DnRFD：用于图像去噪的递进式残差融合密集网络

doi:10.3778/j.issn.1673-9418.2103030

计算机科学与探索 ›› 2022, Vol. 16 ›› Issue (12): 2841-2850.DOI: 10.3778/j.issn.1673-9418.2103030

DnRFD：用于图像去噪的递进式残差融合密集网络

曹义亲¹, 饶哲初¹, 朱志亮¹^,²^,⁺(), 张红斌¹

1.华东交通大学软件学院，南昌 330013
2.中国科学院软件研究所，北京 100190

收稿日期:2021-03-10 修回日期:2021-04-27 出版日期:2022-12-01 发布日期:2021-04-30
通讯作者: +E-mail: rj_zzl@ecjtu.edu.cn
作者简介:曹义亲（1964—），男，江西九江人，硕士，教授，CCF会员，主要研究方向为图像处理、模式识别。
饶哲初（1997—），男，江西丰城人，硕士研究生，主要研究方向为图像处理。
朱志亮（1988—），男，湖北天门人，博士，讲师，CCF会员，主要研究方向为图像信息处理、虚拟现实、人机交互。
张红斌（1979—），男，江苏如皋人，博士，副教授，CCF会员，主要研究方向为计算机视觉、自然语言处理、推荐系统。
基金资助:
国家自然科学基金(61861016);江西省科技支撑计划重点项目(20161BBE50081);江西省青年科学基金项目(20202BABL212006);江西省教育厅科学技术研究项目(GJJ190359)

DnRFD：Progressive Residual Fusion Dense Network for Image Denoising

CAO Yiqin¹, RAO Zhechu¹, ZHU Zhiliang¹^,²^,⁺(), ZHANG Hongbin¹

1. School of Software, East China Jiaotong University, Nanchang 330013, China
2. Institute of Software, Chinese Academy of Sciences, Beijing 100190, China

Received:2021-03-10 Revised:2021-04-27 Online:2022-12-01 Published:2021-04-30
About author:CAO Yiqin, born in 1964, M.S., professor, member of CCF. His research interests include image processing and pattern recognition.
RAO Zhechu, born in 1997, M.S. candidate. His research interest is image processing.
ZHU Zhiliang, born in 1988, Ph.D., lecturer, member of CCF. His research interests include image information processing, virtual reality and human-computer interaction.
ZHANG Hongbin, born in 1979, Ph.D., associate professor, member of CCF. His research interests include computer vision, natural language processing and recommendation systems.
Supported by:
National Natural Science Foundation of China(61861016);Key Project of Science and Technology Support Plan of Jiangxi Province(20161BBE50081);Youth Science Foundation of Jiangxi Province(20202BABL212006);Science and Technology Research Project of Education Department of Jiangxi Province(GJJ190359)

摘要/Abstract

摘要：

基于深度学习的去噪方法能够获得比传统方法更好的去噪效果，但是现有的深度学习去噪方法往往存在网络过深导致计算复杂度过大的问题。针对这个不足，提出一种用于去除高斯噪声的递进式残差融合密集网络（DnRFD）。该网络首先采用密集块来学习图像中的噪声分布，在充分提取图像局部特征的同时大幅降低网络参数；然后利用递进策略将浅层卷积特征依次与深层特征短线连接形成残差融合网络，提取出更多针对噪声的全局特征；最后将各密集块的输出特征图进行融合后输入给重建输出层，得到最后的输出结果。实验结果表明，在高斯白噪声等级为25和50时，该网络都能获得较高的峰值信噪比均值和结构相似性均值，并且去噪平均时间是DnCNN方法的一半，是FFDNet方法的1/3。总的来说，该网络整体去噪性能优于相关对比算法，可有效去除图像中的高斯白噪声和自然噪声，同时能更好地还原图像边缘与纹理细节。

关键词: 图像去噪, 深度学习, 密集块, 残差学习, 递进式残差融合

Abstract:

