The performance of salient object detection (SOD) has been significantly advanced by using deep convolutional networks. However, it largely depends on the high cost of pixel-level annotations. To reduce human effort while improving the prediction accuracy, we propose a novel two-phase learning framework. The weakly supervised information in terms of scribbles is provided as initial labels. Then, as the first phase, high-quality pseudo-labels are generated by mapping scribbles onto object/object-part contours. These contour maps are predicted by the hierarchical contour detection algorithm, providing superior accuracy and smoothness. In the second phase, a deep neural network is alternately trained and predicted. The pseudo-labels are refined in an iterated process, where a conditional random field (CRF) model and a filter module are designed to promote the performance. Extensive experiments on five benchmarks show that our framework can achieve comparable results with the state-of-the-art fully and weakly supervised methods.