Spatiality-Aligned Graph Convolution Hashing for Unsupervised Cross-Modal Retrieval

Unsupervised cross-modal retrieval leveraging hash learning has attracted a lot of interest from academics, primarily due to its minimal storage requirements, swift retrieval speeds, and label-free nature, making it a prime candidate for real-world applications. Nevertheless, the inherent "heterogeneous differences" and "semantic gap" within cross-modal data features pose significant challenges. Despite various attempts by scholars to address these issues, the existing solutions often fall short in terms of effectiveness. This can be attributed to several reasons: Firstly, image features, in contrast to text features, possess a plethora of attributes, making it challenging to extract features that possess strong expressive capabilities. Secondly, each modality of data exhibits not only inter-node correlations but also intricate relationships between individual nodes and their neighboring nodes. Consequently, a straightforward semantic correlation matrix is inadequate in capturing the intricate structural relationships within the data. Thirdly, the neglect of internal structural associations results in insufficient representation capabilities of binary hash codes in the Hamming space, hampering the learning of effective hash functions. To address these limitations, we introduce a Spatiality-Aligned Graph Convolution Hashing (SAGCH) algorithm. Extensive experiments demonstrate that our SAGCH method excels in both "image-to-text retrieval" and "text-to-image retrieval" tasks.