基于物理噪声与时空变换器的RAW视频去噪方法
DOI:
CSTR:
作者:
作者单位:

重庆光电技术研究所

作者简介:

通讯作者:

中图分类号:

基金项目:


RAW Video Denoising via Physics-Guided Noise Modeling and Spatio-Temporal Transformers
Author:
Affiliation:

Chongqing Optoelectronics Research Institute

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    针对微光视频中由泊松–高斯噪声引起的低信噪比以及噪声随机分布导致的时空不稳定问题,本文提出一种噪声模型引导的时空网络(NM-STNet)。该方法通过噪声感知编码器将物理噪声统计特性嵌入特征学习过程,并利用基于变换器(Transformer)的时空融合模块,在无需显式运动估计的情况下实现多帧特征的自适应对齐与聚合。此外,引入基于物理模型的噪声合成及矫正策略,以提升有限训练数据条件下的泛化能力。实验结果表明,NM-STNet在真实弱光视频序列上取得了显著优于现有方法的去噪性能与时空一致性。

    Abstract:

    To address the low signal-to-noise ratio (SNR) induced by Poisson-Gaussian noise and the spatiotemporal instability resulting from the random distribution of noise in low-light videos, this paper proposes a Noise Model-guided Spatiotemporal Network (NM-STNet). The proposed method embeds physical noise statistics into the feature learning process via a noise-aware encoder and employs a Transformer-based spatiotemporal fusion module to achieve adaptive alignment and aggregation of multi-frame features without explicit motion estimation. Furthermore, a physics-based noise synthesis and correction strategy is introduced to enhance generalization capabilities under conditions of limited training data. Experimental results demonstrate that NM-STNet significantly outperforms existing methods in terms of both denoising performance and spatiotemporal consistency on real-world low-light video sequences.

    参考文献
    相似文献
    引证文献
引用本文
分享
相关视频

文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2026-02-02
  • 最后修改日期:2026-02-02
  • 录用日期:2026-02-04
  • 在线发布日期:
  • 出版日期:
文章二维码