Multi-objective Genetic Programming for Figure-Ground Image Segmentation

摘要

Figure-ground segmentation is a crucial preprocessing step in areas of computer vision and image processing. As an evolutionary computation technique, genetic programming (GP) can evolve algorithms automatically for complex problems and has been introduced for image segmentation. However, GP-based methods face a challenge to control the complexity of evolved solutions. In this paper, we develop a novel exponential function to measure the solution complexity. This complexity measure is utilized as a fitness evaluation measure in GP in two ways: one method is to combine it with the classification accuracy linearly to form a weighted sum fitness function; the other is to treat them separately as two objectives. Based on this, we propose a weighted sum GP method and a multi-objective GP (MOGP) method for segmentation tasks. We select four types of test images from bitmap, Brodatz texture, Weizmann and PASCAL databases. The proposed methods are compared with a reference GP method, which is single-objective (the classification accuracy) without considering the solution complexity. The results show that the new approaches, especially MOGP, can significantly reduce the solution complexity and the training time without decreasing the segmentation performance.