Distance variety preserving hashing for large‐scale retrieval

Summary Hashing techniques have been seen huge adoption in large‐scale retrieval owing to the computational and storage efficiencies of binary codes. However, the binary codes might not preserve the semantic similarity powerfully and lead to poor accuracy. To evade binary limitation but remain hashing superiority, this paper introduces a novel asymmetric hashing framework for distance variety preserving, which yields preferable performance via real‐valued query and multi‐valued database with shorter code length. Concretely, a binary sparse representation is proposed for multivalued database encoding to make computation and storage complexity tolerable. By exploiting the supervised information, we learn a projective matrix that project data onto a sub‐space in which the similarity between realvalued query and multi‐valued database can be measured. Due to the discrete constraints, for different discrete variables, we present different meticulous discrete solutions for their sub‐problems. Exhaustive experiments on three multilabel datasets demonstrate our approach outperforms most state‐of‐the‐art hashing methods.