A Supervised Figure-Ground Segmentation Method Using Genetic Programming

Figure-ground segmentation is an important preprocessing phase in many computer vision applications. As different classes of objects require specific segmentation rules, supervised (or top-down) methods, which learn from prior knowledge of objects, are suitable for figure-ground segmentation. However, existing top-down methods, such as model-based and fragment-based ones, involve a lot of human work. As genetic programming (GP) can evolve computer programs to solve problems automatically, it requires less human work. Moreover, since GP contains little human bias, it is possible for GP-evolved methods to obtain better results than human constructed approaches. This paper develops a supervised GP-based segmentation system. Three kinds of simple features, including raw pixel values, six dimension and eleven dimension grayscale statistics, are employed to evolve image segmentors. The evolved segmentors are tested on images from four databases with increasing difficulty, and results are compared with four conventional techniques including thresholding, region growing, clustering, and active contour models. The results show that GP-evolved segmentors perform better than the four traditional methods with consistently good results on both simple and complex images.