«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

期数:

2025年05期

页码:

9-15

栏目:

媒体计算

出版日期:

2025-05-10

文章信息/Info

Title:

Self-supervised Image Denoising with Multi-branch Hybrid Masks

文章编号:

1673-629X(2025)05-0009-07

作者:

王玥乔; 李鑫远

北京邮电大学 人工智能学院,北京 100876

Author(s):

WANG Yue-qiao; LI Xin-yuan

School of Artificial Intelligence,Beijing University of Posts and Telecommunications,Beijing 100876,China

关键词:

图像去噪; 自监督学习; 盲点网络; 多支路混合掩码; 空间自相似注意力

Keywords:

image denoising; self-supervised learning; blind-spot network; multi-branch hybrid masks; spatial self-similarity attention

分类号:

TP183

DOI:

10.20165/j.cnki.ISSN1673-629X.2025.0063

摘要:

近年来,基于深度学习的图像去噪技术取得了较大的进展,然而当图像对难以获取时其将难以发挥优势。 自监督图像去噪算法由于仅需要学习噪声图像本身即可实现噪声的消除,已然成为研究热门。 鉴于图像中像素信号的空间相关性及噪声信号的空间独立性,基于盲点网络的去噪方法已展现出卓越的去噪效果。 然而,当前盲点网络方法在面临噪声相关区域高度相关时,难以有效打破噪声的空间特性。 针对这一挑战,提出了一种创新的多支路混合掩码自监督图像去噪方法。 该方法通过巧妙组合不同形式的掩码,有效打破了噪声的空间相关性,并借助空间自相似注意力机制,深入挖掘了因掩码遮挡而缺失的细节信息。 实验结果表明,该方法在处理高相关性噪声时展现出了出色的恢复性能。 同时,得益于多支路融合与空间自相似注意力的运用,该方法在细节恢复方面相较于当前最先进的方法也表现出更为优越的效果。

Abstract:

In recent years,deep learning-based image denoising techniques have made significant progress. However,their advantages are difficult to fully exploit when image pairs are hard to obtain. Self-supervised image denoising algorithms,which only require learning from noisy images themselves to achieve noise removal,have become a popular research topic. Given the spatial correlation of pixel signals and the spatial independence of noise signals in images,denoising methods based on blind-spot networks have demonstrated ex-ceptional performance. However,current blind - spot network methods struggle to effectively break the spatial characteristics of noise when dealing with highly correlated noisy regions. To address this challenge,we propose an innovative multi-branch hybrid mask self-supervised image denoising method. By cleverly combining different types of masks, this approach effectively breaks the spatial correlation of noise,and through a spatial self-similarity attention mechanism,deeply mines the missing detailed information due to mask occlusion. Experimental results show that the proposed method exhibits excellent recovery performance when handling highly correlated noise. Moreover,thanks to the use of multi-branch fusion and spatial self-similarity attention,the proposed method outperforms the current state-of-the-art techniques in terms of detail restoration.

常用功能

更新日期/Last Update: 2025-05-10