|
发布时间: 2019-08-16 |
学者观点 |
|
|
收稿日期: 2019-04-29; 修回日期: 2019-05-21
基金项目: 国家自然科学基金项目(61662044,61163023,51765042);江西省自然科学基金项目(20171BAB202017)
第一作者简介:
徐少平, 1976年生, 男, 教授, 博士生导师, 主要研究方向为图形图像处理技术、机器视觉、虚拟手术仿真。E-mail:xushaoping@ncu.edu.cn;
刘婷云, 女, 硕士研究生, 主要研究方向为图形图像处理和机器视觉。E-mail:416114517210@email.ncu.edu.cn; 林珍玉, 女, 硕士研究生, 主要研究方向为图形图像处理和机器视觉。E-mail:401030918076@email.ncu.edu.cn; 张贵珍, 女, 硕士研究生, 主要研究方向为图形图像处理和机器视觉。E-mail:406130917331@email.ncu.edu.cn; 李崇禧, 男, 硕士研究生, 主要研究方向为图形图像处理和机器视觉。E-mail:406130917315@email.ncu.edu.cn.
中图法分类号: TP391.41
文献标识码: A
文章编号: 1006-8961(2019)08-1207-08
|
摘要
现有的深度卷积神经网络(DCNN)图像降噪模型受其技术路线内在固有特性的制约,降噪性能仍然有待进一步改进。为了推动现有DCNN图像降噪模型技术的发展,需要正视并及时解决制约其进一步完善的瓶颈问题。本文简要概述了传统的基于自然图像非局部自相似性、稀疏性和低秩性这3种先验知识设计的图像降噪算法的技术路线特点和优缺点,从传统图像降噪算法存在的问题中引出基于DCNN构建图像降噪模型的技术优势,并梳理并总结了DCNN降噪模型未来的发展瓶颈,就相应的解决方案(研究方向)进行详细讨论。通过深入分析发现,可以从扩大卷积核的感受野、降低网络参数与训练集之间的依赖关系以及充分利用DCNN网络的建模能力这3个角度入手,突破现有基于数据驱动的DCNN降噪模型的瓶颈制约,把图像降噪算法的研究水平推向新的高度。
关键词
综述; 图像降噪; 深度卷积神经网络; 瓶颈问题; 感受野; 数据依赖; 参数空间
Abstract
As a representative technique for deep learning, the deep convolution neural network (DCNN) with strong feature learning and nonlinear mapping ability in the field of digital image processing offers a novel opportunity for image denoising research. DCNN-based denoising models show significant advantages over traditional methods in terms of their denoising effect and execution efficiency. However, most of the existing image denoising models are driven by data. Given their inherent restrictions, the denoising performance of these models can be further improved. To promote the development of existing image denoising technologies, some key challenges that restrict their further improvement must be analyzed and addressed. We first summarize the core ideas of traditional image denoising algorithms based on three types of prior knowledge of the natural image, namely, non-local self-similarity, sparsity, and low rank, and then analyze the advantages and disadvantages of these algorithms. Image denoising algorithms modeled with prior knowledge can flexibly deal with distorted images under different noise levels. Unfortunately, they demonstrate the following limitations:1) the limited hand-crafted image priors are not enough to describe all changes in the image structure, thereby limiting the denoising ability of these algorithms; 2) most of the traditional image denoising algorithms iteratively solve their objective functions, thereby resulting in a high computational complexity; and 3) the optimal solution of the objective function needs to adjust several parameters manually according to the actual situation. Based on the above problems, we point out that the technical advantage of the DCNN-based denoising model lies in its strong nonlinear approximation supported by a graphics processing unit. The inherent characteristics of the DCNN-based denoising model are then analyzed, the bottleneck problems that restrict their future development are presented, and the possible solutions (research directions) to these problems are discussed in detail. A thorough analysis reveals many bottleneck problems in data-driven DCNN-based denoising models that need to be solved, including:1) the small receptive field of the DCNN network that limits the range of image feature representation and the ability to fully utilize the priors contained in natural images; 2) the strong dependence of DCNN-based model parameters on the training dataset, that is, the optimal denoising effect can only be obtained if the distortion level of the observed image is close to that of the training images; and 3) the training set cannot be easily constructed and the denoising model cannot be easily trained to denoise an image if both the noise type and level of noisy image are unknown. To solve these problems, we expand the receptive field of the convolution kernel, weaken the dependency between the network parameters and the training set, or fully utilize the modeling ability of the DCNN network. Therefore, the bottlenecks of the existing DCNN denoising models can be addressed, and research on image denoising algorithms can move to a higher level. In this paper, the technical advantages and development bottlenecks of DCNN in the field of image denoising are summarized, and some future research directions for the image denoising method are proposed. This paper should be of interest to readers in the area of image denoising.
