EUVIMEDEUVIMED entdecken
Crossref

Universal Context-Modulated Asymmetric Multiscale Network for Camouflaged Object Segmentation

Xuedong Guo, Lei Deng, Maoyong Li, Yingying Gao, Mingli Dong, Lianqing Zhu

IEEE Transactions on Circuits and Systems for Video Technology · 2026-11 · 36 (11) · 17003-17017

DOI: 10.1109/tcsvt.2026.3718650 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Circuits and Systems for Video Technology, 2026-11. Verfasst von Xuedong Guo und 5 weiteren Autor:innen. Die Arbeit nennt 48 Literaturangaben. Gefördert von: Research Project of Beijing Municipal Natural Science Foundation.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Camouflaged Object Segmentation (COS) aims to extract hidden objects from highly blended backgrounds, necessitating the resolution of critical challenges such as extreme object-background similarity, irregular shapes, and blurred boundaries. Current methods often struggle to dynamically suppress background noise during multi-scale feature interactions and fail to capture complex geometric structures precisely. To overcome these limitations, this paper proposes a Universal Contextual-Modulation Asymmetric Multi-scale Network (UCAMNet). Specifically, we first design a Multi-scale Context Refinement Module (MCRM) to extract asymmetric features and enhance local details. Then, a Cross-layer Context Fusion Module (CCFM) is proposed to integrate adjacent layer information, bridging the semantic gap. Furthermore, a Wide-Context Reconstruction Module (WCRM) is introduced to capture global large-receptive-field information via asymmetric dilated attention for precise layer-by-layer object reconstruction. Subsequently, a Coordinate-Aware Prediction Module (CAPM) provides geometric constraints to generate crisp boundaries. Finally, a Context-Aware Feature Modulation (CAFM) mechanism is specifically designed and embedded into the MCRM, CCFM, and WCRM. By adaptively filtering background interference and enhancing core contexts, it comprehensively ensures the high-precision restoration of camouflaged regions throughout the entire process of feature extraction, fusion, and reconstruction. Extensive experiments on four benchmark datasets (CAMO, COD10K, CPD, and VIAC) demonstrate that UCAMNet achieves a remarkable balance between accuracy and model efficiency, significantly outperforming existing state-of-the-art methods. The source code is available at https://github.com/change68/UCAMNet

Förderung

  • Research Project of Beijing Municipal Natural Science Foundation

Zitieren

APA

Guo, X., Deng, L., Li, M., Gao, Y., Dong, M., & Zhu, L. (2026). Universal Context-Modulated Asymmetric Multiscale Network for Camouflaged Object Segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 36(11), 17003-17017. https://doi.org/10.1109/tcsvt.2026.3718650

Vancouver

Guo X, Deng L, Li M, Gao Y, Dong M, Zhu L. Universal Context-Modulated Asymmetric Multiscale Network for Camouflaged Object Segmentation. IEEE Transactions on Circuits and Systems for Video Technology. 2026;36(11):17003-17017. doi:10.1109/tcsvt.2026.3718650

BibTeX
@article{Guo2026,
  author = {Guo, Xuedong and Deng, Lei and Li, Maoyong and Gao, Yingying and Dong, Mingli and Zhu, Lianqing},
  title = {Universal Context-Modulated Asymmetric Multiscale Network for Camouflaged Object Segmentation},
  journal = {IEEE Transactions on Circuits and Systems for Video Technology},
  year = {2026},
  volume = {36},
  number = {11},
  pages = {17003--17017},
  doi = {10.1109/tcsvt.2026.3718650}
}
Kostenlos starten →