Image registration is an important pre-processing operation to be performed before many image exploitation and processing functions such as data fusion, and super-resolution frame. Given two image frames, obtained from the same sensor or from different sensors, the registration problem involves determining the transformation that most nearly maps (or aligns) one image frame into the other. Typically, image registration requires intensive computational effort and the developed techniques are scene dependent. Furthermore, the problems of multimodal image registration (i.e. problem of registering images acquired from dissimilar sensors) and sub-pixel image registration (i.e. registering two images at sub-pixel accuracy) are highly challenging and no satisfactory solutions exist.This dissertation introduces novel techniques to solve the image registration problem both at the pixel-level and at the sub-pixel level. For pixel-level registration, a procedure is offered that enjoys the advantages that it is not scene dependent and provides the same level of accuracy for registering images acquired from different types of sensors. The new technique is based on obtaining the local frequency content of an image and using this local frequency representation to extract control points for establishing correspondence. To extract the local frequency representation of an image, a computationally efficient scheme based on minimizing the latency of a Gabor filter bank by exploiting certain biological considerations is presented. The dissertation also introduces an extension of using local frequency to solve the sub-pixel image registration problem. The new algorithm is based on using the scaled local frequency representation of the images to be registered, with computationally inexpensive scaling of the local frequency of the images prior to correlation matching. Finally, this dissertation provides a novel approach to solve the problem of multi-modal image registration. The principal idea behind this approach is to employ Computer Aided Design (CAD) models of man-made objects in the scene to permit extraction of regions-of-interest (ROI) whose local frequency representations are computed for extraction of stable matching points. Detailed performance evaluation results from an extensive set of experiments using diverse types of images are presented to highlight the strong points of the proposed registration algorithms.
展开▼
