Mutual Compromised Multi-feature Fusion Method for Cross-modal Hashing Retrieval

Cross-modal hashing retrieval computes similarity based on the Hamming distance among hash codes to facilitate the retrieval of multi-modal data. The primary challenge in cross-modal retrieval is how to eliminate the heterogeneous gap between different modalities. In this paper, we introduce a novel method known as Mutual Compromised Multi-Feature Fusion (MCCMR), which comprehensively combines semantic feature information and semantic structural information to address this challenge and accomplish the cross-modal retrieval task. MCCMR comprises four modules: a semantic feature guidance module, a graph attention feature fusion module, an adversarial learning feature fusion module, and a multi-task learning module. Subsequently, experiments were conducted on three cross-modal benchmark datasets to evaluate the effectiveness of our proposed method. The experimental results demonstrate that MCCMR exhibits superior performance.