Owing to the problems of the absorbing Markov random walk method failing to fully suppress the central background area of the saliency map and losing parts of salient objects near the image boundary, an image saliency detection method based on manifold regularized random walk is proposed. First, a global graph with superpixels from the input image is constructed. An initial saliency map is obtained by using the absorbing Markov chain, and then an adaptive threshold is used to segment the initial saliency map to get robust foreground queries. Second, in order to make effective use of the complementarity of global information and local information, an optimal affinity matrix is obtained by constructing the local regular graph. Finally, the obtained optimal affinity matrix and foreground queries are applied in the manifold regularized framework to obtain the final saliency results. Experimental verifications are carried out on three public datasets. The results show that the precision and recall rate of saliency detection have been improved by the proposed method.