Online Semantic Correlation Framework Enhances Efficiency in Cross-Modal Retrieval

A scalable online semantic correlation learning framework has been developed to enhance the alignment of common representations across vision and language modalities by incorporating high-level multi-label supervisory semantics directly into the correlation learning objective. Addressing a lack of principled paradigms to leverage semantic supervision in online cross-modal retrieval amid surging visual and textual data, the architecture targets the continual alignment of shared latent representations.

To facilitate the continual update of projections for the cross-modal common latent space, the framework builds upon its supervised correlation objective by executing online eigen-basis optimization. Through this design, the system maintains a computational overhead that scales linearly with the size of accumulated samples, supporting continuous adaptation as data accumulates.

Computational efficiency is further reinforced through a closed-form spectral decomposition achieved via a diagonalization strategy. This formulation enables efficient projection updates while eliminating the iterative convergence thresholds and stopping criteria typically required in optimization routines.

In experimental evaluations conducted across four image-text benchmarks, the proposed framework demonstrated competitive retrieval accuracy and practical scalability within the evaluated range. In addition, the source code and data have been made available via an online repository at www.github.com/lia-tao/OSCL.

Subject of Research: Cross-modal retrieval and online semantic correlation learning

Article Title: Online semantic correlation learning for cross-modal retrieval

Article References: Tao, L., Shu, X., Zong, H., Jiang, R., Zhang, S., & Wang, Y. (2026). Online semantic correlation learning for cross-modal retrieval. Applied Intelligence, 56(16), Article 539. https://doi.org/10.1007/s10489-026-07515-4

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07515-4

Keywords: cross-modal retrieval, semantic correlation learning, multi-label supervision, eigen-basis optimization, spectral decomposition, diagonalization strategy, common latent space, image-text benchmarks, Online, semantic, correlation, learning