FSAH: Fast Specific Average Hashing for Cross-Modal Retrieval

Supervised Hashing-based methods for cross-modal retrieval have gained much attention in recent years due to their benefits in storage and computational efficiency. Many existing methods typically learn hash code shared by multiple modalities using pairwise similarity matrix of data, while preserving semantic similarity between heterogeneous modalities in a common hamming space. Although these approaches had achieved encouraging performance, there are still several drawbacks. 1) These methods cannot effectively preserve and combine the distinguishable modality-specific feature and heterogeneous similarity. 2) The computational complexity of the model using the similarity matrix is prohibitively high, which makes these models unscalable to large-scale datasets. 3) Methods that directly project the cross-modal data into a common latent space are unreasonable since data from different modalities have diverse structures and properties. In this paper, we proposed an innovative Fast Specific Average Hashing method for cross-modal retrieval, i.e. FSAH. Extensive experiments conducted on widely used benchmark database demonstrate that the proposed methods outperform state-of-the-art cross-modal hashing methods.