Automated Comparison of Image Similarities and Differences.

An image comparison algorithm for describing the similarities and differences between images is presented. Previous work on the image correspondence problem (involving relaxation algorithms) has focused on determining the similarities between images, with differences treated as noise to be minimized. The image comparison algorithm focuses equally on the description of image differences. With differences of equal concern, the algorithm must counter the tendency of traditionally used measures of similarity (such as cross-correlation) to distort differences into similarities. The mechanisms designed to measure image differences make possible a more complex algorithm which is capable of building a model of both the comparison process and environmental noise. Such a model is possible because the comparison algorithm generates discrete hypotheses concerning difficulties it encounters in the comparison process. The hypotheses are evaluated for consistency with other hypotheses made at different times and involving various spatial relationships within the comparison results. Where inconsistencies are detected, more complex hypotheses or actions are made. The mechanisms by which consistency is maintained are contrasted with the consistency mechanisms used in relaxation. Issues related to human visual processing are explored. For example, the algorithm uses primitive attentional mechanisms. These mechanisms are based upon the algorithm's ability to detect discrepancies in its model of the environment and , based upon that model, focus upon problematic comparison regions requiring further investigation. Finally, the images used to test the comparison algorithm are derived from two-dimensional electrophoetograms. The comparison algorithm uses these images as part of a genetic study to empirically estimate the rate of human mutation.Ph.D.Computer scienceUniversity of Michiganhttp://deepblue.lib.umich.edu/bitstream/2027.42/160041/1/8412253.pd