Robust multi-sensor image registration by using gradient-based statistical information

Abstract

This thesis deals with the robust registration algorithm based on a new similarity measure for aligning multi-sensor images such as charged-couple device (CCD) and infrared (IR) images. Multi-sensor image registration is essential to many applications such as surveillance, remote sensing, and medical imaging. Due to different sensor characteristics and/or different sensor pose in the system, there exist geometrically relative motions such as translation, rotation, scale, and any other transformations between two sensor images. Therefore, the registration can be defined as a process to establish the spatial correspondence between two images. In this thesis, we aim to design a robust similarity measure for registering multi-sensor images to compensate such relative motions which are confined to a global projective model. As a similarity measure for registration, mutual information (MI) based measures have been widely used for medical imaging. These MI-based methods work properly only if the intensity mapping relationship between the two images is global. However, the intensity mapping relationship is often locally different in many other multi-modal images such as ordinary CCD and IR images which are used for surveillance and remote sensing applications. Therefore, MI-based methods cannot cover such various image modalities and their registration performances are limited. Hence, to improve the registration performance, we propose a new similarity measure based on gradient-based statistical information. For robust registration, the proposed similarity measure is based on the entropy obtained from a 3-D joint histogram incorporating edginess and generalized gradient vector flow (GGVF) information. In the measure, in order to make a reliable mapping relationship between the edge regions of two images, the concept of edginess which is normalized version of gradient magnitude is adopted so that registration performance may be affected mainly by edge existences rather ...