A Fast Discrete Two-Step Learning Hashing for Scalable Cross-Modal Retrieval

Recently, some cross-modal hashing methods are proposed to search data for different modality effectively. Hashing has received wide attention because of its low storage and high efficiency. Hashing-based methods project the data instances from different modalities into a Hamming space to learn hash codes for retrieval between different modality. Although obtaining promising performance, hashing-based methods have still several common limitations. First, they learn the hash codes by constructing semantic similarity matrices, resulting in the loss of information. Second, most existing methods simultaneously learn the hash codes and the hash functions, which bring a high computational complexity. Third, they utilize the relaxation-based optimization strategy to generate the hash codes which leads to the large quantization error of the hash codes. To solve the above problems, we propose a novel fast supervised hashing method, termed Fast Discrete Two-Step Learning Hashing (FDTLH) for scalable cross-modal retrieval, which learns the discriminative hash codes by adopting a effective two-step learning scheme. Extensive experiments show that the FDTLH outperforms several state-of-the-art hashing methods in terms of retrieval performance and learning efficiency.