A Novel Saliency-based Object Segmentation Method for Seriously Degenerated Images

Automatically segmenting the salient object based on the saliency information frequently fails on the non-uniform motion blurred images. We propose a novel saliency-based object segmentation method with a self-expansion mechanism to deal with this problem in this paper. Firstly, to improve the initial localization accuracy for expansion, we integrate a modified local autocorrelation congruency into an initial salient object seed for building a combined salient object seed. Secondly, we present a novel method named Normal Expansion to expand the obtained salient object seed to the real boundaries of the target object. At last, we design a strategy based on superpixels to repair the lost degenerated regions. Based on the proposed method, we can more precisely segment the partially motion blurred object boundaries from a uniformly motion blurred background. Our experimental results show that our method outperforms some state-of-the-art saliency-based object segmentation approaches both quantitatively and qualitatively.