Aggregation Enhanced Momentum Contrastive Learning for Unsupervised Cross-Modal Hashing

Unsupervised cross-modal hashing has garnered widespread attention for its support of large-scale cross-modal retrieval. However, the exploration and preservation of inter-modal and intra-modal semantic structures among multi-modal instances remain limited, while the efficiency of learning and the performance of retrieval are affected by modality imbalances and uneven data distributions. In this paper, we propose a novel unsupervised cross-modal hash learning framework, namely Aggregation Enhanced Momentum Contrastive Cross-Modal Hashing (AEMCCH). Firstly, A multi-modal adjacency graph structure is proposed to construct heterogeneous multi-modal semantics correlation, in which significantly enhances the capability of capture and utilization of both global and local information from multi-modal data. Additionally, we introduce momentum contrastive learning to address the issues of data diversity and distribution disparity. Specifically, the hash queue dictionary is constructed by aggregating momentum features obtained from momentum encoders with the designed multi-modal adjacency graph. Simultaneously, a task-specific momentum updating strategy is proposed to ensure encoding consistency among the keys within the hash queue. Sufficient experiments on three benchmark datasets demonstrate that the proposed AEMCCH outperforms existing advanced unsupervised cross-modal hashing methods.