A Genetic Programming Approach for Image Segmentation

This work presents a methodology for using genetic programming (GP) for image segmentation. The image segmentation process is seen as a classification problem where some regions of an image are labeled as foreground (object of interest) or background. GP uses a set of terminals and nonterminals, composed by algebraic operations and convolution filters. A function fitness is defined as the difference between the desired segmented image and that obtained by the application of the mask evolved by GP. A penalty term is used to decrease the number of nodes of the tree, minimally affecting the quality of solutions. The proposed approach was applied to five sets of images, each one with different features and objects of interest. Results show that GP was able to evolve solutions of high quality for the problem. Thanks to the penalty term of the fitness function, the solutions found are simple enough to be used and understood by a human user.