基于光流和多尺度特征融合的视频去噪算法

PDF下载HTML

Title:

Video denoising based on optical flow and multi-scale features

作者:

孙立辉¹, 陈恒¹, 商月平²

1. 河北经贸大学 信息技术学院, 河北 石家庄 050061;
2. 河北经贸大学 数学与统计学学院, 河北 石家庄 050061

Author(s):

SUN Lihui¹, CHEN Heng¹, SHANG Yueping²

1. School of Information Technology, Hebei University of Economics and Business, Shijiazhuang 050061, China;
2. College of Mathematics and Statistics, Hebei University of Economics and Business, Shijiazhuang 050061, China

关键词:

多帧去噪; 视频去噪; 光流对齐; 感知损失; 非局部注意力; 图像处理; 计算机视觉; 深度学习

Keywords:

multi-frame noise reduction; video denoising; optical flow alignment; perceptual loss; non-local attention; image processing; computer vision; deep learning

分类号:

TP391

DOI:

10.11992/tis.202306002

摘要:

为有效地去除视频当中的噪声，减少纹理细节丢失，提出了一种基于光流和多尺度特征融合的级联视频去噪算法。通过分组策略对序列帧进行精准对齐，然后送入集成残差细化和可选择性跳跃连接的多尺度架构，实现细节特征的精确保留与高效融合，进而采用非局部注意力机制以深入挖掘视频帧的时空特征，重建高质量视频。同时为保留更多纹理细节，提出一种联合感知损失的目标函数监督训练。实验结果表明，所提算法的去噪结果可以保留更多的纹理特征，更符合人眼视觉的习惯。该算法在强噪声下具备鲁棒性高、计算量小的特点，可以满足实时去噪的要求。

Abstract:

To effectively eliminate noise from videos while preserving texture details, a cascade video denoising algorithm that integrates optical flow and multi-scale features is proposed. The process begins by accurately aligning sequence frames using a grouping strategy. These frames are then processed through a multi-scale architecture that combines residual refinement and selective skip connection. This approach not only preserves detailed features but also enhances alignment and fusion. Furthermore, a non-local attention mechanism is employed to deeply mine spatiotemporal features, enabling the reconstruction of high-quality videos. To preserve detailed textures, a target function supervision training method that combines perceptual loss is proposed. Experimental results show that the proposed algorithm retains more texture features and aligns well with human visual perception. It is also highly robust, has low computational complexity under strong noise, and meets real-time denoising requirements.