Abstract: Aiming at the problem that the existing cross modal Has h retrieval method could not effectively eliminate the semantic differences betw een the different modal data, we proposed a new retrieval method based on spar se coding Hash. The semantic differences between the lowlevel visual featur es and highlevel semantics were solved, and the effect of cross modal retr ieval was improved. Cross modal similarity retrieval by using sparse coding. Firstly, we used sparse coding to obtain salient features and implicit concepts of images and texts. Secondly, we mapped the latent semantic f eatures of learning to a common abstract space, and then we found correlation between features of the multimodal data by the iterative mechanism . Finally, we obtained the uniform Hash coding by the quantization of highleve l semantic abstract space.

Key words: latent concept; multimedia data, cross modal retrieval, sparse coding