多尺度密集交互注意力残差真实图像去噪网络

doi:10.11996/JG.j.2095-302X.2025020279

图学学报 ›› 2025, Vol. 46 ›› Issue (2): 279-287.DOI: 10.11996/JG.j.2095-302X.2025020279

• 图像处理与计算机视觉 • 上一篇下一篇

多尺度密集交互注意力残差真实图像去噪网络

郭业才¹^,²(), 胡晓伟¹, 毛湘南¹

1.南京信息工程大学电子与信息工程学院，江苏南京 210044
2.无锡学院电子信息工程学院，江苏无锡 214105

收稿日期:2024-07-16 接受日期:2024-12-06 出版日期:2025-04-30 发布日期:2025-04-24
第一作者:郭业才(1962-)，男，教授，博士。主要研究方向为机器学习、气象通信技术、水下通信理论及其应用等。E-mail：guo-yecai@163.com
基金资助:
国家自然科学基金(61673222)

Multiscale dense interactive attention residual real image denoising network

GUO Yecai¹^,²(), HU Xiaowei¹, AMITAVE Saha¹, MAO Xiangnan¹

1. College of Electronic and Information Engineering, Nanjing University of Information Science and Technology, Nanjing Jiangsu 210044, China
2. College of Electronic Information Engineering, Wuxi University, Wuxi Jiangsu 214105, China

Received:2024-07-16 Accepted:2024-12-06 Published:2025-04-30 Online:2025-04-24
First author：GUO Yecai (1962-), professor, Ph.D. His main research interests cover machine learning, meteorological communication technology, underwater communication theory and its applications, etc. E-mail：guo-yecai@163.com
Supported by:
National Natural Science Foundation of China(61673222)

摘要/Abstract

摘要：

针对图像去噪特征提取不全面以及特征利用率低，导致生成图像不够清晰的问题，提出一种多尺度密集交互注意力残差去噪网络(MDIARN)。首先，通过多尺度非对称特征提取模块(MAFM)初步提取浅层信息特征，以确保图像特征的多样性；然后，多尺度级联模块(MSCM)利用多维密集交互残差单元(MDIU)对图像特征进行多维映射，并逐步级联以增强模型之间的信息传递和交互性，充分拟合训练数据；引入双路全局注意力模块(DGAM)对多级特征进行全局联合学习，获取更具有判别性的特征信息；跳跃连接促进结构之间的参数共享，使不同维度的特征充分融合，保证信息的完整性；最后，采用残差学习构建出清晰的去噪图像。结果表明，该算法在真实噪声数据集(DND和SIDD)上峰值信噪比分别为39.80 dB和39.62 dB，结构相似性分别为95.4%和95.8%，均优于主流去噪算法。此外，该算法在低光度场景下应用也能保留更多细节，从而显著提升图像质量。

关键词: 图像去噪, 多尺度特征提取, 多维密集交互, 卷积神经网络, 注意力

Abstract:

To address the problem that the generated image is not clear enough due to incomplete feature extraction and low feature utilization in image denoising, a multi-scale dense interactive attention residual denoising network (MDIARN) was proposed. First, a multi-scale asymmetric feature extraction module (MAFM) was employed to preliminarily extract shallow information features, ensuring diversity of image features. Then, a multi-scale cascade module (MSCM) utilized multi-dimensional dense interactive residual units (MDIU) to perform multi-dimensional mapping of image features. These units were progressively cascaded to enhance the information transmission and interaction between models, fully fitting the training data. A dual-path global attention module (DGAM) was introduced to conduct global joint learning on multi-level features, acquiring more discriminative feature information. Skip connections were integrated to encourage parameter sharing between structures, enabling full integration of features from different dimensions and preserving the completeness of information. Finally, residual learning was employed to construct a clear denoised image. Experimental results demonstrated that this algorithm achieved peak signal-to-noise ratios of 39.80 dB and 39.62 dB on the real noise datasets (DND and SIDD), respectively, and structural similarities of 95.4% and 95.8%, respectively, outperforming mainstream denoising algorithms. In addition, the proposed algorithm demonstrated excellent performance in low-light environments, preserving more details and significantly enhancing image quality.

Key words: image denoising, multi-scale feature extraction, multi-dimensional dense interaction, convolutional neural network, attention

中图分类号:

TP391

郭业才, 胡晓伟, 毛湘南. 多尺度密集交互注意力残差真实图像去噪网络[J]. 图学学报, 2025, 46(2): 279-287.

GUO Yecai, HU Xiaowei, AMITAVE Saha, MAO Xiangnan. Multiscale dense interactive attention residual real image denoising network[J]. Journal of Graphics, 2025, 46(2): 279-287.

