Content-based image retrieval plays an important role in the management of a large image database. However, the results of the state-of-the-art image retrieval approaches are not as satisfactory for the well-known gap between visual features and semantic concepts. Therefore, a novel scheme is proposed, consisting of three major components: prefiltering processing, relevance score computation, and candidate ranking refinement. First, to tackle the problem of the large computation cost involved in a large image database, a prefiltering process is utilized to filter out the most irrelevant images while keeping the most relevant images according to the results of the manifold-ranking algorithm. Second, the relevance between the query image and the remaining images is measured based on probability density estimation, and the obtained relevance scores are stored for a later refinement process. Finally, a transductive model, a random walk with a restart algorithm, is used to refine candidate ranking by taking into account both the pairwise information of unlabeled samples and the relevance scores between the input query sample and unlabeled samples. Experiments conducted on a typical Corel data set demonstrate the effectiveness of the proposed scheme.