A MULTIRESOLUTION HIERARCHICAL APPROACH TO IMAGE SEGMENTATION BASED ON INTENSITY EXTREMA

The aim of our research has been to create a computer algorithm to segment grayscale images into regions of interest (objects). These regions can provide the basis for scene analysis (including shape parameter calculation) or surface-based shaded graphics display. The algorithm creates a tree structure for image description by denning a linking relationship between pixels in successively blurred versions of the initial image. The image description describes the image in terms of nested light and dark regions. This algorithm can theoretically work in any number of dimensions; the implementation works in one, two, or three dimensions. Starting from a mathematical model (developed by Koenderink) describing the technique, our research has shown that • by explicitly addressing the problems peculiar to the discreteness of computer representations the segmentation described by the mathematical model can be successfully approximated; • although the image segmentation performed sometimes contradicts semantic and visual information about the image (e.g., part of the kidney is associated with the liver instead of the rest of the kidney), simple interactive postprocessing can often improve the segmentation results to an extent sufficient to segment the region desired; • the theoretical nesting property of regions, originally thought to hold in all circumstances, does not necessarily apply to all pixels in a region. The interactive postprocessing developed selects regions from the descriptive tree for display in several ways: pointing to a branch of the image description tree (which is shown on a vector display), specifying by sliders the range of scale and/or intensity of all regions which should be displayed, and pointing (on the original image) to any pixel in the desired region. The algorithm has been applied to approximately 15 CT images of the abdomen. While performance is frequently good, work needs to be done to improve performance and to identify and extend the range of images the algorithm will segment successfully. © 1990 IEEE