The graph-based manifold ranking has achieved great success in universal saliency detection. However, it fails to suppress backgrounds when salient objects are surrounded by complex backgrounds. Furthermore, it tends to lose small salient objects. To overcome these problems, we propose an improved saliency detection method by fusing mid-level features and by adaptive updated graph. First, we use simple linear iterative cluster algorithm to segment an image into superpixels, then an initial graph is constructed and saliency values are calculated by the traditional manifold ranking. Second, we take the above saliency values as mid-level features and fuse them with low-level color features into edge weights to update the graph model. Last but not least, soft foreground queries are exploited upon the new graph to compute final saliency values. Extensive evaluations on several challenging datasets reveal that the proposed method not only suppresses backgrounds better, but also achieves higher precision, recall and F-measure than other state-of-the-arts.