Key words
review; image denoising; deep convolutional neural network(DCNN); bottleneck problem; receptive field; data dependencies; parameter space
0 引言
一般来讲,图像受到噪声干扰后的退化模型可以定义为
$ \mathit{\boldsymbol{y}}{\rm{ = }}\mathit{\boldsymbol{x}}{\rm{ + }}\mathit{\boldsymbol{n}} $ | (1) |
式中,
$ \mathit{\boldsymbol{\hat x}} = \arg \;\mathop {\min }\limits_x \left\{ {\frac{1}{2}\mathit{\boldsymbol{y}} - \mathit{\boldsymbol{x}}_2^2 + \lambda \mathit{\boldsymbol{ \boldsymbol{\varPhi} }}(\mathit{\boldsymbol{x}})} \right\} $ | (2) |
最优值的方式实现降噪,即求解由保真项
1) 基于自然图像稀疏表示和低秩所构建的图像先验知识不足以描述所有复杂的图像结构变化,没有充分挖掘出图像中的先验信息,限制了图像的降噪效果;
2) 目标函数的求解通常以复杂的迭代优化过程实现,使得算法时间复杂度较高;
3) 式(2)所定义的目标函数一般是非凸的(non-convex),具体实现时需要手动调整若干个参数才能获得最优的结果。
近年来,深度学习(deep learning)技术因其强大的特征学习和非线性映射能力在图像降噪领域取得了巨大成功[20-29],其中尤以基于深度卷积神经网络(DCNN)构建的图像降噪模型发展迅速。基于DCNN构建的降噪模型[28]通过在大量噪声图像—无失真图像训练数据集上以最小化估计图像
$ \mathit{\boldsymbol{\hat x}} = {\cal F}(\mathit{\boldsymbol{y}};\mathit{\boldsymbol{ \boldsymbol{\varTheta} }})\quad {\rm{ s}}{\rm{.t}}{\rm{. }}\quad \mathop {\min }\limits_{\bf{\Theta }} {\cal L}(\mathit{\boldsymbol{\hat x}}, \mathit{\boldsymbol{x}}) $ | (3) |
基于DCNN的降噪模型依赖网络结构隐式地学习图像中的先验知识,展现出了强大的图像先验知识建模能力(非线性映射能力),能够避免基于稀疏和低秩优化降噪模型中正则项
1 现存问题及研究展望
基于数据驱动的技术路线使得DCNN降噪模型可以获得远胜于其他主流降噪算法的降噪效果,但同时也存在以下瓶颈问题亟待解决:
1) 图像特征提取范围小的问题。DCNN网络依赖大量的卷积核提取图像特征用于后续功能模块处理,然而现有的DCNN网络的卷积核大小通常都设置得比较小,即所谓的感受野(receptive field)非常局限,导致获取的图像特征表征范围有限,不能充分利用自然图像中广泛存在的NSS特性提高降噪效果。
2) 待降噪图像与降噪模型的相互匹配问题。DCNN降噪模型要想获得最佳的降噪效果,必须根据图像降质(失真)的严重程度选用预先训练好的特定DCNN降噪模型才能最大程度地发挥其降噪能力。其本质的原因在于通过训练获得的降噪模型的网络参数
3) 训练数据缺乏问题。现有的DCNN降噪模型常常将噪声模型简化为加性高斯噪声,然而,在现实噪声图像中的噪声往往并不严格符合高斯分布。在待降噪图像噪声信息未知的情况下,构建训练数据集非常困难,导致无法训练降噪模型完成降噪任务。为了解决这些问题,未来可能的解决方案(研究方向)详细描述如下。
