Abstract
To form a comparison of two images of the same object field obtained at different times or under different conditions, it is necessary spatially to register the images pixel-by-pixel, and thereby correct for relative translational shifts, magnification differences and other geometrical distortions. A common technique for registering a pair of image is to form a cross-correlation between the images and determine the location of the maximum correlation. However, there are many cases where one of the pair images is partially covered, for example, such as by clouds or snows in remote sensing images.
In this paper, how the cross-correlation function is distorted when one image is partially covered is analyzed, first. Then, new methods for the registration between partially covered images are proposed.
It is shown that, in the applications, there are three types of such image coverages; 1) partially lacked or darkened, 2) partially added with another pattern, and 3) partially replaced or occluded by another pattern. These types of coverage model are formalized, and the registration methods are proposed for these three types, respectively. Experimental results of the proposed methods applied to example images in remote sensing, medical imaging and scene analysis are also shown. Correct registrations are obtained in spite that one of the pair image to be registered is partially covered.
