An efficient registration and fusion algorithm for large misalignment remote sensing images

摘要

In this paper, an efficient technique to perform automatic registration and fusion for large misalignment remote sensing images is proposed. It complements SIFT features with Harris-affine features, and uses the ratio of the first and second nearest neighbor distance to setup the initial correspondences, then uses the affine invariant of Mahalanobis distance to remove the mismatched feature points. From this correspondence of the points, the affine matrix between two different images can be determined. All points in the sensed image are mapped to the reference using the estimated transformation matrix and the corresponding gray levels are assigned by re-sampling the image in the sensed image. Finally, we develop Burt's match and saliency metric and use neighborhood space frequency to fuse the registrated reference and sensed remote sensing images in NSCT domain. Experiments on remote sensing images with large misalignment demonstrate the superb performance of the algorithm.