The denoising method based on deep learning can achieve better denoising effect than the traditional method, but the existing deep learning denoising methods often have the problem of excessive computational complexity caused by too deep network. To solve this problem, a progressive residual fusion dense network (DnRFD) is proposed to remove Gaussian noise. Firstly, dense blocks are used to learn the noise distribution in the image, and the network parameters are greatly reduced while the local features of the image are fully extracted. Then, a progressive strategy is used to connect the shallow convolution features with the deep features to form a residual fusion network to extract more global features for noise. Finally, the output characteristic images of each dense block are fused and input to the reconstructed output layer to get the final output result. Experimental results show that, when the Gaussian white noise level is 25 and 50, the network can achieve higher mean PSNR and mean structural similarity, and the average time of denoising is half of the DnCNN method and one third of the FFDNet method. In general, the overall denoising performance of the network is better than that of the correlative comparison algorithms, and it can effectively remove the white Gaussian noise and natural noise in the image, and can restore the edge and texture details of the image better.

Key words: image denoising, deep learning, dense block, residual learning, progressive residual fusion

中图分类号:

TP391

曹义亲, 饶哲初, 朱志亮, 张红斌. DnRFD：用于图像去噪的递进式残差融合密集网络[J]. 计算机科学与探索, 2022, 16(12): 2841-2850.

CAO Yiqin, RAO Zhechu, ZHU Zhiliang, ZHANG Hongbin. DnRFD：Progressive Residual Fusion Dense Network for Image Denoising[J]. Journal of Frontiers of Computer Science and Technology, 2022, 16(12): 2841-2850.

图/表 17

图1 Dense Block模块结构

Fig.1 Dense Block modular structure

图2 Bottleneck结构

Fig.2 Bottleneck structure

图3 递进式残差融合密集网络整体结构

Fig.3 Whole structure of progressive residual fusion dense network

图4 基于RFD的图像去噪方法整体框架

Fig.4 Overall framework of image denoising method based on RFD

图5 用于低等级噪声去噪的测试原图

Fig.5 Original test images for low level noise denoising

表1 不同算法得到的PSNR结果（噪声等级为25）单位：dB

Table 1 PSNR results obtained by different algorithms (noise level is 25)

Methods	House	Pepper	Ship	Man	Landscape	Airplane	Average
NLM^[5]	30.532	28.380	27.921	27.996	30.661	35.476	30.161
BM3D^[6]	32.860	30.162	29.910	29.623	32.542	36.932	32.004
WNNM^[8]	33.230	30.400	30.030	29.770	32.710	37.260	32.233
DnCNN^[12]	32.930	30.660	30.150	30.030	33.110	38.270	32.525
FFDNet^[21]	33.060	30.720	30.210	30.040	33.130	38.320	32.580
Ours	33.083	30.799	30.222	30.049	33.198	38.389	32.623

表2 不同算法得到的SSIM结果（噪声等级为25）

Table 2 SSIM results obtained by different algorithms (noise level is 25)

Methods	House	Pepper	Ship	Man	Landscape	Airplane	Average
NLM^[5]	0.820	0.820	0.730	0.740	0.820	0.900	0.805
BM3D^[6]	0.860	0.870	0.800	0.810	0.890	0.950	0.863
WNNM^[8]	0.861	0.870	0.802	0.810	0.890	0.960	0.866
DnCNN^[12]	0.961	0.973	0.951	0.940	0.960	0.920	0.951
FFDNet^[21]	0.862	0.879	0.812	0.822	0.896	0.969	0.873
Ours	0.962	0.976	0.954	0.942	0.963	0.940	0.956

表3 不同算法的处理时间（噪声等级为25）单位：s

Table 3 Processing time of different algorithms (noise level is 25)

Methods	$256 × 256$		$256 × 256$	$512 × 512$		$512 × 512$	$418 × 321$		$418 × 321$	Overall mean
Methods	House	Pepper	Mean	Ship	Man	Mean	Landscape	Airplane	Mean	Overall mean
NLM^[5]	0.051	0.034	0.043	0.166	0.157	0.162	0.080	0.094	0.087	0.097
BM3D^[6]	1.000	0.800	0.900	3.600	3.600	3.600	2.200	2.300	2.250	2.250
WNNM^[8]	134.925	131.224	133.075	552.875	596.001	574.438	361.181	323.310	342.246	349.920
DnCNN^[12]	0.176	0.173	0.175	0.678	0.668	0.673	0.417	0.405	0.411	0.419
FFDNet^[21]	0.480	0.336	0.408	1.107	0.904	1.006	0.527	0.499	0.513	0.642
Ours	0.099	0.087	0.093	0.398	0.381	0.389	0.238	0.218	0.228	0.237

