LEARNING UNIFIED SPARSE REPRESENTATIONS FOR MULTI-MODAL DATA

Cross-modal retrieval has become one of interesting and important research problem recently, where users can take one modality of data (e.g., text, image or video) as the query to retrieve relevant data of another modality. In this paper, we present a Multi-modal Unified Representation Learning (MURL) algorithm for cross-modal retrieval, which learns unified sparse representations for multi-modal data representing the same semantics via joint dictionary learning. The l(1)-norm is imposed on the unified representations to explicitly encourage sparsity, which makes our algorithm more robust. Furthermore, a constraint regularization term is imposed to force the representations to be similar if their corresponding multi-modal data have must-links or to be far apart if their corresponding multi-modal data have cannot-links. An iterative algorithm is also proposed to solve the objective function. The effectiveness of the proposed method is verified by extensive results on two real-world datasets.