A point-based algorithm for multi-spectral image registration using mutual information of regions

A point-based image registration algorithm for multi-spectral imagery is described in this paper. On the basis of point mapping techniques, it deals with the registration of images with affine geometric distortion. First, the initial set of correspondent points is acquired from SUSAN corners using the similarity of regional mutual information (RMI): If the RMI value of two points is larger than the threshold, they belong to the initial set of correspondent points (multivalued mapping). Next, a relaxation algorithm restricted by the invariance of Mahalanobis distance in affine geometric distortion is proposed: first the matching intensity and definitude grade of the initial set of correspondent points is calculated; and the one to one correspondent points with high matching intensity and high definitude grade are obtained by relaxation algorithm; next, correct correspondent points are exracted by the invariance of Mahalanobis distance in affine geometric distortion; and incorrect correspondent points are deleted from the initial set to get a new initial set; then, the relaxation is practised again and again until no new correct correspondent points can be extracted. Finally, initial transformation parameters are estimated based on the sufficient control points. An iterative refined parameter process is devised to improve the accuracy of registration. Experimental result shows high accuracy registration can be acquired by the algorithm.