Asymmetric and discriminative hashing for cross-modal retrieval

Abstract

Cross-modal hashing has garnered substantial interest in the domain of information retrieval owing to its advantages in storage and computation. However, prevalent cross-modal hashing techniques frequently struggle to adequately exploit the discriminative information inherent in modalities, leading to suboptimal performance in retrieval tasks. Additionally, the symmetric factorization methods encounters tricky optimization problems and information loss when solving discrete variables. To address these limitations, we propose an Asymmetric and Discriminative Cross-modal Hashing(ADCH) method. On the one hand, ADCH incorporates a projection learning module with kernel discriminant analysis to find a discriminative common subspace. On the other hand, an asymmetric hash learning approach that preserves the pairwise label distance relations is developed to enhance the discriminative power of the learned hash functions. The extensive experiments were carried out on four benchmark datasets and the experimental results demonstrate that ADCH achieves significant improvements on retrieval precision and outperforms state-of-the-art cross-modal hashing methods.