| Image registration is one of the crucial steps in the analysis of remotely sensed data. A new acquired image must be transformed, using image registration techniques, to match the orientation and scale of previous related images. Image registration requires intensive computational effort not only because of its computational complexity, but also due to the continuous increase in image resolution and spectral bands. Thus, high-performance computing techniques for image registration are critically needed. Very few works have addressed image registration on contemporary high-performance computing systems. Furthermore, issues of load balancing, scalability, and formal analysis of algorithmic efficiency were seldom considered.; This dissertation introduces high-performance automatic image registration (HAIR) algorithms. High performance is achieved by: (1) reduction in search data, (2) reduction in search space, and (3) parallel processing.; Reduction in search data is achieved by performing registration using only subimages. A new metric called registrability is used to select those subimages such that accuracy is maintained. In addition, a histogram comparison is used to discard anomalous subimages, such as those with clouds. Further data reduction is obtained using an iterative refinement search (IRA), which exploits the wavelet multi-resolution representation. This technique starts searching images with lower resolution first, then refining the results using higher resolution images to use the least possible data points in the overall registration task.; Reduction of search space is achieved through two methods. First, iterative refinement reduces dramatically the number of solutions examined. In addition, genetic algorithms were also used to further expedite the search.; Parallel processing techniques have been utilized to provide coarse-grain load-balanced parallel algorithms based on iterative refinement as well as genetic algorithms. Two hybrid algorithms have been also devised in order to integrate the strengths of iterative refinement, genetic algorithms, automatic subimage selection, and parallelism.; The proposed algorithms were shown, experimentally and analytically, to provide substantial improvements in both accuracy and performance when applied to remote sensed images. Test images included LandSat Thematic Mapper (TM), Advanced Very High Resolution Radiometry (AVHRR), Geostationary Operational Environmental Satellite-8 (GOES-8), and Synthetic Aperture Radar (SAR). The parallel algorithms have exhibited good load-balancing scalability on contemporary parallel computers, including Cray T3E and Beowulf Parallel Clusters. |