图/表 15

图1 MDIARN整体框图

Fig. 1 Overall block diagram of MDIARN

图2 多尺度非对称特征提取模块((a) MAFM增强骨架特征细节操作；(b) MAFM详细结构)

Fig. 2 Multi-scale asymmetric feature extraction module((a) MAFM enhanced skeleton feature detail operation; (b) Detailed structure of MAFM)

图3 MSCM详细结构

Fig. 3 Detailed structure of the MSCM

图4 DGAM详细结构

Fig. 4 Detailed structure of the DGAM

表1 环境配置

Table 1 Environment configuration

名称	配置
Operating system	Windows 10
GPU	NVIDIA GeForce GTX 3060 Ti
CPU	Intel Core i5- 12400K @3.40 GHz
Deep learning framework	PyTorch 1.7.1+cu 101
Python Version	3.7

表2 不同算法在SIDD和DND数据集上的去噪对比结果

Table 2 Comparison results of different algorithms for denoising on SIDD and DND datasets

算法	SIDD		DND
算法	PSNR/dB	SSIM/%	PSNR/dB	SSIM/%
BM3D	26.65	68.5	34.51	85.1
RIDNet	38.73	95.4	39.25	95.0
AINDNet	38.96	95.2	39.37	95.1
MSGAN	39.11	95.5	39.59	95.5
DCBDNet	38.94	95.3	39.37	95.1
DCANet	39.27	95.6	39.57	95.3
MRIDN	39.43	95.7	39.49	95.1
Ours	39.62	95.8	39.80	95.4

图5 不同算法在SIDD数据集上去噪效果对比

Fig. 5 Comparison of denoising effects of different algorithms on the SIDD dataset ((a) Noisy image; (b) BM3D; (c) RIDNet; (d) DCBDNet; (e) DCANAet; (f) MRIDN; (g) Ours; (h) Clean image)

图6 不同算法在DND数据集上去噪效果对比

Fig. 6 Comparison of denoising effects of different algorithms on the DND dataset ((a) Noisy image; (b) BM3D; (c) RIDNet; (d) DCBDNet; (e) DCANAet; (f) MSGAN; (g) MRIDN; (h) Ours)

图7 不同算法在MIDD数据集上去噪效果对比

Fig. 7 Comparison of denoising effects of different algorithms on the MIDD dataset ((a) Noisy image; (b) BM3D; (c) RIDNet; (d) DCBDNet; (e) DCANAet; (f) MSGAN; (g) Ours; (h) Clean image)

图8 不同算法在手机拍摄的图像上去噪效果对比

Fig. 8 Comparison of denoising effects of different algorithms on images taken by cell phones ((a) Noisy image; (b) BM3D; (c) RIDNet; (d) DCBDNet; (e) AINDNet; (f) DCANAet; (g) MSGAN; (h) Ours)

表3 卷积核性能比较

Table 3 Convolutional kernel performance comparison

模块	类型	参数/M	PSNR/dB	SSIM/%
MAFM	非对称	9.67	39.62	95.8
3×3 Conv	对称	9.60	39.53	95.5

表4 消融实验结果

Table 4 Results of ablation experiments

变体	模块				PSNR/dB	SSIM/%
变体	MAFM	MSCM	DGAM	Skip connection	PSNR/dB	SSIM/%
MDIARN-a	√	-	-	-	35.36	91.0
MDIARN-b	√	√	-	-	39.45	95.4
MDIARN-c	√	-	√	-	37.53	94.0
MDIARN-d	√	√	√	-	39.54	95.7
MDIARN-e	√	√	-	√	39.53	95.7
Ours	√	√	√	√	39.62	95.8

表5 MSCM数量性能比较

Table 5 Performance comparison of the number of MSCM

数量	参数/M	PSNR/dB	SSIM/%
1	2.68	39.25	95.5
2	5.01	39.43	95.7
3	7.34	39.50	95.7
4	9.67	39.62	95.8
5	12.00	39.64	95.8

表6 MDIU中卷积数量性能对比

Table 6 Performance comparison of the number of convolutions in MDIU

数量	参数/M	PSNR/dB	SSIM/%
3	4.14	39.39	95.3
4	6.66	39.53	95.5
5	9.67	39.62	95.8
6	13.28	39.63	95.8
7	17.47	39.64	95.8

表7 不同模型在一个256×256噪声图像上运行的时间以及模型参数数量

Table 7 Runtimes and model parameters for different models on a 256×256 noisy image

方法	时间/s	参数/M
BM3D	2.165(CPU)	-
RIDNet	0.046(GPU)	1.50
DCBDNet	0.054(GPU)	1.10
DCANet	0.059(GPU)	1.40
AINDNet	0.085(GPU)	13.76
MRIDN	0.051(GPU)	6.92
Ours	0.048(GPU)	9.67

