Edge Detection of Petrographic Images Using Genetic Programming

摘要

This paper discusses work in progress that uses genetic programming to evolve edge detectors for petrographic images. Microscopic images of thin sections from mineral samples are obtained using a rotating polarizer microscope. These images are then processed using a number of filters, resulting in a set of nine filtered image parameters. In order to be useful for higher-level analysis, such as automatic mineral identification, the grain boundaries within these images must be identified. Using genetic programming, edge detecting functions are evolved for this purpose. The edge detectors may use as any of the filtered image parameters as input. Since the source images are large, a subset of the images is sampled for training, and the remainder of the image is used for testing. This training data is selected with a biased random sampling strategy. The complexity of the images dictates that a generic edge detector for all mineral specimens is infeasible. Rather, the most useful edge detectors will be those that are specialized for particular families of mineral specimens.