| Image registration is a procedure of finding a spatial deformation to match two images (2D or 3D), and it has received considerable attention in the areas of image analysis and computer vision.; Image registration methods can be divided into landmark-based and intensity-based methods. Landmark-based methods need to first identify homologous features (points, curves, or surfaces) as landmarks and evaluate the deformation that matches the landmarks. Intensity-based methods are always based on intensity similarity measures and try to find a deformation to optimize the similarity measure between two images. In an image registration system, three issues must be considered: validity, accuracy, and efficiency. It is difficult to achieve all of these issues with previous methods.; This dissertation describes a deformable model for image registration and object matching. The deformable model outlined in this dissertation is based on the concepts of free-form deformation, hierarchical B-splines, and scattered data approximation. It can be used for both landmark-based and intensity-based image registration. The method is adaptive in that it puts more effort in complex regions that need more attention. This coarse-to-fine strategy not only makes the registration system efficient, but also guarantees a valid matching. It has enough degrees of freedom for an accurate matching as well. |
