COUPLED FEATURE SELECTION FOR MODALITY-DEPENDENT CROSS-MEDIA RETRIEVAL

With the explosive growth of the multi-media data, the cross media retrieval technology has drawn much attention. Previous methods usually used the 12-norm for the regularization constraint when learning the projection matrices, which can't use the informative and discriminative features to reach the better performance. In this paper, we propose the coupled feature selection model for cross-media retrieval(CFSCR) based on the modality-dependent method. In details, the proposed framework learns two couples of projection matrices for two retrieval sub-tasks(I2T and T21), and uses the l(2,1)-norm for coupled feature selection when learning the mapping matrices, which not only considers the the measure of relevance but also aims to select informative and discriminative features from image and text feature spaces. Experiment results on three different datascts demonstrate that our method performs better than the state-of-the-art methods.