表3 不同算法的处理时间（噪声等级为25）单位：s

Table 3 Processing time of different algorithms (noise level is 25)

Methods	$256 × 256$		$256 × 256$	$512 × 512$		$512 × 512$	$418 × 321$		$418 × 321$	Overall mean
Methods	House	Pepper	Mean	Ship	Man	Mean	Landscape	Airplane	Mean	Overall mean
NLM^[5]	0.051	0.034	0.043	0.166	0.157	0.162	0.080	0.094	0.087	0.097
BM3D^[6]	1.000	0.800	0.900	3.600	3.600	3.600	2.200	2.300	2.250	2.250
WNNM^[8]	134.925	131.224	133.075	552.875	596.001	574.438	361.181	323.310	342.246	349.920
DnCNN^[12]	0.176	0.173	0.175	0.678	0.668	0.673	0.417	0.405	0.411	0.419
FFDNet^[21]	0.480	0.336	0.408	1.107	0.904	1.006	0.527	0.499	0.513	0.642
Ours	0.099	0.087	0.093	0.398	0.381	0.389	0.238	0.218	0.228	0.237

图6 不同算法处理House图的结果（噪声等级为25）

Fig.6 Performance comparison of different algorithms on House image (noise level is 25)

图7 不同算法处理Man图的结果（噪声等级为25）

Fig.7 Performance comparison of different algorithms on Man image (noise level is 25)

图8 用于高等级噪声去噪的测试原图

Fig.8 Original test images for high level noise denoising

表4 不同算法得到的PSNR结果（噪声等级为50）单位：dB

Table 4 PSNR results obtained by different algorithms (noise level is 50)

Methods	Starfish	Butterfly	Lena	Room	Desert	Coral	Average
NLM^[5]	22.843	23.853	26.643	24.048	27.211	24.549	24.858
BM3D^[6]	24.842	25.662	28.962	21.099	28.483	25.674	25.785
WNNM^[8]	25.430	26.320	29.250	26.640	28.560	25.920	27.020
DnCNN^[12]	25.703	26.871	29.435	26.884	28.833	26.211	27.323
FFDNet^[21]	25.680	26.920	29.439	27.040	28.840	26.290	27.368
Ours	25.706	26.932	29.452	27.094	28.854	26.303	27.390

表5 不同算法得到的SSIM结果（噪声等级为50）

Table 5 SSIM results obtained by different algorithms (noise level is 50)

Methods	Starfish	Butterfly	Lena	Room	Desert	Coral	Average
NLM^[5]	0.630	0.720	0.680	0.570	0.620	0.600	0.637
BM3D^[6]	0.740	0.820	0.810	0.710	0.730	0.690	0.751
WNNM^[8]	0.760	0.830	0.810	0.710	0.720	0.690	0.753
DnCNN^[12]	0.930	0.950	0.920	0.900	0.850	0.910	0.911
FFDNet^[21]	0.775	0.858	0.821	0.734	0.731	0.727	0.774
Ours	0.932	0.953	0.921	0.920	0.860	0.912	0.916

图9 不同算法处理Room图的结果（噪声等级为50）

Fig.9 Performance comparison of different algorithms on Room image (noise level is 50)

图10 不同算法处理Lena图的结果（噪声等级为50）

Fig.10 Performance comparison of different algorithms on Lena image (noise level is 50)

