航空学报 > 2023, Vol. 44 Issue (22): 627476-627476   doi: 10.7527/S1000-6893.2022.27476

基于属性引导的多源遥感舰船目标可解释融合关联网络

熊振宇(), 崔亚奇, 董凯, 李孟洋, 熊伟   

  1. 海军航空大学 信息融合研究所,烟台  264001
  • 收稿日期:2022-05-20 修回日期:2022-06-21 接受日期:2022-07-29 出版日期:2023-11-25 发布日期:2022-08-03
  • 通讯作者: 熊振宇 E-mail:x_zhen_yu@163.com
  • 基金资助:
    国家青年科学基金(62001499);国家自然科学基金(61790554)

Interpretable fusion association network for multi-source remote sensing ship target based on attribute guidance

Zhenyu XIONG(), Yaqi CUI, Kai DONG, Mengyang LI, Wei XIONG   

  1. Institute of Information Fusion,Naval Aviation University,Yantai  264001,China
  • Received:2022-05-20 Revised:2022-06-21 Accepted:2022-07-29 Online:2023-11-25 Published:2022-08-03
  • Contact: Zhenyu XIONG E-mail:x_zhen_yu@163.com
  • Supported by:
    National Science Fund for Young Scholars(62001499);National Natural Science Foundation of China(61790554)

摘要:

多源遥感舰船目标关联作为前期大范围预警探测的重要手段为海上态势研判提供重要情报支撑,现有关联算法面临关联结果可解释性差,异构特征难度量,多源目标关联精度低等问题。提出了一种基于属性引导的可解释融合网络用于解决多源遥感舰船目标的关联问题。首先,提出全局关联模块,利用跨模态度量损失函数将图像特征映射到共同空间中度量,用于解决多源图像内容差异大,特征难对齐问题。然后,提出包含多头注意力模型和属性监督函数的可解释模块,提升关联精度并输出可解释的关联结果。其中多头注意力模型让网络关注到舰船目标显著性区域,属性监督函数引导模型关注舰船图像中判别性属性特征,利用属性特征帮助网络解释输出关联结果的决策依据,并以量化的形式可视化属性特征对关联结果的贡献度。最后,利用知识蒸馏的思想减小全局关联模块和可解释模块输出特征距离的差异,使得网络实现精准关联并提供可解释的关联结果。在实验部分,构建了首个多源遥感舰船目标数据集,在该数据集上的测试结果显示本文算法不仅在关联精度上优于现有算法,同时能够为关联过程提供清晰和直观的可视化关联结果。

关键词: 多源目标, 遥感图像, 关联学习, 注意力模型, 可解释特征

Abstract:

Multi-source remote sensing ship target correlation, as an important means for early-stage large-scale early warning and detection, provides important information support for maritime situation research and judgment. Existing association algorithms face the problems of poor interpretability of association results, difficulty in measuring heterogeneous features and low accuracy of multi-source target association. In this paper, an interpretable fusion network based on attribute guidance is proposed to solve the problem of ship target association in multi-source remote sensing. Firstly, to solve the problem of large difference in multi-source image content and difficulty in feature alignment, a global association module is proposed, which uses the cross modal measurement loss function to map image features into the common space. Then, an interpretable module including the multi head attention model and the attribute supervision function is proposed to improve the correlation accuracy and output interpretable correlation results. The multi-head attention model makes the network pay attention to the salient region of ship targets, and the attribute supervision function enables the model to pay attention to the discriminant attribute features in ship images. Finally, the idea of knowledge distillation is used to reduce the difference between the output feature distance of the global correlation module and the interpretable module, so that the network can realize accurate correlation and provide interpretable correlation results. In the experimental part, this paper constructs the first multi-source remote sensing ship target data set. The test results on this data set show that this algorithm is not only better than the existing algorithms in correlation accuracy, but also can provide clear and intuitive visual correlation results for the correlation process.

Key words: multi-source ship target, remote sensing image, association learning, attention model, interpretable feature learning

中图分类号: