Supervised online multi-modal discrete hashing

Abstract

Multi-modal hashing has been proposed due to its flexibility and effectiveness in multimedia retrieval tasks. Nevertheless, the majority of multi-modal hashing methods now in use acquire hash codes and hash functions through batch-based learning, which is unsuitable to handle streaming data. Online learning can be used for multi-modal hashing, but still exists in some issues that need to be addressed, such as how to properly employ the modal semantic information and reduce hash learning loss. To address these issues mentioned above, we propose a multi-modal hashing method, called Supervised Online Multi-modal Discrete Hashing (SOMDH). SOMDH first imposes a multi-modal weight to obtain the integrated multi-modal feature representation and then leverages matrix factorization to directly obtain hash codes. In addition, the correlations between the new data and existing data are established with a similarity matrix. Finally, SOMDH can learn the hash codes by discrete optimization strategy. Experimental results on two benchmark datasets demonstrate that SOMDH outperforms state-of-the-art offline and online multi-modal hashing methods in terms of retrieval accuracy.