Multi-head Hashing with Orthogonal Decomposition for Cross-modal Retrieval

Recently, cross-modal hashing has become a promising line of research in cross-modal retrieval. It not only takes advantage of complementary multiple heterogeneous data modalities for improved retrieval accuracy, but also enjoys reduced memory footprint and fast query speed due to efficient binary feature embedding. With the boom of deep learning, convolutional neural network (CNN) has become the de facto method for advanced cross-model hashing algorithm. Recent research demonstrates that dominant role of CNN is challenged by increasingly effective Transformer architectures due to their advantages of long-range modeling by relaxing local inductive bias. However, the absence of inductive bias shatters the inherent geometric structure, which inevitably leads to compromised neighborhood correlation. To alleviate this problem, in this paper, we propose a novel cross-modal hashing method termed Multi-head Hashing with Orthogonal Decomposition (MHOD) for cross-modal retrieval. More specifically, with the multi-modal Transformers used as the backbones, MHOD leverages orthogonal decomposition for decoupling local cues and global features, and further captures their intrinsic correlations through our designed multi-head hash layer. In this way, the global and local representations are simultaneously embedded into the resulting binary code, leading to a comprehensive and robust representation. Extensive experiments on popular cross-modal retrieval benchmarking datasets demonstrate the proposed MHOD method achieves advantageous performance against the other state-of-the-art cross-modal hashing approaches.