Color image segmentation using graph-based semi-supervised learning

A novel color image foreground/background segmentation model by semi-supervised learning is proposed. The essence is how to use the labeled pixels to achieve the whole image segmentation. Combining the color similarity between neighboring pixels and the color similarity between the unknown pixel and the known foreground/background pixels, a double-Gaussian function for the weight of graph nodes is constructed. And an adaptive parameter selection strategy and an energy model of semi-supervised segmentation are presented. The energy model is used to predict the labels of the unlabeled points by an optimization process. The experiments demonstrate the better segmentation accuracy than the competing algorithms.