几何配准是影像后续处理的重要前提，是遥感信息处理领域研究的热点之一。复杂地形区多时相遥感影像的高精度配准一直是难以突破的难题，光流估计法通过逐像素位移增量解算为此提供了可行的解决思路，但光流法对地物变化异常敏感，经常导致计算的光流场及配准影像存在异常。为此，本文提出一种基于光流校正的复杂地形区多时相遥感影像配准方法，采用亮度和梯度双重约束获取光流场初值，在此基础上使用高斯拉普拉斯算子对异常光流进行检测，然后通过Delaunay三角形曲面插值对异常光流进行校正处理，从而得到各像素精准位移。实验表明，本文提出方法对存在地物变化的复杂地形区多时相遥感影像，可实现高保真、高精度的配准。

Image registration is a process of geometric alignment of two or more images acquired at different time, different sensors or under different conditions (weather, illumination, camera position and angle, etc.). Remote sensing image registration is an important prerequisite for subsequent processing, such as image fusion, image stitching, long time sequence analysis etc., and it is one of spotlights in the field of remote sensing information processing. High-precision registration of multi-temporal remote sensing images covering complex-terrain region is always a problem to break through. The conventional registration algorithms guiding by the transformation model, is enable to take the pixel-level geometric distortion into consideration, which means that the displacements of a pair of corresponding pixels is different from that of the other pair.Under this circumstance, the global or local mapping function could not describe the geometric deformation between two images covering the complex-terrain region. Optical flow estimation calculates per-pixel displacements considering the very local distortions, even the pixel-level deformation in the computer vision field, providing a feasible and creative solution. It estimates displacement in x- and y-directions for a pair of corresponding pixels under the intensity and gradient consistency constraints, with resistant to the change of illumination. However, it is sensitive to land cover changes, which often lead to abnormal optical flow field and further affect the registered image after the coordinate transformation and resampling. To this end, a registration algorithm based on the optical flow modification for multi-temporal remote sensing images covering the complex-terrain region is proposed. On the preliminary optical flow field, Laplace of Gaussian operator is employed to detect the abnormal optical flow in Munsell color system. With the mask of abnormal optical flow based on the detection results, the Delaunay triangle curved surface interpolation is utilized to correct them, which is calculated by the around accurate pixel displacements. The coordinates in the sensed image are transformed, and the new pixel value is put on the corresponding pixel with the specified resampling method. Ultimately, the aligned image is generated. Experiments based on multi-temporal remote sensing images covering the complex-terrain region with land cover changes demonstrate that the proposed method achieves high-fidelity and high-precision registration compared with the results of the conventional methods.Nevertheless, for registration of the image with sub-meter spatial resolution or image registration of different sensors, the difference of imaging angle, imaging mechanism, noise type etc. have an impact on the accuracy of the proposed algorithm. These remote sensing images are important data guarantee for fine research of earth surface and disaster assessment under poor imaging conditions in disaster region. How to realize high fidelity and high efficiency registration of the ultra-high resolution image or the multi-model image, is a problem that needs an in-depth study for us. In our future work, based on the proposed algorithm in this paper, research will be carried out specifically for the aforementioned problem. The aligned complex topographic region images will be applied to disaster monitoring, assessment, land use change analysis and other fields for an assessment to further improve our proposed method.