上海科技大学知识管理系统

Unsupervised cross-modal hashing techniques have gained significant attention in large-scale multimedia retrieval tasks due to their ability to maintain high retrieval efficiency and minimal storage costs without requiring labeled data. However, despite the substantial progress made by existing methods, two primary issues persist: (1) Insufficient structural information in the similarity matrices and the reliance on fixed weight coefficients for generating inter-modal matrices, resulting in limited generalization capabilities; (2) Inadequate constraint capabilities in the feature fusion process, leading to suboptimal information integration. To overcome these limitations, this paper proposes a novel unsupervised cross-modal hashing method based on graph-guided deep feature fusion, named GDFH. Specifically, an adaptive similarity matrix construction module is introduced, which dynamically constructs similarity matrices through nonlinear transformations and learnable adaptive weight parameters to capture latent modal associations and semantic relationships. To further enhance feature representation, GDFH incorporates a heterogeneous graph neural network (HGNN) aimed at effectively capturing fine-grained semantic relationships between features. This approach allows for the explicit creation of more representative joint-modal features, which collaborate with existing features to facilitate the learning of hash functions. Experimental results on the MIRFlickr-25K and NUS-WIDE public datasets demonstrate that GDFH outperforms baseline methods, delivering superior performance. Specifically, for the text-to-image retrieval subtask on the MIRFlickr-25K dataset with a 64-bit hash code length, the retrieval accuracy improved by 2.81%.