Multilevel similarity model for high-resolution remote sensing image registration

Abstract Image registration is a prerequisite and the basis for many important applications of remote sensing images. Compared with medium-/low-resolution images, high-resolution (HR) remote sensing images exhibit considerable resolution differences, complex distortions and repeatable textures. Most of the existing registration methods are designed for images with medium/low resolutions. However, these methods typically suffer from many false matches of keypoints when working with HR images. This problem often causes automatic registration to fail in applications. To address these problem, we propose a multilevel similarity model for HR remote sensing image registration. The multilevel similarity model includes three progressive levels of elements: the similarity of keypoint physical size (i.e., point-like similarity), the similarity of textures between two keypoints (i.e., line-like similarity) and the similarity of keypoint space relationship (i.e., plane-like similarity). First, a candidate match set of keypoints is identified depending upon the physical sizes of the keypoint blob-like structures, so that many useless keypoints can be significantly excluded. Then, a minimum spanning tree is developed to discover the false matches, where the weights of the tree are estimated based on the similarities of image windows created between two target keypoints and their candidate homologous keypoints. Finally, a spatial relationship matrix is constructed to further refine the matches between images by efficiently coding the relative spatial locations among keypoints. Experiments were conducted on various HR remote sensing images with different resolutions and distortions, and the experimental results demonstrate the effectiveness of our method.