Comparison study of controlling bloat model of GP in constructing filter for cell image segmentation problems

The final goal of this research is to construct a cell image analysis system for supporting corneal regenerative medicine. Existing image analysis software requires knowledge about image processing of users because users have to combine several image processing on its analysis. Therefore, several types of methods to construct the objective image processing automatically using genetic programming (GP) have been proposed. However, in conventional researches, only canonical GP models were utilized. In this paper, GP models suited to cell image segmentation are investigated applying proposed controlling bloat model of GP. Applied models were six types in addition to the canonical model; those are Double Tournament, Tarpeian, Non-Destructive Crossover (NDC), Recombinative Hill-Climbing (RHC), Spatial Structure + Elitism (SS+E). The combination of image processing obtained by these GP models and the robustness are examined by comparative experiments, using corned endothelium cell image. The experiment results showed that SS+E is superior to other models in both robustness and image processing constructed for cell image segmentation, without depending on parameters of tree depth limit and penalty.