SALIENCY DETECTION VIA LOCAL SINGLE GAUSSIAN MODEL

Saliency detection has been long researched. However, most existing algorithms can not uniformly highlight salient objects. To approach this problem, we propose a novel saliency detection algorithm based on the Local Single Gaussian Model (LSGM). First, we utilize a bottom-up model to generate an initial saliency map and construct a background dictionary and a foreground dictionary based on the initial saliency map, respectively. Then, a LSGM is used to obtain a LSGM-based map. Note that we construct a corresponding LSGM for each superpixel region and thus the LSGM is a dynamic model with geometric structure information. Finally, we integrate the LSGM-based saliency map and the initial bottom-up map with global information as the final saliency map. Extensive experiments on four public datasets show that our algorithm outperforms state-of-the-art methods.