Unsupervised cross-modal similarity via Latent Structure Discrete Hashing Factorization

Abstract

To date, large amounts of discriminative semantics-preserving discrete hash models are enjoying great popularity in cross-modal hashing community. Most of them, however, distill the shared hash codes from semantic tags, which ignores the intrinsic structure of features. Therefore, this paper proposes a direct extraction of discrete hash representation framework by factorizing latently similar structures for cross-modal retrieval in an unsupervised manner, which we refer to as Latent Structure Discrete Hashing Factorization (LSDHF). Concretely, for different modalities, assisted by the Hadamard matrix, LSDHF aligns all eigenvalues of the similarity matrix to generate a hash dictionary, and then straightly distills the shared hash codes from modalities’ intrinsic structure rather than just preserving the original geometry, so as to strengthen modal connection. In addition, a hyperbolic tangent kernel function is exploited to make the original feature more close to the hash code, thus reducing the mapping loss from the original space to the Hamming space. In the optimization phase, a discrete iterative algorithm is designed for binary optimization without introducing any intermediate variables, or utilizing relaxation strategies. The simulation results on widely used data sets demonstrate the superiority of the proposed strategy competing with state-of-the-art methods, including some supervised cross-modal hashing methods.