Wrapper Feature Construction for Figure-Ground Image Segmentation Using Genetic Programming

Figure-ground segmentation is a process of separating regions of interest from unimportant backgrounds. It is challenging to separate objects from target images with high variations (e.g. cluttered backgrounds), which requires effective feature sets to capture the discriminative information between object and background regions. Feature construction is a process of transforming a given set of features to a new set of high-level features, which considers the interactions between the previous features, thus the constructed features can be more meaningful and effective. As Genetic programming (GP) is a well-suited algorithm for feature construction (FC), it is employed to conduct both multiple FC (MFC) and single FC (SFC), which aims to improve the segmentation performance for the first time in this paper. The cooperative coevolution technique is introduced in GP to construct multiple features from different types of image features separately while conducting feature combination simultaneously, called as CoevoGPMFC. One wrapper method (wrapperGPSFC) is also designed, and one well-performing embedded method (embeddedGPSFC) is introduced as a reference method. Compared with the original features extracted by existing feature descriptors, the constructed features from the proposed methods are more robust and performance better on the test set. Moreover, the features constructed by the three methods achieve similar performance for the given segmentation tasks.