A new salient region detection algorithm is proposed via KL divergence between color probability distributions of super-pixels and merging multi-scale saliency maps. Firstly, multi-scale super-pixel segmentations of an input image are computed. In each segmentation scale, an undirected close-loop connected graph is constructed, in which nodes are the super-pixels and the adjacent regions are expanded reasonably relying on the total number of super-pixels. Then, all the color values in each super-pixel are clustered in terms of their discriminative power to get the statistical probability distribution of the cluster labels for each super-pixel. Next, the edges between all adjacent super-pixel pairs are weighted with the harmonic-mean of KL divergence of their probability distributions, and then the multi-scale saliency maps are calculated according to boundary connectivity and region contrast. The final saliency map is obtained by calculating and optimizing the mean map of all the saliency maps with different scales. Experimental results on some large benchmark datasets demonstrate that the proposed algorithm outperforms some state-of-the-art methods, and has higher precision and recall rates. The proposed algorithm can also produce smooth saliency maps.