图11 本文算法处理自然图像的结果

Fig.11 Performance of proposed algorithm on natural images

图12 本文算法处理彩色图像的结果

Fig.12 Performance of proposed algorithm on color images

参考文献 23

[1]	TIAN C, FEI L, ZHENG W, et al. Deep learning on image denoising: an overview[J]. Neural Networks, 2020, 131: 251-275. DOI PMID
[2]	GETIS A, GRIFFITH D A. Comparative spatial filtering in regression analysis[J]. Geographical Analysis, 2002, 34(2): 130-140. DOI URL
[3]	ARMSTRONG J. Peak-to-average power reduction for OFDM by repeated clipping and frequency domain filtering[J]. Electronics Letters, 2002, 38(5): 246-247. DOI URL
[4]	CANDAN C, KUTAY M A, OZAKTAS H M. The discrete fractional Fourier transform[J]. IEEE Transactions on Signal Processing, 2000, 48(5): 1329-1337. DOI URL
[5]	BUADES A, COLL B, MOREL J M. A non-local algorithm for image denoising[C]// Proceedings of the 2005 IEEE Com-puter Society Conference on Computer Vision and Pattern Recognition, San Diego, Jun 20-26, 2005. Washington: IEEE Computer Society, 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. PMID
[7]	KNAUS C, ZWICKER M. Progressive image denoising[J]. IEEE Transactions on Image Processing, 2014, 23(7): 3114-3125. PMID
[8]	GU S, ZHANG L, ZUO W, et al. Weighted nuclear norm minimization with application to image denoising[C]// Pro-ceedings of the 2014 IEEE Conference on Computer Vision and Pattern Recognition, Columbus, Jun 23-28, 2014. Washington: IEEE Computer Society, 2014: 2862-2869.
[9]	陈强, 郑钰辉, 孙权森, 等. 片相似性各项异性扩散图像去噪[J]. 计算机研究与发展, 2010, 47(1): 33-42.
	CHEN Q, ZHENG Y H, SUN Q S, et al. Patch similarity based anisotropic diffusion for image denoising[J]. Journal of Computer Research and Development, 2010, 47(1): 33-42.
[10]	窦诺, 赵瑞珍, 岑翼刚, 等. 基于稀疏表示的含噪图像超分辨重建方法[J]. 计算机研究与发展, 2015, 52(4): 943-951.
	DOU N, ZHAO R Z, CEN Y G, et al. Superresolution reconstruction of noisy images based on sparse representation[J]. Journal of Computer Research and Development, 2015, 52(4): 943-951.
[11]	王洪雁, 王拓, 潘勉, 等. 基于伽马范数最小化的图像去噪算法[J]. 通信学报, 2020, 41(10): 222-230. DOI
	WANG H Y, WANG T, PAN M, et al. Image denoising algorithm based on gamma norm minimization[J]. Journal of Communications, 2020, 41(10): 222-230.
[12]	ZHANG K, ZUO W, CHEN Y, et al. Beyond a Gaussian denoiser: residual learning of deep CNN for image denoising[J]. IEEE Transactions on Image Processing, 2017, 26(7): 3142-3155. DOI PMID
[13]	LEFKIMMIATIS S. Universal denoising networks: a novel CNN architecture for image denoising[C]// Proceedings of the 2018 IEEE Conference on Computer Vision and Pattern Recognition, Salt Lake City, Jun 18-22, 2018. Washington: IEEE Computer Society, 2018: 3204-3213.
[14]	HE K, ZHANG X, REN S, et al. Deep residual learning for image recognition[C]// Proceedings of the 2016 IEEE Con-ference on Computer Vision and Pattern Recognition, Las Vegas, Jun 26-Jul 1, 2016. Washington: IEEE Computer Society, 2016: 770-778.
[15]	HUANG G, LIU S, VAN DER MAATEN L, et al. Condensenet: an efficient densenet using learned group convolutions[C]// Proceedings of the 2018 IEEE Conference on Computer Vision and Pattern Recognition, Salt Lake City, Jun 18-22, 2018. Washington: IEEE Computer Society, 2018: 2752-2761.
[16]	HUANG G, LIU Z, VAN DER MAATEN L, et al. Densely connected convolutional networks[C]// Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, Jul 21-26, 2017. Washington: IEEE Computer Society, 2017: 4700-4708.
[17]	郭恒意, 贾振堂. 结合残差密集块的卷积神经网络图像去噪方法[J]. 计算机工程与设计, 2020, 41(7): 1998-2003.
	GUO H Y, JIA Z T. Image denoising method based on convolutional neural network combined with residual-dense block[J]. Computer Engineering and Design, 2020, 41(7): 1998-2003.
[18]	ZHANG Y, TIAN Y, KONG Y, et al. Residual dense network for image restoration[J]. IEEE Transactions on Pattern Analy-sis and Machine Intelligence, 2021, 43(7): 2480-2495.
[19]	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 evaluation model based on PSNR and SSIM[J]. Journal of Image and Graphics, 2006, 11(12): 1758-1763.
[20]	HAHN J, HAUSMAN J, KUERSTEINER G. Estimation with weak instruments: accuracy of higher-order bias and MSE approximations[J]. The Econometrics Journal, 2004, 7(1): 272-306. DOI URL
[21]	ZHANG K, ZUO W, 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. DOI URL
[22]	CHENG Y, LIU Z. Image denoising algorithm based on structure and texture part[C]// Proceedings of the 2016 12th International Conference on Computational Intelligence and Security, Wuxi, Dec 16-19, 2016. Piscataway: IEEE, 2016: 147-151.
[23]	GNANADURAI D, SADASIVAM V. An efficient adaptive thresholding technique for wavelet based image denoising[J]. International Journal of Signal Processing, 2006, 2(2): 114-119.

