【作者】 王卫东；

【导师】 刘波；

【作者基本信息】 西安电子科技大学 ， 电路与系统， 2021， 硕士

【摘要】 变化检测(Change Detection,CD)是通过对同一场景在不同时间获取的两幅或者多幅遥感图像进行对比,分析图像间的差异,从而获得所需的变化信息。合成孔径雷达(Synthetic Aperture Radar,SAR)是遥感技术中比较常见的获取遥感图像的方式,利用合成孔径原理实现高分辨的微波成像,不受天气等外界条件影响,可以全天候、全天时工作,因此SAR图像变化检测成为研究的热点。随着研究的深入,SAR图像变化检测已经有了很多优秀的方法,但是仍有一些问题存在。针对现有变化检测方法存在的问题,本文提出了基于自监督学习和深度关系网络的SAR图像变化检测,主要包括以下三个研究内容:1.针对传统变化检测方法特征学习能力较弱,以及深度学习方法依赖大量标记数据的的问题,本文提出了一种基于自监督学习和深度广义典型相关分析(Deep Generalized Canonical Correlation Analysis,DGCCA)的变化检测方法。该方法受自监督学习思想和变化检测任务特性的启发,借鉴深度广义典型相关性分析方法来设计前置任务,DGCCA的无监督多视图表示学习特性,使其具有强大的提取非线性特征以及多视图之间相关联的共享特征的能力,能够学习出有利于下游任务即SAR图像变化检测的判别性特征。通过在四组真实的SAR数据集上进行实验,验证了本方法能够显著减少标记样本的使用,并取得了较好的检测结果。2.针对传统方法在多分布跨域变化检测上的泛化性差、检测精度下降的问题,提出了一种基于细粒度类别遍历和深度关系网络的变化检测方法。该方法提出细粒度类别遍历模型,根据不同分布数据进行更加精细的类别划分,再通过类别遍历模块提取这些类别中本质的特征,从而更好地提取多分布数据的特征。同时借助深度关系网络,通过比较支持集与查询集中的样本,学习其中的非线性相似性度量,从而精确地表示复杂变化区域数据分布并提高判别性能,提高在多分布跨域变化检测的泛化性。通过在四组真实的SAR数据集上进行实验,验证了本章方法的有效性和鲁棒性。3.针对传统方法对于SAR图像边界细节信息检测较差的问题,提出了一种基于图神经网络和细粒度深度关系网络的变化检测方法。图神经网络可以处理非欧氏空间的数据结构,善于提取图的拓扑信息。引入图神经网络,可以充分提取SAR图像非变化类与变化类的边界特征,从而提高对边界细节检测的性能。该方法还将标签进行One-Hot编码,并与对应标签的样本拼接到一起,将现有数据的标签信息通过GNN进行传递到测试数据上,进一步提高在多分布跨域数据集上的泛化性。通过在四组真实的SAR数据集上进行实验,验证了本章方法检测图像边界细节的性能。更多还原

【Abstract】 Change Detection(CD)compares two or more remote sensing images acquired at different times in the same scene and analyzes the differences between them to obtain the required change information.Synthetic Aperture Radar(SAR)is a relatively common way to acquire remote sensing images in remote sensing technology.It uses the principle of synthetic aperture to achieve high-resolution microwave imaging.It is not affected by external conditions such as weather and can work all-weather and all day.Therefore,SAR image change detection has become a research hotspot.With the deepening of research,there have been many excellent SAR image change detection methods,but there are still some problems.Aiming at the problems of existing change detection methods,this paper proposes SAR image change detection based on self-supervised learning and deep relation network,which mainly includes the following three research contents:1.Given the weak feature learning ability of traditional change detection methods and the problem that deep learning methods rely on a large amount of labeled data,this paper proposes a change detection method based on self-supervised learning and deep generalized canonical correlation analysis(DGCCA).This method is inspired by the idea of self-supervised learning and the characteristics of change detection tasks.It draws on the deep generalized canonical correlation analysis method to design pretext.The unsupervised multi-view representation learning feature of DGCCA gives it a powerful ability to extract non-linear features and shared features associated with multiple views and learn discriminative features beneficial to downstream tasks.Through experiments on four sets of real SAR data sets,it is verified that this method can significantly reduce the use of labeled samples and obtain better detection results.2.Aiming at the problem of poor generalization performance and decreased detection accuracy of traditional methods in multi-distribution and cross-domain change detection,a change detection method based on fine-grained category traversal and deep relation network is proposed.This method proposes a fine-grained category traversal model,which divides more refined categories according to different distribution data,and then extracts the key features of these categories through the category traversal module,to better extract the feature of multi-distributed data.At the same time,with the help of the deep relation network,the non-linear similarity measure is learned by comparing the samples in the support set and the query set,to accurately represent the data distribution of the complex change area and improve the discrimination performance of the method,and improve the generalization performance in multi-distribution and cross-domain change detection.Through experiments on four sets of real SAR data sets,the effectiveness and robustness of the method in this chapter are verified.3.Aiming at the problem of poor detection of SAR image boundary details by traditional methods,a change detection method based on graph neural network and fine-grained deep relation network is proposed.The graph neural network can process the data structure of non-Euclidean space and is good at extracting the topological information of the graph.The introduction of a graph neural network can fully extract the relationship information between pixels in SAR images,thereby improving the performance of boundary detail detection.This method also performs One-Hot encoding of the labels and splices them with the corresponding sample together to pass the label information of existing data to the test data,which further improves the generalization performance in multi-distributed and cross-domain change.Through experiments on four sets of real SAR data sets,the effectiveness and robustness of the method in this chapter are verified.更多还原