EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Hub-Removal Sparse Hashing for Cross-Modal Retrieval

Zhikai Hu, Yiu-Ming Cheung, Donglin Zhang, Xin Liu

IEEE Transactions on Circuits and Systems for Video Technology · 2026-11 · 36 (11) · 17378-17389

DOI: 10.1109/tcsvt.2026.3716504 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Circuits and Systems for Video Technology, 2026-11. Verfasst von Zhikai Hu und 3 weiteren Autor:innen. Die Arbeit nennt 67 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: General Research Fund of Research Grants Council, RGC Senior Research Fellow Scheme, RGC Junior Research Fellow Scheme.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Supervised cross-modal hashing has been widely studied for efficient large-scale retrieval. Most existing methods construct a complete similarity graph from labels and preserve all pairwise relations in it as global supervision for hash code learning. However, this strategy becomes less reliable in multi-label scenarios, where multi-label samples tend to act as hub nodes that connect different semantic clusters. A small number of such hubs can serve as useful bridges and help maintain global connectivity. However, when hub nodes become numerous, hub-hub connections introduce excessive cross-cluster coupling, distort the supervision structure, and weaken local discriminability in the finite Hamming space, thus degrading retrieval performance. To address this issue, we propose a novel method called Hub-Removal Sparse Hashing (HRSH), which first identifies hub nodes and reconstructs the supervision graph by suppressing hub-hub connections. To avoid simply breaking the graph structure, a hub soft-landing strategy is further introduced to restrict each hub to a single semantic region while preserving necessary global connectivity in a more controlled manner. Extensive experiments on three benchmark datasets demonstrate that HRSH consistently outperforms existing dense and sparse baselines. Code is available at https://github.com/hutt94/HRSH

Förderung

  • General Research Fund of Research Grants Council
  • RGC Senior Research Fellow Scheme
  • RGC Junior Research Fellow Scheme
  • Seed Funding for Collaborative Research Grants of Hong Kong Baptist University
  • Guangdong and Hong Kong Universities “1+1+1” Joint Research Collaboration Scheme
  • National Science Foundation of China
  • Xiamen Science and Technology Project

Zitieren

APA

Hu, Z., Cheung, Y., Zhang, D., & Liu, X. (2026). Hub-Removal Sparse Hashing for Cross-Modal Retrieval. IEEE Transactions on Circuits and Systems for Video Technology, 36(11), 17378-17389. https://doi.org/10.1109/tcsvt.2026.3716504

Vancouver

Hu Z, Cheung Y, Zhang D, Liu X. Hub-Removal Sparse Hashing for Cross-Modal Retrieval. IEEE Transactions on Circuits and Systems for Video Technology. 2026;36(11):17378-17389. doi:10.1109/tcsvt.2026.3716504

BibTeX
@article{Hu2026,
  author = {Hu, Zhikai and Cheung, Yiu-Ming and Zhang, Donglin and Liu, Xin},
  title = {Hub-Removal Sparse Hashing for Cross-Modal Retrieval},
  journal = {IEEE Transactions on Circuits and Systems for Video Technology},
  year = {2026},
  volume = {36},
  number = {11},
  pages = {17378--17389},
  doi = {10.1109/tcsvt.2026.3716504}
}
Kostenlos starten →