Multimodality registration using statistical and spatial information

Abstract

Multimodality registration is important in many research fields such as remote sensing, medicine, and computer vision for fusing information of different characteristics from multimodal images. This thesis deals with two multimodality registration algorithms, or multi-sensor image registration and model-to-image registration. Multi-sensor image registration aims to geometrically align multimodal images such as visible and infrared images of the same scene acquired by different sensors. Registration is achieved by iteratively minimizing an objective function (or maximizing a similarity measure) through updating transformation parameters. Recently, for registering medical multimodal images, mutual information is widely used as a similarity measure. However, since the mutual information does not include spatial information, the mutual information based registration suffers from local optima or incorrect global optimum problems. This thesis presents a new objective function for multi-sensor image registration by incorporating intensity information and edge orientation information. Experimental results show that the proposed objective function provides more robust registration than the existing ones. This thesis also deals with model-to-image registration. Accurate registration of a 3-D model and a 2-D camera image is useful in many applications. The accurate model-to-image registration can be achieved by precisely estimating both intrinsic and extrinsic camera parameters. Among them, intrinsic parameters are usually calculated in advance. Hence, this thesis attempts to accurately register the corresponding 2-D projected image of a given 3-D model to an input camera image by finding extrinsic camera parameter values. The proposed algorithm for accurate registration consists of two parts; initial registration and precise registration. In the initial registration, camera orientation parameters are estimated based on a generalized Hough transform by using roughly ob...