Hadamard matrix-guided multi-modal hashing for multi-modal retrieval

Abstract

Multi-modal hashing can encode heterogeneous multi-modal data into compact binary codes, which has been extensively studied to solve large-scale multi-modal retrieval. However, since pioneer methods do not exploit fully the potential discriminative information in category labels and the rich complementary information between multi-modal data, their retrieval performance is limited. To address this problem, we propose a novel multi-modal hashing method that performs subspace learning and target feature learning in an overall framework. On the one hand, the proposed method captures the complementary information between multi-modal data by adaptive projection learning. To enhance the feature representation ability, the multi-modal spaces are reconstructed via the collective matrix factorization. On the other hand, the target binary codes that are predefined by the Hadamard matrix are softened into the learnable target features, which can promote the inter-class separability and preserve the intra-class difference. The extensive experiment results conducted on three public datasets show that the proposed method outperforms state-of-the-art methods.