Fast image segmentation based on multilevel banded closed-form method

Abstract

In this paper, a fast interactive image segmentation method is developed. The method combines the GrabCut algorithm with the multilevel banded closed-form (MLBCF) technique to achieve the acceleration. The GrabCut method is first applied on a low-resolution image to obtain the segmentation. The coarse labeling is then propagated to the higher-resolution level by using the banded closed-form method with the locally linear assumption. Some post-processing, such as alpha thresholding, probability classification and multi-seeds banded Graph Cuts, is applied to assign the final labeling. For experimental comparison, we also implement the multilevel banded Graph Cuts (MLBGC) method based on GrabCut algorithm. Experiments using synthesized noisy images and real natural scene images demonstrate the superior performance of the proposed method in terms of segmentation accuracy, computation efficiency and memory usage.