航空学报 > 2025, Vol. 46 Issue (7): 331003-331003   doi: 10.7527/S1000-6893.2024.31003

基于地貌类别信息指导的SAR图像仿真方法

孟令捷1, 李红光2(), 李新军2   

  1. 1.北京航空航天大学 电子信息工程学院,北京 100191
    2.北京航空航天大学 无人系统研究院,北京 100191
  • 收稿日期:2024-07-29 修回日期:2024-09-07 接受日期:2024-10-09 出版日期:2024-11-21 发布日期:2024-10-29
  • 通讯作者: 李红光 E-mail:lihongguang@buaa.edu.cn
  • 基金资助:
    国家自然科学基金(62076019);国家重点研发计划(2022YFB3904303)

SAR image simulation method guided by geomorphic category information

Lingjie MENG1, Hongguang LI2(), Xinjun LI2   

  1. 1.School of Electronic Information Engineering,Beihang University,Beijing 100191,China
    2.Institute of Unmanned System,Beihang University,Beijing 100191,China
  • Received:2024-07-29 Revised:2024-09-07 Accepted:2024-10-09 Online:2024-11-21 Published:2024-10-29
  • Contact: Hongguang LI E-mail:lihongguang@buaa.edu.cn
  • Supported by:
    National Natural Science Foundation of China(62076019);National Key Research and Development Program of China(2022YFB3904303)

摘要:

深度学习SAR图像仿真方法一般没有考虑SAR图像不同地貌类别特征差异,导致仿真图像地貌区分失真。针对这一情况,提出一种地貌类别信息指导的可见光到SAR图像转换算法。算法设计了地貌类别提取分支,使用注意力机制,从多个维度采集地貌类别信息,指导SAR图像仿真。设计了图像内容提取分支,使用对比学习,增强网络对可见光和SAR图像共有的内容信息的特征提取能力。设计了图像生成模块,在地貌类别信息的指导下,将内容信息转化为SAR图像,使生成的SAR图像具有对应地貌类别的特征,并使用路径正则化细分可见光到SAR图像的完整转换过程,降低实现难度。建立了具有多种不同地貌的可见光和SAR图像配对数据集,通过实验对比6类评价指标,所提算法较其他代表性算法均表现出较好性能,其中结构相似度至少提升了9.24%。同时,仿真SAR图像的视觉效果真实度更高,能够有效保留地貌类别特征。

关键词: 合成孔径雷达, 深度学习, 多类别地貌, 注意力机制, 对比学习, 路径正则化

Abstract:

The current deep learning SAR image simulation methods generally do not consider the feature differences of different geomorphic categories in SAR images, resulting in distortion of geomorphic differentiation in simulated images. To address this issue, this paper proposes a visible-to-SAR image translation algorithm guided by geomorphic category information. A topographic category extraction branch is designed, and the attention mechanism is used to collect topographic category information from multiple dimensions to guide SAR image simulation. Image content extraction branches are designed, and contrast learning is used to enhance the feature extraction capability of the network for common content information of visible light and SAR images. An image generation module is designed to convert content information into SAR images under the guidance of geomorphic category information, so that the generated SAR images have the features corresponding to geomorphic categories, and path regularization is used to subdivide the complete translation process from visible light to SAR images to reduce the difficulty of implementation. A pair dataset of visible light and SAR images with different terrains is established. Experimental comparison of 6 evaluation indexes shows that the proposed algorithm has better performance than other representative algorithms, with the structural similarity being improved by at least 9.24%. In addition, the simulated SAR image shows a higher degree of realism in the visual effect, and can effectively retain the features of landform categories.

Key words: SAR, deep learning, multi-category terrain, attention mechanism, contrastive learning, path regularization

中图分类号: