Article: Refining cross-modal retrieval via two stage graph hashing with latent subspace alignment Journal: International Journal of Information and Communication Technology (IJICT) 2026 Vol.27 No.91 pp.23 - 53 Abstract: Cross-modal hashing facilitates efficient retrieval across diverse modalities by mapping shared semantics into a compact binary space. Despite significant progress, conventional approaches frequently encounter two bottlenecks: 1) a preoccupation with Euclidean geometry that overlooks intrinsic graph structures, making them susceptible to data noise/outliers; 2) a reliance on a single latent subspace which fails to capture the complexity of multi-modal distributions. To circumvent these limitations, we introduce refining crossing modal retrieval via two stage graph hashing (RCRT). In this initial phase, collective matrix factorisation is integrated with a customised graph convolutional network (GCN) to extract modality-specific latent representations. By merging Euclidean features with non-Euclidean structural priors, this stage robustly mitigates the impact of anomalous data points. Subsequently, a global approximation strategy aligns these heterogeneous subspaces to distil high-level common semantics, followed by the derivation of discrete hash codes that preserve semantic consistency. In the second phase, a linear hash function is optimised via a local similarity preservation mechanism, ensuring the fine-grained topology is maintained within the Hamming space. Across multiple benchmark datasets, RCRT demonstrates superior performance over all current SOTA methods. Inderscience Publishers - linking academia, business and industry through research

Title: Refining cross-modal retrieval via two stage graph hashing with latent subspace alignment

Authors: Yuhang Fang; Chaoye Li; Tao Liao; Zhaohui Liao

Addresses: The School of Automation, Guangdong University of Technology, Guangzhou, 510006, Guangdong, China ' The School of Automation, Guangdong University of Technology, Guangzhou, 510006, Guangdong, China ' The School of Automation, Guangdong University of Technology, Guangzhou, 510006, Guangdong, China ' The School of Computer Science and Technology, Guangdong University of Technology, Guangzhou, 510006, Guangdong, China

Abstract: Cross-modal hashing facilitates efficient retrieval across diverse modalities by mapping shared semantics into a compact binary space. Despite significant progress, conventional approaches frequently encounter two bottlenecks: 1) a preoccupation with Euclidean geometry that overlooks intrinsic graph structures, making them susceptible to data noise/outliers; 2) a reliance on a single latent subspace which fails to capture the complexity of multi-modal distributions. To circumvent these limitations, we introduce refining crossing modal retrieval via two stage graph hashing (RCRT). In this initial phase, collective matrix factorisation is integrated with a customised graph convolutional network (GCN) to extract modality-specific latent representations. By merging Euclidean features with non-Euclidean structural priors, this stage robustly mitigates the impact of anomalous data points. Subsequently, a global approximation strategy aligns these heterogeneous subspaces to distil high-level common semantics, followed by the derivation of discrete hash codes that preserve semantic consistency. In the second phase, a linear hash function is optimised via a local similarity preservation mechanism, ensuring the fine-grained topology is maintained within the Hamming space. Across multiple benchmark datasets, RCRT demonstrates superior performance over all current SOTA methods.

Keywords: discrete optimisation; crossing modal retrieval hashing; graph convolutional.

DOI: 10.1504/IJICT.2026.155789

International Journal of Information and Communication Technology, 2026 Vol.27 No.91, pp.23 - 53

Received: 07 Feb 2026
Accepted: 02 Jun 2026
Published online: 13 Aug 2026 *

Title: Refining cross-modal retrieval via two stage graph hashing with latent subspace alignment

Keep up-to-date