Computation of object cores from grey-level images

Pixels in discrete images represent not just dimensionless points but the integration of image information over spatial areas. The size of this measurement aperture (the scale of image measurements) limits our ability to discern finer-scale structure. Changing the magnification of objects in images changes the level of detail that can be described. Oversensitivity to such detail causes difficulty in making shape-based comparisons for such tasks as identification or registration. To create object representations that are invariant to magnification, one must tie the measurement scale to the size of the objects involved. Research at UNC-CH has led to a representation known as a core that establishes this relationship. Cores are related to previous medial representations of shape but go farther to capture the central essence of objects. This object-relevant use of scale separates general shape properties of objects from specific detail (allowing more robust representations when comparing objects) and provides a basis for object-based tradeoff between noise and loss of detail (allowing finer, but noisier, representations for small objects while simultaneously producing less noisy representations for larger, more stable, objects). This dissertation presents four distinct algorithms that, taken together, compute cores for objects in images. They derive cores directly from the image and do not require a prior segmentation, thus producing image segmentations as well as object representations. They include: (1) An algorithm, similar to the Hough transform, for transforming boundary-likelihood information (graded boundary values, not binary contours) to medial-likelihood information at all spatial positions and scales; (2) An algorithm for finding curves in this scale space of medial information that are locally optimal with respect to both position and scale in directions transverse to the curve; (3) A competitive credit attribution (feature-binding) algorithm that constrains the medial-likelihood computation to produce a more coherent interpretation of the image; (4) A cooperative algorithm that encourages longer connected cores and completion of cores across gaps. These algorithms are evaluated using both qualitative testing on various types of images and quantitative study of their behavior under systematically controlled variations in images. Comparisons are also made between computed cores and the results of previous and ongoing psychophysical studies.