参考文献 25

[1]	SOLOMON C, BRECKON T. Fundamentals of digital image processing: a practical approach with examples in MATLAB[M]. Hoboken: Wiley-Blackwell, 2011: 6-7.
[2]	AFONSO M V, BIOUCAS-DIAS J M, FIGUEIREDO M A T. Fast image recovery using variable splitting and constrained optimization[J]. IEEE Transactions on Image Processing, 2010, 19(9): 2345-2356. DOI PMID
[3]	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. DOI PMID
[4]	ANWAR S, BARNES N. Real image denoising with feature attention[C]// 2019 IEEE/CVF International Conference on Computer Vision. New York: IEEE Press, 2019: 3155-3164.
[5]	YAO C, TANG Y B, SUN J, et al. Multiscale residual fusion network for image denoising[J]. IET Image Processing, 2022, 16(3): 878-887.
[6]	吴文聪. 基于卷积神经网络的图像去噪方法研究[D]. 昆明: 云南师范大学, 2023.
	WU W C. Research on convolutional neural network for image denoising[D]. Kunming: Yunnan Normal University, 2023 (in Chinese).
[7]	YU X J, FU Z X, GE C K. A multi-scale generative adversarial network for real-world image denoising[J]. Signal, Image and Video Processing, 2022, 16(1): 257-264.
[8]	DING X H, GUO Y C, DING G G, et al. ACNet: strengthening the kernel skeletons for powerful CNN via asymmetric convolution blocks[C]// 2019 IEEE/CVF International Conference on Computer Vision. New York: IEEE Press, 2019: 1911-1920.
[9]	HEIDE F, STEINBERGER M, TSAI Y T, et al. FlexISP: a flexible camera image processing framework[J]. ACM Transactions on Graphics, 2014, 33(6): 231.
[10]	WOO S, PARK J, LEE J Y, et al. CBAM: convolutional block attention module[C]// The 15th European Conference on Computer Vision. Cham: Springer, 2018: 3-19.
[11]	满开亮, 汪友生, 刘继荣. 基于稠密残差网络的图像超分辨率重建算法[J]. 图学学报, 2021, 42(4): 556-562.
	MAN K L, WANG Y S, LIU J R. Image super-resolution reconstruction algorithm based on dense residual network[J]. Journal of Graphics, 2021, 42(4): 556-562 (in Chinese).
[12]	CRNJANSKI J, KRSTIĆ M, TOTOVIĆ A, et al. Adaptive sigmoid-like and PReLU activation functions for all-optical perceptron[J]. Optics Letters, 2021, 46(9): 2003-2006. DOI PMID
[13]	JIANG D C, QU H, ZHAO J H, et al. Aggregating multi-scale contextual features from multiple stages for semantic image segmentation[J]. Connection Science, 2021, 33(3): 605-622.
[14]	HE X, CHEN Y S. Modifications of the multi-layer perceptron for hyperspectral image classification[J]. Remote Sensing, 2021, 13(17): 3547.
[15]	魏敏, 姚鑫. 基于多尺度与注意力机制的两阶段风暴单体外推研究[J]. 图学学报, 2024, 45(4): 696-704. DOI
	WEI M, YAO X. Two-stage storm entity prediction based on multiscale and attention[J]. Journal of Graphics, 2024, 45(4): 696-704 (in Chinese). DOI
[16]	CHARBONNIER P, BLANC-FERAUD L, AUBERT G, et al. Two deterministic half-quadratic regularization algorithms for computed imaging[C]// The 1st International Conference on Image Processing. New York: IEEE Press, 1994: 168-172.
[17]	JIANG K, WANG Z Y, YI P, et al. Multi-scale progressive fusion network for single image deraining[C]// 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2020: 8343-8352.
[18]	ABDELHAMED A, LIN S, BROWN M S. A high-quality denoising dataset for smartphone cameras[C]// 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2018: 1692-1700.
[19]	PLÖTZ T, ROTH S. Benchmarking denoising algorithms with real photographs[C]// 2017 IEEE Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2017: 2750-2759.
[20]	FLEPP R, IGNATOV A, TIMOFTE R, et al. Real-world mobile image denoising dataset with efficient baselines[C]// 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2024: 22368-22377.
[21]	佟雨兵, 张其善, 祁云平. 基于PSNR与SSIM联合的图像质量评价模型[J]. 中国图象图形学报, 2006, 11(12): 1758-1763.
	TONG Y B, ZHANG Q S, QI Y P. Image quality assessing by combining PSNR with SSIM[J]. Journal of Image and Graphics, 2006, 11(12): 1758-1763 (in Chinese).
[22]	CAZENAVE T, SENTUC J, VIDEAU M. Cosine annealing, Mixnet and Swish activation for computer go[C]// The 17th International Conference on Advances in Computer Games. Cham: Springer, 2022: 53-60.
[23]	常禧龙, 梁琨, 李文涛. 深度学习优化器进展综述[J]. 计算机工程与应用, 2024, 60(7): 1-12. DOI
	CHANG X L, LIANG K, LI W T. Review of development of deep Learning optimizer[J]. Computer Engineering and Applications, 2024, 60(7): 1-12 (in Chinese). DOI
[24]	KIM Y, SOH J W, PARK G Y, et al. Transfer learning from synthetic to real-noise denoising with adaptive instance normalization[C]// 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE Press, 2020: 3479-3489.
[25]	冯妍舟, 刘建霞, 王海翼, 等. 基于多级残差信息蒸馏的真实图像去噪方法[J]. 计算机工程, 2024, 50(3): 216-223. DOI
	FENG Y Z, LIU J X, WANG H Y, et al. Real image denoising method based on multi-level residual information distillation[J]. Computer Engineering, 2024, 50(3): 216-223 (in Chinese). DOI

