EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Language-Guided Mixture of Spectral Experts for Cross-Scenario Hyperspectral Anomaly Detection

Yimin Fu, Peiyuan Ma, Yinghao Xu, Michael K. Ng, Peng Ren

IEEE Transactions on Circuits and Systems for Video Technology · 2026-11 · 36 (11) · 16502-16514

DOI: 10.1109/tcsvt.2026.3719083 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Circuits and Systems for Video Technology, 2026-11. Verfasst von Yimin Fu und 4 weiteren Autor:innen. Die Arbeit wurde bisher 2-mal zitiert (Crossref) und nennt 73 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Department of Science and Technology of Guangdong Province (GDSTC): Guangdong and Hong Kong Universities “1+1+1” Joint Research Collaboration Scheme, National Key Research and Development Program of China, Research Grants Council of Hong Kong (RGC) General Research Fund.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Hyperspectral anomaly detection (HAD) aims to identify atypical targets in hyperspectral images (HSIs) without requiring prior knowledge of anomaly characteristics. Existing HAD methods typically distinguish anomalies from background patterns in a self-supervised manner, where reconstruction errors are used as the anomaly indicator. However, their practical deployment is severely constrained by the reliance on the one-for-one detection scheme, which necessitates scenario-specific retraining for inference. Otherwise, the detection performance can severely degrade due to distribution shifts between training and testing scenarios. To address these limitations, we propose language-guided mixture of spectral experts (LG-MoSE) for cross-scenario hyperspectral anomaly detection. Specifically, LG-MoSE leverages language priors to provide comprehensive spectral–spatial characterizations of background components, thereby preventing the reconstruction from collapsing into identical shortcut solutions based solely on spatial representations of single-image inputs. During encoding, diverse spectral representations corresponding to different spectral subspaces are adaptively aggregated under the guidance of scene-level textual descriptions. Then, the aggregated spectral representations are combined with spatial representations to enhance the understanding of the underlying background pattern. Moreover, a transposable state space block is embedded into each decoding stage to mitigate the over-dependence on specific structural regularities. Extensive experiments are conducted on eight commonly used HAD datasets, and the proposed method consistently achieves state-of-the-art performance across various scenarios and settings. The code will be released at https://github.com/fuyimin96/LG-MoSE upon acceptance.

Förderung

  • Department of Science and Technology of Guangdong Province (GDSTC): Guangdong and Hong Kong Universities “1+1+1” Joint Research Collaboration Scheme
  • National Key Research and Development Program of China
  • Research Grants Council of Hong Kong (RGC) General Research Fund

Zitieren

APA

Fu, Y., Ma, P., Xu, Y., Ng, M. K., & Ren, P. (2026). Language-Guided Mixture of Spectral Experts for Cross-Scenario Hyperspectral Anomaly Detection. IEEE Transactions on Circuits and Systems for Video Technology, 36(11), 16502-16514. https://doi.org/10.1109/tcsvt.2026.3719083

Vancouver

Fu Y, Ma P, Xu Y, Ng MK, Ren P. Language-Guided Mixture of Spectral Experts for Cross-Scenario Hyperspectral Anomaly Detection. IEEE Transactions on Circuits and Systems for Video Technology. 2026;36(11):16502-16514. doi:10.1109/tcsvt.2026.3719083

BibTeX
@article{Fu2026,
  author = {Fu, Yimin and Ma, Peiyuan and Xu, Yinghao and Ng, Michael K. and Ren, Peng},
  title = {Language-Guided Mixture of Spectral Experts for Cross-Scenario Hyperspectral Anomaly Detection},
  journal = {IEEE Transactions on Circuits and Systems for Video Technology},
  year = {2026},
  volume = {36},
  number = {11},
  pages = {16502--16514},
  doi = {10.1109/tcsvt.2026.3719083}
}
Kostenlos starten →