Multilevel Deep Semantic Feature Asymmetric Network for Cross-Modal Hashing Retrieval

Cross-modal hash retrieval has been widely applied due to its efficiency and low storage overhead. In the domain of supervised cross-modal hash retrieval, existing methods exhibit limitations in refining data features, leading to insufficiently detailed semantic information extraction and inaccurate reflection of data similarity. The challenge lies in utilizing multi-level deep semantic features of the data to generate more refined hash representations, thereby reducing the semantic gap and heterogeneity caused by different modalities. To address this challenging problem, we propose a multilevel deep semantic feature asymmetric network structure (MDSAN). Firstly, this architecture explores the multilevel deep features of the data, generating more accurate hash representations under richer supervised information guidance. Secondly, we investigate the preservation of asymmetric similarity within and between different modalities, allowing for a more comprehensive utilization of the multilevel deep features to bridge the gap among diverse modal data. Our network architecture effectively enhances model accuracy and robustness. Extensive experiments on three datasets validate the significant improvement advantages of the MDSAN model structure compared to current methods.