A CRITICAL-VIEW OF PYRAMID SEGMENTATION ALGORITHMS

Although pyramid segmentation algorithms have been reported to be successful in a large number of fields, some of their basic properties have not been systematically investigated. They were especially used in applications involving compact objects, in which moreover the connectivity of the final segments was of no great importance. We will show that most pyramid segmentation algorithms are shift-, rotation- and scale-variant and hence that the results of a pyramid segmentation are very data-dependent and not reproducible. We will demonstrate that the fundamental reason for this shortcoming is the subsampling introduced in the higher levels of the pyramid. We will also discuss why multi-resolution algorithms in general have a fundamental and inherent difficulty in analyzing elongated objects and ensuring connectivity. Finally, we will discuss the way in which a new pyramid algorithm using irregular tessellations, reduces some of those problems but introduces some new ones. © 1990.