Abstract: Traditional interactive image segmentation methods require users giving out background as well as foreground scribbles. Aiming at this problem, this paper proposes a novel image segmentation framework, named image segmentation with only positive and unlabeled examples. By combining learning from only positive and unlabeled examples method with graph-based semi-supervised learning technique, this method only needs users labeling a small number of pixels on interest region for segmentation. Experiments on the BJUT Eyebrow Database show that the proposed method achieves analogous results to graph-based semi-supervised learning, Random Walk as well as Lazy Snapping method, and is suitable for eyebrow recognition preprocessing.


Key words: earning from only positive and unlabeled examples, graph-based semi-supervised learning, interactively image segmentation, naive Bayes, expectation-maximization