Extracting nodes that reflect image content and assigning initial labels for these nodes are two critical technologies for saliency detection. A novel method of saliency detection is proposed by this work. It consists of two main parts, one is self organizing map (SOM), and the other is manifold learning (ML). Hundreds of nodes are obtained by the SOM. These nodes can capture not only the color, but also the contour of image content. By means of embedding a two dimension map into higher Euclid space, a weighted undirected graph is constructed. In consideration of edge symmetry in undirected graph, a manifold learning method, which combines undirected graph and semi-supervision, is further proposed. With supplied initial saliency values for nodes along image borders, the saliency values are computed for all nodes. Experimental results demonstrate the proposed model not only achieves high performance on precision and recall, but also presents a pleasing visual effect.