Image registration is a process by which the most accurate match is determined between two images, which may have been taken at same or different times, by the same or different sensors, from the same or different viewpoints. It is one of the crucial steps in remote sensing. Image registration requires intensive computational effort, not only because of its computational complexity but also the increasing resolution of images. For high accuracy and robustness as well as low computational cost a suitable similarity metric, a reduction in search data and a robust search space strategy is needed.
Reduction in search data can be achieved by using a few subimages for image registration. We propose a new measure, called alignability, which shows the ability of subimages to provide robust and reliable results. We compare this feature to entropy, variance and gradient magnitude respectively. We show that using alignability produces not only reliable but robust results in image registration.
In the dissertation, we investigate mutual information as a similarity metric. We show the effect of bin size on mutual information. Increasing the number of bins discriminates more image intensities, on the other hand, in decreasing the number of bins mutual information becomes noisy or even fail. We propose two methods which formalize the selection of bin size. We show that using the proposed methods increases the robustness of mutual information.
Mutual information has emerged in recent years as a popular similarity metric in the registration of images. Unfortunately, it ignores the spatial information contained in the images such as edges and corners that might be useful in the matching of images. It takes into account only the relationships between corresponding individual pixels and not those of each pixel's neighborhood. However, it is essential to consider both quantitative and qualitative information in the registration of images. We propose a new similarity metric, called spatial mutual information, which combines mutual information and a weighting function based on image gradient, image variance, and image entropy of local regions. Salient pixels in the in regions with high gradient, high variance and high entropy contribute more in the estimation of mutual information of image pairs being registered. We show that spatial mutual information is more robust to noise than mutual information. We also demonstrate that the spatial mutual information is not only more robust than mutual information but more reliable in the registration of multitemporal images.
A search space strategy based on Robbins-Monro stochastic approximation algorithm is also introduced. Results show that the algorithm is not only fast but robust as well.