Sparse Manifold Embedded Hashing for Multimedia Retrieval

Hashing has become more and more attractive in the large-scale multimedia retrieval community, due to its fast search speed and low storage cost. Most hashing methods focus on finding the inherent data structure, and neglect the sparse reconstruction relationship. Besides, most of them adopt a two-step solution for the structure embedding and the hash codes learning, which may yield suboptimal results. To address these issues, in this paper, we present a novel sparsity-based hashing method, namely, Sparse Manifold embedded hASHing, SMASH for short. It employs the sparse representation technique to extract the implicit structure in the data, and preserves the structure by minimizing the reconstruction error and the quantization loss with constraints to satisfy the independence and balance of the hash codes. An alternative algorithm is devised to solve the optimization problem in SMASH. Based on it, SMASH learns the hash codes and the hash functions simultaneously. Extensive experiments on several benchmark datasets demonstrate that SMASH outperforms some state-of-the-art hashing methods for the multimedia retrieval task.