的方法与单任务检测算法YOLOv5相比,mAP提升了4.4个百分点,虚警率降低了3.4个百分点。实验结果表明本文算法对陆地虚警抑制有效。

关键词: 关键词:舰船检测, 海陆分割, 多任务学习, 损失函数

Abstract: Abstract: In the task of inshore warship detection in remote sensing optical images, this paper proposes an inshore warship detection method based on multi-task learning for the false alarms problem of similar features in complex scenes. By constructing a parallel dual-branch task framework for the sea-land segmentation mission and the warship detection mission, this method optimizes the traditional task of serial processing into parallel processing mode. Secondly, we propose a joint loss constraint for dual path optimum training, which improves the stability of model training. Finally, the dataset made by Google Earth remote sensing images is used for experiments. The detection results in land mask are eliminated by the dual-branch fusion model, and the land false alarm filter is realized. Compared with the single task detection algorithm YOLOv5, the mAP of the proposed method increased by 4.4 percentage points and the false alarm rate decreased by 3.4 percentage points. The experimental results show that the proposed algorithm is effective in suppressing false alarm on land.

Key words: Key words: warship detection, sea-land segmentation, multi-task learning, loss function

中图分类号: