Statistically Consistent and Novel Image Registration Methods for Various Transformations in Presence of Zooming

Abstract

In image processing literature, image registration is the process of spatially matching two or more images of the same scene or object obtained at various times, from different perspectives, or by separate sensors in order to simplify comparison, fusion, or analysis. Its purpose is to identify the spatial transformation that minimizes the differences between related characteristics or intensities in the reference and moved images. Image registration techniques find a wide range of applications in several imaging fields, namely, medical imaging, satellite imaging, fingerprint matching, remote sensing, and image comparison, to name a few. The growing popularity of this imaging field motivates numerous researchers to develop novel and improved methods of image registration. In this dissertation, we develop several image registration techniques under various image transformations, especially when zooming or scaling is involved. In our first research-work, we develop an intensity-based image registration technique that registers two images efficiently, where one is a zoomed-in version of the other. In order to improve the performance and also to reduce the computational complexity of the method at the same time, we opt for a feature-based approach for the same problem in the second research-work of this dissertation. Motivated by the success of the feature-based approach, we use similar approach to image registration for a more complex image transformation where translation, rotation, and zooming are all involved in our final research-work. All our proposed methods are backed by extensive numerical studies and comparisons with several state-of-the-art methods, which show our methods' superiority over them. We also provide theoretical justifications when the image resolutions are high.