Abstract: Nowadays, the multimedia retrieval technology is moving forward from single media retrieval based on similarity of content feature to multimedia integrated retrieval based on semantic correlation which can be used for cross-media retrieval.In this paper, the architecture of cross-media retrieval is studied.Based on the analysis of semantic information among various media objects with different modalities, we propose and design a cross-media search engine, its query relaxation approach and query result integration method.The experiments' results show a great improvement for increasing recall ratio and precision ratio.