Image Segmentation by Bilayer Superpixel Grouping

The task of image segmentation is to group image pixels into visually meaningful objects. It has long been a challenging problem in computer vision and image processing. In this paper we address the segmentation as a super pixel grouping problem. We propose a novel graph-based segmentation framework which is able to integrate different cues from bilayer super pixels simultaneously. The key idea is that segmentation is formulated as grouping a subset of super pixels that partitions a bilayer graph over super pixels, with graph edges encoding super pixel similarity. We first construct a bipartite graph incorporating super pixel cue and long-range cue. Furthermore, mid-range cue is also incorporated in a hybrid graph model. Segmentation is solved by spectral clustering. Our approach is fully automatic, bottom-up, and unsupervised. We evaluate our proposed framework by comparing it to other generic segmentation approaches on the state-of-the-art benchmark database.