多尺度密集交互注意力残差真实图像去噪网络

Multiscale dense interactive attention residual real image denoising network

RichHTML

PDF (PC)

可视化

摘要/Abstract

引用本文

使用本文

图/表 15

参考文献 25

相关文章 15

编辑推荐

Metrics

本文评价

[1]	张立立, 杨康, 张珂, 魏薇, 李晶, 谭洪鑫, 张翔宇. 面向柴油车辆排放黑烟的改进型YOLOv8检测算法研究[J]. 图学学报, 2025, 46(2): 249-258.
[2]	崔克彬, 耿佳昌. 基于EE-YOLOv8s的多场景火灾迹象检测算法[J]. 图学学报, 2025, 46(1): 13-27.
[3]	吴亦奇, 何嘉乐, 张甜甜, 张德军, 李艳丽, 陈壹林. 基于多重特征提取和点对应关系的三维点云非刚配准[J]. 图学学报, 2025, 46(1): 150-158.
[4]	陈冠豪, 徐丹, 贺康建, 施洪贞, 张浩. 基于转置注意力和CNN的图像超分辨率重建网络[J]. 图学学报, 2025, 46(1): 35-46.
[5]	张文祥, 王夏黎, 王欣仪, 杨宗宝. 一种强化伪造区域关注的深度伪造人脸检测方法[J]. 图学学报, 2025, 46(1): 47-58.
[6]	苑朝, 赵明雪, 张丰羿, 冯晓勇, 李冰, 陈瑞. 基于点云特征增强的复杂室内场景3D目标检测[J]. 图学学报, 2025, 46(1): 59-69.
[7]	卢洋, 陈林慧, 姜晓恒, 徐明亮. SDENet：基于多尺度注意力质量感知的合成缺陷数据评价网络[J]. 图学学报, 2025, 46(1): 94-103.
[8]	胡凤阔, 叶兰, 谭显峰, 张钦展, 胡志新, 方清, 王磊, 满孝锋. 一种基于改进YOLOv8的轻量化路面病害检测算法[J]. 图学学报, 2024, 45(5): 892-900.
[9]	刘义艳, 郝婷楠, 贺晨, 常英杰. 基于DBBR-YOLO的光伏电池表面缺陷检测[J]. 图学学报, 2024, 45(5): 913-921.
[10]	吴沛宸, 袁立宁, 胡皓, 刘钊, 郭放. 基于注意力特征融合的视频异常行为检测[J]. 图学学报, 2024, 45(5): 922-929.
[11]	翟永杰, 李佳蔚, 陈年昊, 王乾铭, 王新颖. 融合改进Transformer的车辆部件检测方法[J]. 图学学报, 2024, 45(5): 930-940.
[12]	刘丽, 张起凡, 白宇昂, 黄凯烨. 结合Swin Transformer的多尺度遥感图像变化检测研究[J]. 图学学报, 2024, 45(5): 941-956.
[13]	章东平, 魏杨悦, 何数技, 徐云超, 胡海苗, 黄文君. 特征融合与层间传递：一种基于Anchor DETR改进的目标检测方法[J]. 图学学报, 2024, 45(5): 968-978.
[14]	谢国波, 林松泽, 林志毅, 吴陈锋, 梁立辉. 基于改进YOLOv7-tiny的道路病害检测算法[J]. 图学学报, 2024, 45(5): 987-997.
[15]	熊超, 王云艳, 罗雨浩. 特征对齐与上下文引导的多视图三维重建[J]. 图学学报, 2024, 45(5): 1008-1016.