Background prior and label propagation have been widely advocated for salient region detection. However, traditional background prior based models heuristically assume that all or parts of the pixels on the image boundary are background. And the label propagation based models only consider the pairwise smoothness in optimization. To tackle these two shortcomings, we propose a framework which utilizes background prior and label propagation to generate more reliable saliency maps. Firstly, a novel optimal seeds estimation strategy is proposed to adaptively and robustly choose the most informative seeds from refined background map and foreground prior. Then, a new label propagation model which takes into account both the pairwise and local smoothness constraint is proposed to learn the saliency score according to the estimated background and foreground seeds. Last but not least, we present a new application of salient region detection named attention driven image abstraction. Both quantitative and qualitative evaluations on three widely used datasets demonstrate the superiority of the proposed method to other several state-of-the-art methods.