Figure-ground image segmentation using genetic programming and feature selection

Yuyu Liang, Mengjie Zhang, Will N. Browne

Abstract: Figure-ground segmentation is an essential but difficult preprocessing step for many computer vision and image preprocessing tasks, such as object recognition. One challenge is to separate objects from backgrounds on images with high variations (e.g. in object shapes), which requires both effective feature sets and powerful segmentors. This paper develops a GP based segmentation method, which transforms segmentation tasks into pixel classification based problems. To control the complexity of evolved solutions, parsimony pressure is introduced in GP. Tested on two datasets with high variations (the Weizmann and Pascal datasets), the proposed method achieves similar performance in F\ score with much simpler solutions, compared with a reference GP based method that does not consider solution complexity. Moreover, it is the first time that the occurrence rates of the features used by the evolved solutions are studied to conduct feature selection for figure-ground segmentation. Compared with the whole feature set using traditional classifier based segmentation methods, the selected feature subsets can improve the segmentation performance. Moreover, analyses on the evolved solutions reveal how they function and why specific features are selected.