基于光流法和核线约束的SAR与光学遥感图像密集匹配方法研究

Applications that jointly use multi-sensor images can overcome limitations of single-sensor images; the fundamental task of these remote sensing applications is accurate image registration. Specifically, dense matching aims to generate pixelwise correspondence, which is the essential prerequisite for dense 3D surface reconstruction, pixel-level image fusion and change detection. For dense matching of homologous remote sensing image pairs, the key is to perform the epipolar geometry to reduce the matching problem from 2D to 1D. However, a rigorous epipolar-line does not exist for optical and Synthetic Aperture Radar (SAR) image pairs, which forces us to find alternatives for dense matching between optical and SAR images. Moreover, the registration between optical and SAR images is still an open problem due to the drastic differences. In this study, we aim at solving dense matching between optical and SAR image based on optical flow and epipolar constraint. Starting from the image domain, we investigate the robust feature description that satisfies the brightness constancy assumption in optical flow, then estimate the pixelwise flow field and obtain the disparity map. Then considering the imaging geometry, we analyze the property of epipolar line based on the imaging model and compensate the errors for the uncritical epipolar line to make it fulfill the requirement of traditional dense matching methods. This project firstly proposes an improved optical flow model for SAR and optical images, and then analyzes the geometric properties of the epipolar line between heterogeneous images. While solving the dense matching problem, it also provides the basis and new ideas for subsequent heterogeneous remote sensing applications.

异源遥感图像的联合应用能够克服单一传感器的限制，而同时使用异源图像的前提是图像匹配。许多应用，如三维地表重建、图像融合等，都需要用到图像中像素的一一对应，即密集匹配。同源遥感立体像对的密集匹配可以通过核线约束来将匹配问题从二维减至一维。但合成孔径雷达（SAR）和光学图像是两种典型的异源图像，不存在严格的核线约束，不能直接使用传统密集匹配方法，且由于它们之间的差异较大，其匹配尚未很好解决。因此，本项目旨在提出解决SAR和光学图像密集匹配问题的方法。首先从图像域出发，研究满足光流法模型亮度恒定假设的特征描述，估计像素的运动流向，得到逐点对应图；其次从成像几何出发，利用成像模型分析核线性质，研究对不严格核线进行误差补偿的方法，使其满足传统算法的要求。本项目首次提出适用于SAR和光学图像的改进光流法模型，分析这两种图像之间的核线约束，在解决密集匹配问题的同时，也为后续遥感应用提供了基础和新思路。

光学和SAR图像密集匹配是多源遥感影像联合应用的前提之一。针对光学和SAR影像表征差异过大引起匹配失效和异源影像成像模型不同引起核线约束不严格的问题，我们首先提出了一种针对光学和SAR影像的一致性特征构建模型，能够显著减小这两种异源遥感影像之间的辐射和几何差异，然后基于一致性特征提出了一种结合三维相位相关的亚像素精度光学和SAR图像配准算法，验证了一致性特征在图像配准任务下的有效性。进而我们提出了基于一致性特征模型和光流法的全局化和局部化框架用于逐像素光学和SAR影像密集匹配，可以有效解决几何校正处理引入的局部畸变。最后我们研究了光学和SAR卫星严密成像模型和通用成像模型下的几何性质和核线约束理论模型，提出了基于异构模型的光学和SAR异源影像的联合定位精度提升方法和一种核线约束不严格条件下的核线影像生成算法。通过本项目的研究，有望推动基于异源影像的三维重建和变化检测应用。