An analysis of node linking in overlapped pyramids

Abstract
An image segmentation algorithm is described that uses an overlapped pyramid to represent an image at multiple levels of resolution. The procedure `lifts' objects to levels of lower and lower resolution until they become `spot' or `streaklike' and are identifiable by local processing (using 3 by 3 operators). They are then `rooted.' The result is a forest embedded within the pyramid, with the single tree rooted at the pyramid apex representing the background regions and the remaining trees representing compact object regions. In addition to the definition of the pyramid linking algorithm, the convergence of the algorithm is proved and optimal rooting rules for binary images are analyzed.