Size-dependent image resampling for mutual information based remote sensing image registration

Abstract

Registration consistency has been used as a performance evaluation criterion for mutual information based image registration techniques when the ground truth is not known. In practice, when the spatial resolutions of the two images to be registered are different, the low resolution image is often chosen as the floating image to expedite the registration process because it involves fewer pixels. However, we have found that this choice introduces problems when the difference in spatial resolution is large. This is because the resulting mutual information registration function calculated through linear interpolation or partial volume interpolation can be extremely rough that makes the optimization hard to perform and the registration result unreliable. The main contribution of this paper is the development of a size-dependent kernel to resample the high resolution reference image for joint histogram estimation. Since the size of the support of the kernel can be very large, the computational load of this approach is high and loses the advantage of using the low resolution image as the floating image. As an alternate approach, an offline preprocessing of the high resolution image is proposed in this paper. After preprocessing the high resolution reference image, conventional linear and partial volume interpolations can be employed to estimate the joint histogram efficiently. A HyMap image (6.8m/pixel) and a digital aerial photograph (0.15m/pixel) are used in our experiments to demonstrate the effectiveness of the proposed approach.