Automatic generation of morphological set recognition algorithms.

A system for automatically generating simple morphological set-recognition algorithms is described and implemented. The program accepts a group of binary or grey level images, along with two "truth" masks for each, indicating the pixels to be accepted and those to be rejected. It then searches an image operator/parameter space for possible solutions, using domain knowledge about the uses and effects of individual operators to reduce the search. Solutions are presented in terms of an image algebraic target language. Currently the system can solve problems which have perfect of nearly perfect solutions in a few operator steps from a repertoire of several thous and operator/parameter combinations. The system supports roughly a dozen different operator groups drawn primarily from mathematical morphology (e.g., erosion, dilation, opening, closing, etc.), covering such image characteristics as absolute position, distance from the foreground or background, size class of pixels, and width, roughness and length of connected particles and holes. The system is shown to be equivalent to an exhaustive search, in that it cannot miss any solutions which fall within its defined scope, yet efficient, in that it accomplishes its task significantly faster than would an exhaustive search technique. Research into this area offers the potential to enhance our formal underst and ing of image operators and when to use them, to provide better, more powerful tools for testing and evaluation of image algorithms by algorithm developers, and to partially or even completely automate some algorithm development tasks via systems that can intelligently manage the search and testing process. Building such systems also provides a way to model the algorithm development process as carried out by humans, which can improve our underst and ing of how to make such efforts more efficient.PhDComputer scienceUniversity of Michiganhttp://deepblue.lib.umich.edu/bitstream/2027.42/161878/1/8813015.pd