Deep Consistency Preserving Network for Unsupervised Cross-Modal Hashing

Given the proliferation of multimodal data in search engines and social networks, unsupervised cross-modal hashing has gained traction for its low storage consumption and fast retrieval speed. Despite the great success achieved, unsupervised cross-modal hashing still suffers from lacking reliable similarity supervision and struggles with reducing information loss caused by quantization. In this paper, we propose a novel deep consistency preserving network (DCPN) for unsupervised cross-modal hashing, which sufficiently utilizes the semantic information in different modalities. Specifically, we gain consistent features to fully exploit the co-occurrence information and alleviate the heterogeneity between different modalities. Then, a fusion similarity matrix construction method is proposed to capture the semantic relationship between instances. Finally, a fusion hash code reconstruction strategy is designed to fit the gap between different modalities and reduce the quantization error. Experimental results demonstrate the effectiveness of the proposed DCPN on unsupervised cross-modal retrieval tasks.