基于深度学习和层次语义模型的极化SAR分类

发布时间：2018-08-03 17:21

【摘要】：针对复杂场景的极化合成孔径雷达(Synthetic aperture radar,SAR)图像,堆叠自编码模型能够自动学习高层特性,有效表示城区、森林等复杂地物的结构,然而,却难以保持图像的边界和细节.为了克服该缺点,本文结合深度自编码器和极化层次语义模型(Polarimetric hierarchical semantic model,PHSM),提出了新的无监督的极化SAR图像分类算法.该方法根据极化层次语义模型,将复杂的极化SAR图像划分为聚集、匀质和结构三大区域.对聚集区域,采用堆叠自编码模型进行高层特征表示,并构造字典得到稀疏特征进行分类;对匀质区域,采用层次模型进行分类;对于结构区域,进行线目标保留和边界定位.实验结果表明,该算法通过不同的分类策略优势互补,能够得到区域一致性好且边界保持的分类结果.
[Abstract]:For the polarimetric synthetic Aperture Radar (Synthetic aperture) images of complex scenes, the stacked self-coding model can automatically learn the high-level characteristics and effectively represent the structures of complex objects such as urban areas and forests. However, it is difficult to maintain the boundaries and details of the images. In order to overcome this shortcoming, a new unsupervised polarimetric SAR image classification algorithm is proposed by combining the depth self-encoder with the polarimetric hierarchical semantic model (Polarimetric hierarchical semantic model PHSM). According to the semantic model of polarization hierarchy, the complex polarimetric SAR images are divided into three regions: aggregation, homogeneity and structure. For the clustering region, the stacked self-coding model is used for high-level feature representation, and the dictionary is constructed to obtain sparse features for classification; for homogeneous region, hierarchical model is used to classify; for structural region, line object reservation and boundary location are carried out. The experimental results show that the proposed algorithm can obtain the classification results with good regional consistency and boundary preservation by complementing the different classification strategies.
【作者单位】：西安电子科技大学计算机学院;西安理工大学计算机科学与工程学院;西安电子科技大学智能感知与图像理解教育部重点实验室;福建农林大学计算机与信息学院;
【基金】：国家重点基础研究发展计划(973计划)(2013CB329402) 国家自然科学基金(61573267,61571342,61572383);国家自然科学基金青年科学基金项目(31300473) 教育部“长江学者和创新团队发展计划”(IRT1170) 高等学校学科创新引智计划(B07048) 福建省自然科学基金(2014J01073)资助~~
【分类号】：TN957.52

【参考文献】