Feature Construction Using Genetic Programming for Figure-Ground Image Segmentation

Abstract

Figure-ground segmentation is a process of separating regions of interest from unimportant backgrounds. For complex images (e.g. images with high variations), feature construction (FC) is necessary, which can produce high-level features, to help achieve accurate segmentation performance. Genetic programming (GP) is considered as a well-suited FC technique, which is employed for the first time to build FC methods that aim to improve the segmentation performance in this paper. One filter GP method (FGP), in which a novel entropy based fitness function is developed, and one embedded GP method (EGP), in which the error rate is used as the fitness function, are proposed. The single constructed feature and the combined features (the constructed feature + original features) are tested on two standard image datasets with high variations, i.e. Weizmann and Pascal datasets. Compared with the original features extracted by existing feature descriptors, both methods can construct useful features from the original ones with the combined features improving the segmentation performance on both datasets generally. Moreover, EGP is more efficient and perform better than FGP.