FINDING OPTIMIZED OBJECT ALPHABET USING GA

摘要

Some researches show that people recognize some types of object as the combination of neurons which ignite on small figures piece by piece. That mechanism is called "Figure Alphabet hypothesis", and those small figures are called "Object Alphabet". Object Alphabet is considered as simple figures (e.g. circle, triangle, and so on). However, it is not clearly defined for the shape and the numbers of Object Alphabet. This research, therefore, tries to find optimized Object Alphabet for the study images by Genetic Algorithm (GA) as a first step of modeling the mechanism that human recognizes the figure. We assume that Object Alphabet is image constructed by the dot patterns of n * n. However, the number of those patterns is huge, so we utilize GA for solving the speed issue of full search. We calculate the difference rate between original images and those dot patterns, and we use those difference rates as amount of features for classification. We apply our method to a lot of fonts and hand-written characters. As a result, we obtain simple and height classification accuracy dot patterns.