A CAViT-UNet Hybrid Architecture Method for Radio Frequency Interference Segmentation
Research in Astronomy and Astrophysics · 01.01.2027 · 27 (1) · 015024
DOI: 10.1088/1674-4527/aea1ef ↗Kurz gefasst
Fachartikel in der Zeitschrift Research in Astronomy and Astrophysics, online erschienen am 10.11.2026. Verfasst von Hongmei Tang und 10 weiteren Autor:innen. Die Arbeit nennt 23 Literaturangaben.
Abstract (Originalsprache)
Radio frequency interference (RFI) severely contaminates radio astronomical observation data, particularly in high-sensitivity pulsar surveys, where its complex patterns and fluctuating intensity make traditional methods ineffective for precise removal. To address this, this study proposes a ViT-UNet hybrid architecture integrating cross-attention mechanisms (CAViT-UNet). A standard single-scale Vision Transformer encoder extracts global semantic features, while a CNN encoder with strip pooling captures local spatial details. Cross-attention fusion dynamically aligns global semantics with local features. Experiments on LOFAR real data and seven types of simulated RFI datasets demonstrate that the method achieves 91.09% IoU and 95.34% F1 score on simulated data, while achieving comparable performance to U-Net on real LOFAR data (F1 = 0.5887 versus 0.5916). Visualization results further confirm that CAViT-UNet accurately segments complex interference patterns with clear structure, effectively suppressing false positives and missed detections in controlled scenarios.
Zitieren
APA
Tang, H., Zhang, H., Jiao, Y., Zhang, Y., Wang, J., Ye, X., Li, J., Wang, W., Zhang, T., Du, X., & Cai, W. (2027). A CAViT-UNet Hybrid Architecture Method for Radio Frequency Interference Segmentation. Research in Astronomy and Astrophysics, 27(1), 015024. https://doi.org/10.1088/1674-4527/aea1ef
Vancouver
Tang H, Zhang H, Jiao Y, Zhang Y, Wang J, Ye X, et al. A CAViT-UNet Hybrid Architecture Method for Radio Frequency Interference Segmentation. Research in Astronomy and Astrophysics. 2027;27(1):015024. doi:10.1088/1674-4527/aea1ef
BibTeX
@article{Tang2027,
author = {Tang, Hongmei and Zhang, Hailong and Jiao, Yuyue and Zhang, Yazhou and Wang, Jie and Ye, Xinchen and Li, Jia and Wang, Wanqiong and Zhang, Ting and Du, Xu and Cai, Wenna},
title = {A CAViT-UNet Hybrid Architecture Method for Radio Frequency Interference Segmentation},
journal = {Research in Astronomy and Astrophysics},
year = {2027},
volume = {27},
number = {1},
pages = {015024},
doi = {10.1088/1674-4527/aea1ef}
}
EUVIMED