Ordinal-Preserving Latent Graph Hashing

Current research on semantic-preserving hashing for similarity search primarily focuses on exploring semantic similarities among pointwise or pairwise samples in the visual space to generate discriminative hash codes. However, these approaches neglect the intrinsic latent features within the high-dimensional feature space, making it challenging to capture the underlying topological structure of data and resulting in suboptimal hash codes for image retrieval. This chapter introduces the Ordinal-preserving Latent Graph Hashing (OLGH) method, which formulates hash codes from the latent space and preserves the high-order locally topological structure of data in the learned hash codes. The approach introduces a triplet-constrained topology-preserving loss to unveil ordinal-inferred local features in binary representation learning, enabling the learning system to implicitly capture high-order similarities during the feature learning process. Additionally, well-designed latent subspace learning is incorporated to acquire noise-free latent features based on sparse-constrained supervised learning, fully leveraging the latent under-explored characteristics of data in subspace construction. Latent ordinal graph hashing is formulated by jointly exploiting latent space construction and ordinal graph learning, with an efficient optimization algorithm developed to achieve the optimal solution. Comprehensive experiments on diverse datasets demonstrate the effectiveness and superiority of the proposed OLGHOrdinal-preserving Latent Graph Hashing (OLGH) method compared to advanced learning-to-hash algorithms for fast image retrieval.