SAMM: Surroundedness and Absorption Markov Model Based Visual Saliency Detection in Images

Abstract

In this paper, we propose a saliency detection method (SAMM) by using the surroundedness and absorption Markov model. First, the approximate area of the salient object is predicted by the surroundedness to the eye fixation point prediction. Second, a simple linear iterative clustering algorithm is applied to the original image to calculate superpixels, and a two-ring image graph model is formed. We calculate two initial saliency maps $S_{1}$ and $S_{2}$ . Prior map $S_{1}$ is calculated by applying the absorption Markov chain, as the superpixel-based region of the two boundaries farthest from the predicted salient object is taken as the background region, while map $S_{2}$ is calculated by using the absorption Markov chain to detect the superpixels in the approximate region of the salient object as a foreground region. The final saliency map is obtained by combining $S_{1}$ and $S_{2}$ . Finally, a guided filter is used to reduce the background noise from the saliency map. For the evaluation, experiments are performed on six publicly available test datasets (MSRA, ECSSD, Imgsal, DUT-OMRON, PASCAL-S, and MSRA10k), and the results are compared against 10 state-of-the-art saliency detection algorithms. Our proposed saliency detection algorithm (SAMM) performs better with higher precision_recall, AUC, ${F}$ -measure, and minimum mean absolute error values.