Out of Core Voxel Data Registration

Abstract

Image Registration is the task of aligning two different images of the same scene taken from different viewpoints, at different times or with different sensors. Image Registration is a task often performed in processing of two and three dimensional images before further processing of the images takes place. This thesis will develop a method for registering large three dimensional images from the PRESIOUS project. These three dimensional images are created from X-ray CT scans of stones from cultural heritage sites. The goal is to measure the effects of erosion on these stones. The high resolution of these datasets makes it necessary to develop a registration approach that can register images too big for the main memory. The registration method developed in this thesis is separated into two phases. The first phase does a rough registration on downscaled versions of the images to get an initial alignment, and the second phase does a more accurate registration with the full scale datasets, using the external hard drive to cache data too large for the main memory. The first phase is quite fast, and in most cases provides a highly accurate initial alignment. However, in some cases the registration method fails as the method employed is not guaranteed to find the global optimal solution. The second phase of the registration process is very time consuming because a simple but expensive optimization strategy is employed. The second phase registration method produces accurate results when provided with a good initial alignment.