基于前馈卷积神经网络的多模块自监督降噪模型

doi:10.13229/j.cnki.jdxbgxb.20250138

摘要/Abstract

摘要：

针对降噪过程中成对的噪声图像和干净图像难以获取的问题，提出了一种基于前馈降噪卷积神经网络的多模块自监督降噪模型（MMS-DNet）。该模型采用改进的DnCNN架构，通过添加结合原图下采样和全局噪声水平估计的多元输入模块，生成多张输入子图的同时弥补了生成过程中的信息丢失；通过添加子像素卷积模块，避免了图像重建过程中冗余信息的引入。实验结果表明：本文模型在灰度数据集BSD68和Set12上的效果优于现有的自监督学习模型，在彩色数据集CBSD68和Kodak24上同样表现出良好的效果。

关键词: 图像降噪, 前馈神经网络, 自监督, 多模块

Abstract:

In order to solve the problem that pairwise noisy images and clean images are difficult to obtain in the process of denoising， this paper proposes a multi-modular self-supervised image denoising network based on feed-forward convolutional neural network（MMS-DNet）. The algorithm uses the improved DnCNN architecture. By adding a multivariate input module that combines original image down-sampling and global noise level estimation， multiple input sub-images are generated while compensating for the information loss in the generation process. The introduction of redundant information in the image reconstruction process is avoided by adding a sub-pixel convolution module. Experimental results show that the network outperforms existing self-supervised learning algorithms on the BSD68 and Set12 datasets of intensity images， and also achieves good results on CBSD68 and Kodak24 datasets of color images.

Key words: image denoising, feed-forward neural network, self-supervised, multi-modular

中图分类号:

TP751

郭晓新,李广宇,杨广奇. 基于前馈卷积神经网络的多模块自监督降噪模型[J]. 吉林大学学报(工学版), 2026, 56(9): 2467-2475.

Xiao-xin GUO,Guang-yu LI,Guang-qi YANG. Multi-modular self-supervised image denoising network based on feed-forward convolutional neural network[J]. Journal of Jilin University(Engineering and Technology Edition), 2026, 56(9): 2467-2475.

图/表 16

图1

图2

图3

表1

图4

表2

在BSD68数据集上的不同噪声强度对比实验 (dB/-)"

方法	$σ$ =15	$σ$ =25	$σ$ =50
方法	PSNR/SSIM	PSNR/SSIM	PSNR/SSIM
BM3D	31.07/0.82	28.57/0.80	25.62/0.69
WNNM	31.37/0.82	28.83/0.81	25.87/0.70
DnCNN	31.72/0.83	29.23/0.83	26.23/0.72
N2N	30.89/0.85	28.86/0.82	25.77/0.70
Blind2Unblind	31.44/0.89	28.99/0.82	26.09/0.72
SwinIA	31.07/0.86	29.17/0.80	26.61/0.71
MMS-DNet	31.66/0.87	29.28/0.82	26.30/0.72

表2

图5

表3

在Set12数据集上的不同噪声强度对比实验 (dB/-)"

方法	$σ$ =15	$σ$ =25	$σ$ =50
方法	PSNR/SSIM	PSNR/SSIM	PSNR/SSIM
BM3D	32.37/0.87	29.97/0.85	26.72/0.77
WNNM	32.70/0.85	30.26/0.86	27.05/0.78
DnCNN	32.86/0.91	30.43/0.87	27.18/0.78
N2N	32.60/0.90	30.02/0.85	26.83/0.77
Blind2Unblind	32.46/0.90	30.09/0.85	26.01/0.78
SwinIA	30.37/0.86	28.72/0.82	26.03/0.74
MMS-DNet	32.75/0.89	30.47/0.88	27.41/0.79

表3

图6

表4

在CBSD68数据集上的不同噪声强度对比实验 (dB/-)"

方法	$σ$ =15	$σ$ =25	$σ$ =50
方法	PSNR/SSIM	PSNR/SSIM	PSNR/SSIM
BM3D	33.52/0.93	30.71/0.86	27.38/0.79
WNNM	33.67/0.93	31.02/0.87	27.79/0.79
DnCNN	33.89/0.93	31.23/0.88	27.92/0.79
N2N	-/-	30.79/0.87	-/-
Blind2Unblind	33.67/0.92	30.86/0/86	27.88/0.78
SwinIA	33.45/0.91	28.40/0.79	27.34/0.72
MMS-DNet	33.87/0.93	31.21/0.88	27.95/0.79

表4

图7

表5

在Kodak24数据集上的不同噪声强度对比实验 (dB/-)"

方法	$σ$ =15	$σ$ =25	$σ$ =50
方法	PSNR/SSIM	PSNR/SSIM	PSNR/SSIM
BM3D	34.28/0.92	31.68/0.87	28.46/0.78
WNNM	34.35/0.92	32.00/0.88	28.74/0.79
DnCNN	34.48/0.92	32.03/0.88	28.85/0.80
N2N	-/-	32.08/0.88	-/-
Blind2Unblind	34.12/0.90	32.27/0.88	28.92/0.81
SwinIA	33.50/0.87	30.12/0.82	28.33/0.75
MMS-DNet	34.62/0.92	32.11/0.88	29.00/0.80

表5

图8

表6

表7

不同主体网络消融实验 (dB)"

