| Image registration is the process of geometrically matching the corresponding pixels in images captured for the same scene at different times and/or sensors from various perspectives. Image registration approaches can be classified according to many viewpoints, such as various transformation models, spatial versus frequency domain, single- versus multi-modal, and intensity-based versus feature-based. Image registration plays a central role in several applications, such as remote sensing, medical imaging, computer vision, change detection, and super-resolution.;In this dissertation, we address the impacts of the arbitrarily-shaped locally variant illuminations on the geometric registration precision. Given a perfect camera following a pin-hole camera model and no object motions exist in the scene, we propose an intensity-based image registration model that can handle arbitrarily-shaped local illumination variations, ASLIV. Then, the ASLIV model is cast in a registration approach whose idea is based on iteratively segmenting the absolute image difference between the input images into distinct illumination regions. Assuming gain and offset uniformity along with each region, the proposed approach applies different M-estimators simultaneously. Each estimator has its own objective function that is assigned to a certain illumination region. The residuals of an illumination region are then differently penalized in accordance with their own objective function to minimize the registration error. In addition, some areas located on the boundary of each region could be mis-segmented due to the iterative process of creating the illumination regions. To lessen the negative impacts of such misclassified areas, a weighting function has been used. The proposed approach is cast in an iterative coarse-to-fine scheme to allow for large motions.;Experiments show that the proposed approach yields clear improvements in terms of geometric registration precision and illumination correction with a slight increase in computational time compared to competing approaches. As well, the proposed approach shows more resistance against segmentation perturbations as opposed to others. Real and simulated image pairs are employed in the experiments. The performance of competing approaches is evaluated using: normalized cross-correlation (NCC), structural similarity (SSIM) index, and peak signal-to-noise ratio (PSNR).;The problem is that the geometric registration precision could be impacted due to local illumination variations. Thus, any subsequent processing would be easily negatively affected. In early research, the registration process assumed brightness constancy. Recently, some research has incorporated illumination variations in the registration process in a limited manner, such as using a global or an affine illumination model. |
