Semantic-Enhanced Graph Regularization Hashing for Large-Scale Cross-Modal Retrieval

Abstract In the area of large-scale data, cross-modal hashing has attracted significant attention for fast similarity search task due to its low storage requirements and fast retrieval speed. Existing supervised cross-modal hashing learning methods have made extensive attempts to fully exploit supervisory information. Different from previous work, we propose a method named S emantic- E nhanced G raph R egularization H ashing ( SEGRH ), which contributes a new and novel exploration to this research domain. Specifically, our method consists of two core components: contrastive cross-modal fusion and robust semantic preservation. We adopt the pre-trained CLIP model as the backbone to extract modality-specific features, which are subsequently fused via a multimodal Transformer. Robust semantic preservation is designed to mine and integrate high-quality semantic information. SEGRH first leverages positive pointwise mutual information to convert label information into semantic association metrics. Second, the graph Laplacian is incorporated to absorb such semantic cues and further constrain the structure of the hash space. Our SEGRH is capable of performing high-quality multimodal learning and possesses strong semantic consistency modeling capabilities, thereby achieving satisfactory generalization performance. Experimental results on three standard datasets demonstrate that our approach significantly outperforms previous state-of-the-art methods.