DnRFD：用于图像去噪的递进式残差融合密集网络

DnRFD：Progressive Residual Fusion Dense Network for Image Denoising

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 17

参考文献 23

相关文章 15

编辑推荐

Metrics

[1]	张璐, 芦天亮, 杜彦辉. 人脸视频深度伪造检测方法综述[J]. 计算机科学与探索, 2023, 17(1): 1-26.
[2]	王仕宸, 黄凯, 陈志刚, 张文东. 深度学习的三维人体姿态估计综述[J]. 计算机科学与探索, 2023, 17(1): 74-87.
[3]	梁佳利, 华保健, 吕雅帅, 苏振宇. 面向深度学习算子的循环不变式外提算法[J]. 计算机科学与探索, 2023, 17(1): 127-139.
[4]	王剑哲, 吴秦. 坐标注意力特征金字塔的显著性目标检测算法[J]. 计算机科学与探索, 2023, 17(1): 154-165.
[5]	杨才东, 李承阳, 李忠博, 谢永强, 孙方伟, 齐锦. 深度学习的图像超分辨率重建技术综述[J]. 计算机科学与探索, 2022, 16(9): 1990-2010.
[6]	张祥平, 刘建勋. 基于深度学习的代码表征及其应用综述[J]. 计算机科学与探索, 2022, 16(9): 2011-2029.
[7]	李冬梅, 罗斯斯, 张小平, 许福. 命名实体识别方法研究综述[J]. 计算机科学与探索, 2022, 16(9): 1954-1968.
[8]	任宁, 付岩, 吴艳霞, 梁鹏举, 韩希. 深度学习应用于目标检测中失衡问题研究综述[J]. 计算机科学与探索, 2022, 16(9): 1933-1953.
[9]	吕晓琦, 纪科, 陈贞翔, 孙润元, 马坤, 邬俊, 李浥东. 结合注意力与循环神经网络的专家推荐算法[J]. 计算机科学与探索, 2022, 16(9): 2068-2077.
[10]	安凤平, 李晓薇, 曹翔. 权重初始化-滑动窗口CNN的医学图像分类[J]. 计算机科学与探索, 2022, 16(8): 1885-1897.
[11]	曾凡智, 许露倩, 周燕, 周月霞, 廖俊玮. 面向智慧教育的知识追踪模型研究综述[J]. 计算机科学与探索, 2022, 16(8): 1742-1763.
[12]	刘艺, 李蒙蒙, 郑奇斌, 秦伟, 任小广. 视频目标跟踪算法综述[J]. 计算机科学与探索, 2022, 16(7): 1504-1515.
[13]	赵小明, 杨轶娇, 张石清. 面向深度学习的多模态情感识别研究进展[J]. 计算机科学与探索, 2022, 16(7): 1479-1503.
[14]	夏鸿斌, 肖奕飞, 刘渊. 融合自注意力机制的长文本生成对抗网络模型[J]. 计算机科学与探索, 2022, 16(7): 1603-1610.
[15]	孙方伟, 李承阳, 谢永强, 李忠博, 杨才东, 齐锦. 深度学习应用于遮挡目标检测算法综述[J]. 计算机科学与探索, 2022, 16(6): 1243-1259.