Relation-Guided Dual Hash Network for Unsupervised Cross-Modal Retrieval

Yuanchao Zheng, Yan Dong, Xiaowei Zhang

Abstract: Cross-modal hashing has attracted a great deal of attention due to its unique low storage cost and high retrieval efficiency. However, these existing cross-modal retrieval methods fail to deal with global semantic relational redundancy, leading to an unsatisfactory performance on such data. In this paper, to address this issue, we propose a novel cross-modal hashing, namely Relation-guided Dual Hash Network (RDHN) for unsupervised cross-modal retrieval. It captures both long-range dependencies within modalities and enhances the relevance of semantic relations between different modalities through a heterogeneous feature fusion module. Besides, we designs a dual hash network for image modality and text modality, two for each, which can effectively highlight the useful global semantic relations while suppressing the redundant information. In the cross-modal retrieval tasks, our proposed RDHN improves by 4.9\(\%\) mAP@50 for \(I \rightarrow T\) and 9.1\(\%\) mAP@50 for \(T\rightarrow I\) in 128-bit compared to the state-of-the-art AGCH on the MIRFlickr-25k dataset, respectively. Code is available at https://github.com/Z000204/RDHN.