SUBPIXEL IMAGE REGISTRATION (IMAGE PROCESSING, IMAGE REGISTRATION, MATCHING)

This dissertation introduces and analyzes a new image registration topic--subpixel registration. Four algorithms for subpixel registration are presented and then analyzed according to their accuracy and computational requirements. One algorithm requires the resampling of an image. The accuracy of this resampling process is evaluated by comparing the frequency response of an interpolation function sampled with a high sampling frequency to an ideal low-pass filter. Using bilinear interpolation of simulated image intensities, a 0.005 pixel registration accuracy can be achieved. The four registration algorithms can be ordered according to accuracy as follows: intensity interpolation algorithm, difference method, correlation one-dimensional interpolation, correlation two-dimensional interpolation and phase correlation.; A formal mathematical description of the iterative intensity interpolation algorithm is then presented. Analyses indicate that compared to a direct intensity interpolation algorithm, the iterative algorithm saves computations by a factor of more than ten thousand. Furthermore, iterative hill-climbing algorithms (a pick-Ng coarse-search method and three climbing strategies) have been successfully used for measuring object displacements with an accuracy of 0.05 pixel. These use two successive frames of speckle images of an object and achieve an additional factor of 10 savings in computations.; This fast and accurate method for subpixel registration can be used in a variety of application areas, such as motion estimation, nondestructive evaluation, image sequence analysis, etc. It provides a way to accurately measure the displacement of individual points of a plane, without any contact or disturbance. To our knowledge, a method with these characteristics has not been previously available.; A feature-based registration algorithm using the discrete cosine transform is proposed to be used when position-dependent noise appears in images.