首页 >

可见光与SAR多源遥感图像关联学习算法-信息融合大会增刊

王子玲1,熊振宇2,顾祥岐1   

  1. 1. 海军航空大学
    2. 海军航空大学信息融合研究所
  • 收稿日期:2022-04-06 修回日期:2022-05-06 出版日期:2022-05-19 发布日期:2022-05-19
  • 通讯作者: 王子玲
  • 基金资助:
    国家自然科学基金

Correlation learning algorithm of visible light and SAR cross modal remote sensing images

1,Zhen-Yu XIONG1,   

  • Received:2022-04-06 Revised:2022-05-06 Online:2022-05-19 Published:2022-05-19

摘要: 针对可见光和SAR图像由于成像机理不同导致图像内容差异大,深度特征难对齐,关联速度慢,提出一种深度多源哈希网络模型实现SAR图像和可见光图像间关联。首先,针对SAR与光学遥感图像颜色信息差异大,提出图像变换机制,将光学图像转换生成四种不同类型的光谱图像输入到网络中,打乱颜色通道,让网络更加关注于图像的纹理和轮廓信息,而对颜色信息不敏感;其次,针对SAR图像噪声大,同一场景下两种模态图像内容异构,提出图像对训练策略,减小多源图像间的特征差异;然后,针对关联效率低,存储消耗大,提出三元组哈希损失函数,提升模型的关联准确率,降低关联时间;最后,构建了一个SAR与光学双模态遥感数据集SODMRSID,填补了可见光与SAR多源遥感图像关联数据的空白,同时实验部分验证了数据集的实用性和提出算法的有效性。

关键词: 多源异构, 遥感图像, 关联学习, 深度神经网络

Abstract: Due to the different imaging mechanism of visible and SAR images, the content of images is different, the depth features are difficult to align and the correlation speed is slow. A depth cross modal hash network model is proposed to realize the cross modal correlation between SAR images and optical images. Firstly, aiming at the great difference of color information between SAR and optical remote sensing images, an image transformation mechanism is proposed. Four different types of spectral images are generated from optical image conversion and input into the network, which disrupts the color channel, so that the network pays more attention to the texture and contour information of the image, but is not sensitive to the color information; secondly, aiming at the high noise of SAR image, two modal images in the same scene are generated. For example, the content is heterogeneous, image pair training strategy is proposed to reduce the feature difference between cross-modal images; then, aiming at the low correlation efficiency and high storage consumption, a triple hash loss function is proposed to improve the association accuracy of the model and reduce the association time. Finally, a SAR and optical dual-mode remote sensing data set is constructed to make up for the lack of data in this field. The experimental part verifies the practicability of the data set and the effectiveness of the proposed algorithm.

Key words: Multi-source heterogeneous, remote sensing image, association learning, deep neural network

中图分类号: