Cross-modal image-text search via Efficient Discrete Class Alignment Hashing

Hashing has produced enormous potentials in cross-modal image-text search, which learns compact binary codes by exploring the correlations between distinct modalities. However, there still exist some limitations. First, most existing methods neglect the relation between the data characteristics and supervised information. Second, a relaxation strategy results in large quantization errors. Third, constructing large n x n (a.k.a. training size) similarity graphs increases computational load. To address these issues, we propose a novel discrete supervised hashing method, termed Efficient Discrete Class Alignment Hashing (EDCAH), which integrates class alignment and matrix factorization for hashing learning. Specifically, it exploits the semantic consistency of data instances and informative labels to simultaneously learn the hash codes and hash functions. Meanwhile, a discrete optimization strategy is developed to solve the EDCAH, which is beneficial to generate high-quality hash codes. Furthermore, to improve the learning efficiency of EDCAH, we propose a fast and efficient variant dubbed EDCAH-t that utilizes a two-step hashing strategy. Extensive experiments demonstrate the superiority of EDCAH and EDCAH-t in both search accuracy and learning efficiency.