主体网络	$σ$ =15	$σ$ =25	$σ$ =50
DnCNN	31.66	29.28	26.30
U-Net	30.79	28.99	26.16
VDSR	30.77	28.64	26.12

表7

图9

参考文献 24

[1]	Su J, Xu B, Yin H. A survey of deep learning approaches to image restoration[J]. Neurocomputing, 2022, 487: 46-65.
[2]	Liu D, Wen B X, Liu X M, et al. When image denoising meets high-level vision tasks: A deep learning approach[C]∥Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, Stockholm, Sweden, 2018: 842-848.
[3]	Tomasi C, Manduchi R. Bilateral filtering for gray and color images[C]∥Proceedings of Sixth International Conference on Computer Vision, Bombay, India, 1998: 839-846.
[4]	Gu S H, Zhang L, Zuo W M, et al. Weighted nuclear norm minimization with application to image denoising[C]∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Columbus, USA, 2014: 2862-2869.
[5]	Buades A, Coll B, Morel J M. A non-local algorithm for image denoising[C]∥2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, San Diego, USA, 2005: 60-65.
[6]	Dabov K, Foi A, Katkovnik V, et al. Image denoising by sparse 3-D transform-domain collaborative filtering[J]. IEEE Transactions on Image Processing, 2007, 16(8): 2080-2095.
[7]	Zhang K, Zuo W M, Gu S H, et al. Learning deep CNN denoiser prior for image restoration[C]∥2017 IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, USA, 2017: 2808-2817.
[8]	Valsesia D, Fracastoro G, Magli E. Deep graph-convolutional image denoising[J]. IEEE Transactions on Image Processing, 2020, 29: 8226-8237.
[9]	Lehtinen J, Munkberg J, Hasselgren J, et al. Noise2Noise: learning image restoration without clean data[C]∥Proceedings of International Conference on Machine Learning, Stockholm, USA, 2018: 2965-2974.
[10]	Goodfellow I, Pougetabadie J, Mirza M, et al. Generative adversarial networks[J]. Communications of the ACM, 2020, 63(11): 139-144.
[11]	Chen J W, Chen J W, Chao H Y, et al. Image blind denoising with generative adversarial network based noise modeling[C]∥2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, USA, 2018: 3155-3164.
[12]	Chen H T, Wang Y H, Guo T Y, et al. Pre-trained image processing Transformer[C]∥2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Nashville, USA,2021:12294-12305.
[13]	Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need[J]. Advances in Neural Information Processing Systems, 2017, 30: 5998-6008.
[14]	Deng J, Dong W, Socher R, et al. ImageNet: A large-scale hierarchical image database[C]∥2009 IEEE Conferenceon Computer Vision and Pattern Recognition, Miami, USA,2009: 248-255.
[15]	Pang T Y, Zheng H, Quan Y H, et al. Recorrupted-to-recorrupted: Unsupervised deep learning for image denoising[C]∥Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Nashville, USA,2021: 2043-2052.
[16]	Martin D, Fowlkes C, Tal D, et al. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics[C]∥Proceedings Eighth IEEE International Conference on Computer Vision, Vancouver, Canada, 2001: 416-423.
[17]	Zhang K, Zuo W M, Chen Y J, et al. Beyond a gaussian denoiser: residual learning of deep CNN for image denoising[J]. IEEE Transactions on Image Processing, 2017, 26: 3142-3155,
[18]	Franzen R. Kodak lossless true color image suite[EB/OL]. [2026-08-23].
[19]	Huynh-Thu Q, Ghanbari M. Scope of validity of PSNR in image/video quality assessment[J]. Electronics Letters, 2008, 44(13): 800-801.
[20]	Wang Z, Bovik A C, Sheikh H R, et al. Image quality assessment: from error visibility to structural similarity[J]. IEEE Transactions on Image Processing, 2004, 13(4): 600-612.
[21]	Zhang K, Zuo W M, Zhang L. FFDNet: toward a fast and flexible solution for CNN-based image denoising[J]. IEEE Transactions on Image Processing, 2018, 27(9): 4608-4622.
[22]	Gu S H, Zhang L, Zuo W M, et al. Weighted nuclear norm minimization with application to image denoising[C]∥IEEE Conference on Computer Vision and Pattern Recognition, Columbus, USA,2014: 2862-2869.
[23]	Wang Z J, Liu J Z, Li G Q, et al. Blind2unblind: Self-supervised image denoising with visible blind spots[C]∥Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, New Orleans, USA,2022: 2017-2026.
[24]	Mikhail P, Chizhov P, Parts L. SwinIA: self-supervised blind-spot image denoising with zero convolutions[C]∥2025 IEEE/CVF Winter Conference on Applications of Computer Vision, Tucson, Arizona, USA,2023: 7071-7080.

Metrics

Viewed

Full text

Abstract

Cited

Shared

Discussed

项目	网络层数	通道数	噪声范围	子图大小
灰度图	14	64	［15，50］	70×70
彩色图	12	96	［15，50］	50×50

实验序号	架构模块	PSNR/dB
1	DnCNN	29.23
2	DnCNN+池化下采样 +反池化	29.19
3	DnCNN+多元输入模块+反池化	29.25
4	DnCNN+池化下采样+子像素卷积	29.24
5	DnCNN+多元输入模块+子像素卷积	29.28