Figure-ground image segmentation via semantic information

—Image segmentation is a technique to segment the image into particular regions which are consistent in some specific characteristics. It is still a non-trivial task to segment the foreground from the background completely, as these regions are seldom uniform in color or texture which are defined as commonly-used low-level features adopted in the traditional techniques. In this work we present a new segmentation framework by employing the semantic information, which is classified as the high-level visual features and holds important clue to indicate the foreground from the background. Firstly, the image is segmented hierarchically by using the semantic edge constraint, and the result is represented as an Ultrametric Contour Map (UCM). Then the UCM is employed in the segmentation energy function as the data term and smoothness term. Finally, the optimal segmentation labels are obtained via the Graph-cut. Experiments achieve good performance testing on the MSRC 21 dataset.