基于超像素和CRF模型的高分辨率极化SAR图像语义分割技术研究

The semantic segmentation of polarimetric synthetic aperture radar (PolSAR) images has been widely used in both military and civil fields. The high-resolution (HR) PolSAR images can acquire abundant detailed information of terrains. However, the existing semantic segmentation approaches for PolSAR images mainly treat single pixels as processing cells, which can not effectively employ the terrains’ regional distribution property that generally exists in HR PolSAR images. Moreover, these methods are often influenced by the inaccurate statistical modeling of PolSAR images, and are also insufficient for integrating the multi-feature and information. To solve the problems in the existing semantic segmentation of PolSAR images, some key issues are researched in this project, such as the fast and robust superpixel generation method of HR PolSAR images, the accurate statistical modeling of HR PolSAR images and the design of conditional random field (CRF) model on HR PolSAR image superpixels. And then a novel semantic segmentation algorithm for HR PolSAR images is proposed. The project is of great theoretical and practical significance for promoting the development and application of PolSAR image interpretation technologies.

极化合成孔径雷达(Synthetic Aperture Radar，SAR)图像语义分割在军事和民用领域具有广泛应用。高分辨率极化SAR图像能获取地物丰富的细节信息，但现有极化SAR图像语义分割方法主要以像素为处理单元，难以充分利用高分辨率极化SAR图像地物的区域分布特性，并受极化SAR图像统计建模不准确的影响，此外还存在多特征信息融合能力不足等问题。本项目针对现有极化SAR图像语义分割技术存在的问题，重点研究高分辨率极化SAR图像超像素快速鲁棒生成、高分辨率极化SAR图像精确统计建模以及面向高分辨率极化SAR图像超像素的条件随机场(Conditional Random Field，CRF)模型设计等问题，提出高分辨率极化SAR图像语义分割新方法。本项目研究对推动极化SAR图像解译技术发展及应用具有重要的理论与实际意义。

本项目围绕高分辨率极化合成孔径雷达(Synthetic Aperture Radar，SAR)图像语义分割，针对传统方法不能充分利用区域分布特性且效率较低、多特征融合不足以及人工特征有效性不强等问题，基于超像素和条件随机场(Conditional Random Field，CRF)模型研究了三方面内容。1.为有效利用区域分布特性并提升算法效率，设计了基于相干斑滤波与线性谱聚类的极化SAR图像超像素生成算法，研究了面向超像素的极化SAR图像语义分割方法；2.为充分发挥不同特征的优势，设计了基于多特征和CRF模型的高分辨率极化SAR图像语义分割算法；3.为提取和利用表征性强的特征，设计了基于卷积神经网络的高分辨率极化SAR图像特征提取与语义分割算法。本项目提供了高分辨率极化SAR图像语义分割新方法，推进了SAR图像解译技术的发展与应用，具有一定理论与实际意义。