LEARNING OPTIMAL DATA REPRESENTATION FOR CROSS-MEDIA RETRIEVAL

Cross-media retrieval is an interesting and challenging issue in content-based multimedia retrieval. Cross-media data representation, which is the fundamental problem for cross-media retrieval, is mainly discussed in this paper. First, heterogeneous low-level features are analyzed with Kernel Canonical Correlation Analysis; and then the Laplacian Space is constructed for data representation and correlation estimation; thirdly, multimodal semantic representation is calculated by solving the objective function learned from pairwise constraints. Extensive experiments have validated the proposed methods with encouraging results, and demonstrated the superiority of our method over several existing algorithms.