摘要
极化合成孔径雷达(PolSAR)图像配准在地物分类、变化检测、图像融合中都具有广泛应用。现有的PolSAR图像配准方法,无论是基于深度学习还是传统方法,大多采用PolSAR幅度影像信息进行处理。这种处理方式导致大量极化信息丢失,同时在PolSAR图像固有相干斑噪声影响下,配准精度和可靠性表现不佳。为此,本文首先发展了一种有效的基于极化白化滤波(PWF)精细化处理的关键点检测器,利用PWF对PolSAR图像进行相干斑噪声抑制,通过阈值约束、形态学腐蚀及非极大值抑制来选取显著且分布均匀的匹配关键点。进一步地,本文设计了一种孪生简单稠密胶囊网络(SimSD-CapsuleNet)来快速提取PolSAR图像的浅层纹理特征和深层语义特征,同时为了充分利用极化信息,本文将极化协方差矩阵作为输入数据。本文计算了胶囊形式特征描述符之间的距离,并将其输入硬L2损失函数用于模型的训练。本文方法在不同传感器获取的不同分辨率PolSAR图像上进行验证。结果表明,该方法能够在更短的时间内获取更加均匀且数量更多的匹配关键点,结合PWF和深度神经网络可以实现快速准确的PolSAR图像配准。
Polarimetric synthetic aperture radar(PolSAR)image registration has a wide range of applications in feature classification,change detection,and image fusion.Existing PolSAR image registration methods,whether based on deep learning or conventional methods,use PolSAR magnitude image information for processing.This processing leads to a large amount of polarization information loss,and at the same time,the registration accuracy and reliability perform poorly under the influence of the inherent coherent speckle noise of PolSAR images.To this end,this paper first develops a novel and effective key point detector based on polarization whitening filter(PWF)refinement processing,which uses PWF to suppress coherent speckle noise in PolSAR images and selects significant and uniformly distributed matching key points by threshold constraint,morphological erosion,and non-extreme value suppression.Further,in this paper,we design a Siamese simple dense capsule network(SimSD-CapsuleNet)to quickly extract the shallow texture features and deep semantic features of the data,and we use the polarization covariance matrix as the input data in order to make full use of the polarization information.In this paper,the distances between the capsule form feature descriptors are calculated and fed into a hard L2 loss function for the training of the model.The method in this paper is validated on PolSAR images acquired by different sensors with different resolutions,and the results show that the method can acquire more uniform and a larger number of matching key points in a shorter time,and the combination of PWF and deep neural network can achieve fast and accurate PolSAR image registration.
作者
项德良
丁怀跃
管冬冬
程建达
孙晓坤
XIANG Deliang;DING Huaiyue;GUAN Dongdong;CHENG Jianda;SUN Xiaokun(College of Information Science and Technology,Beijing University of Chemical Technology,Beijing 100029,China;Beijing Advanced Innovation Center for Soft Matter Science and Engineering,Beijing University of Chemical Technology,Beijing 100029,China;College of Operational Support,Rocket Force University of Engineering,Xi'an 710025,China;School of Automation and Electrical Engineering,Linyi University,Linyi 276000,China)
出处
《测绘学报》
EI
CSCD
北大核心
2024年第3期450-462,共13页
Acta Geodaetica et Cartographica Sinica
基金
国家自然科学基金(62171015)。
作者简介
第一作者:项德良(1989—),男,博士,教授,研究方向为城市遥感、SAR极化、SAR图像处理、人工智能、模式识别等。E-mail:xiangdeliang@buct.edu.cn;通信作者:管冬冬,E-mail:gdd@whu.edu.cn。
