Band registration is one of the most critical steps in the production of multi/hyperspectral images and determines the accuracy of applications directly. Because of the characteristics of imaging devices in some multi/hyperspectral satellites, there may be a time difference between bands during push-broom imaging, which leads to displacements of moving clouds with respect to the ground. And a large number of feature points may gather around cloud contours due to the high contrast and rich texture, resulting in building a transformation more suitable for moving clouds and making ground objects ghosted and blurred. This brings a big challenge for registration methods based on feature extraction and matching. In this paper, we propose a novel coarse-to-fine band registration framework for multi/hyperspectral images containing moving clouds. In the coarse registration stage, a cloud mask is generated by grayscale stretching, morphology and other operations. Based on this mask, a coarse matching of cloud-free regions is performed to eliminate large misalignment between bands. In the refinement stage, low-rank analysis and RASL (Robust Alignment by Sparse and Low-rank decomposition) are used to optimize the rank of coarse results to achieve fine registration between bands. After experiments on a total of 102 images (83 cloudy images and 19 cloud-free images with all 32 bands) from Zhuhai-1 hyperspectral satellite, our method can achieve a registration accuracy of 0.6 pixels in cloudy images, 0.41 pixels in cloud-free images, which is enough for subsequent applications.
展开▼
