Elastic deformable models and deformable image matching methods have been developed in recent years. Because elastic deformable models can describe the object perfectly in a physical way, the image matching methods based on them can remove the detailed structural variation between individuals by matching a study image to a target image. Thus these techniques have widely been used in various fields, for example, computer graphics, computer vision, medical image processing and so on. The author performed a deep study of this area and proposed some new methods. The main contributions in this thesis can be summarized as follows: [1] Propose a new elastic matching method incorporating geometry based shape information. This elastic deformable model integrates two kinds of methods to compute the driving forces: a method based on region information and a method based on feature information. A new computing external forces method, which incorporating geometry shape information, is proposed. These external forces drive the deformation more correctly and robustly. [2] Propose a new robust elastic matching method based on a hybrid elastic model. This method can jointly estimate the correspondence and nor-rigid matching, which need not extract features, works directly on gray level images. This method also takes a multiresolution strategy and linear model to reduce the computation complexity, and approaches better matching. The method can not only be used for medical image matching field, but also be used for other deformable matching. [3] Compared with the former hybrid elastic model, a new improved model is presented for deformable image matching and hominid morphology studies. This improved method not only uses the gray level information, but also involves the weighted gray histogram feature information and the image gradient information to realize deformable image matching. This new improved method takes a multiresolution strategy to search tile initial position of the corresponded spring net grid, which improve the deformation result more correctly. [4] It is a hot research field for dealing with the 3D data. The traditional methods work directly on original data in the 3D space. Because of its huge data, it is usually time consuming. A new method is proposed to deal with the 3D data. The 3D data is first described in polar representation, so that, the 3D data is converted into 2.5D data and can be described in 2D space: In this case, we can apply our deformable matching method to do non-rigid image matching in the 2D space. Finally, the matching result can be converted into 3D space again. With this method, we can realize the 3D deformable matching in the 2D space.
修改评论