1.1 扩大卷积核的感受野
自然图像在内容上存在显著的NSS非局部自相似性,研究者利用这个特性或将这个特性与图像的稀疏和低秩特性结合,实现了很多降噪性能不错的降噪算法[16-18, 30-33]。
目前DCNN网络中的卷积核通常都设置得非常小,仅仅能够提取图像中非常有限的局部特征信息,其处理视野有待进一步扩大。为扩大DCNN网络的感受野,常规的策略是增大卷积核尺寸和增加网络深度。然而,增大卷积核的尺寸会引入大量的网络参数并导致网络训练难度增大,同时执行效率也将降低。另外,设计层次非常深的神经网络结构容易出现过拟合和非凸优化问题。最近提出的膨胀卷积(dilated convolution)[34]、沙漏状结构(hourglass-shaped architectures)[35]等技术可以达到增大感受野的目的。其实,引入图像的NSS非局部自相似性是扩大DCNN网络卷积视野更好的方式,有望获得更高的性能提升。
最近这方面探索性的工作有:Ahn等人[36]提出了一种基于NSS先验知识的块匹配卷积神经网络(BMCNN)算法。该算法首先采用现有的降噪算法(如3维块匹配算法(BM3D)、去噪卷积神经网络(DnCNN)[24])对噪声图像进行预处理获得对应的参考图像,然后从噪声图像及参考图像中找到相似的图块,并将它们堆叠成与BM3D算法类似的3维矩阵,最后通过DnCNN网络架构完成图块降噪。BMCNN算法可以看做是传统基于非局部自相似性的算法的扩展,其以数据驱动方式训练的框架能够学习更精确的降噪模型。与现有的各种DCNN的降噪模型相比,在图像结构比较丰富的自然图像上能获得更好的降噪性能。然而,BMCNN算法的性能受限于获取参考图像的降噪算法,所获得的参考图像与原始无失真图像毕竟尚有差距,而且也只使用了1幅参考图像,从而会影响到相似图块搜索的正确性,在一定程度上制约了其性能。类似地,在文献[37]中,Cruz等人将基于卷积神经网络的降噪模型与基于非局部自相似性的传统降噪算法级联组合使用。总之,上述两个工作初步实现了在DCNN网络中结合NSS非局部自相似性,进一步提高了DCNN网络模型的降噪效果,但是上述工作还存在一定的发展空间。
由上述DCNN网络降噪模型的发展态势可以看出:将自然图像的NSS特性集成到DCNN模型中,有效地增大DCNN网络的感受野是大势所趋,而将传统的基于NSS的降噪技术完全无缝地集成到DCNN框架下才能最大限度地利用各自的优势。近期,Yang等人[38]就传统降噪技术CNN网络化的问题进行了探索,他们将BM3D算法的执行过程等价展开为一个卷积神经网络结构,提出了一种称为BM3D-Net的降噪模型。BM3D-Net网络模型依次由提取层(extraction)、卷积层、非线性变换层、卷积层和聚集层(aggregation)堆叠而成。其中,提取层对应于BM3D中的块匹配操作(block-matching),卷积、非线性变换和卷积这3层实现了BM3D算法中的3维小波变换收缩(3D wavelet transform shrinkage),聚集层则实现了BM3D中的块聚集(patch group aggregation)操作。BM3D-Net算法获得了比传统BM3D算法更佳的降噪性能,且能够在GPU上执行,从而可以获得极高的执行效率。但该网络模型仅包含2个卷积层,其非线性映射能力有待进一步加强。尽管如此,文献[38]的工作表明:传统的基于NSS特性的降噪算法的处理过程完全可以在网络化和模块化后作为嵌入网络模块(ENB)集成到现有的DCNN网络中。ENB模块可以看做卷积神经网络结构中类似于卷积层的一种新型网络结构,将其集成到DCNN网络架构中可增大DCNN网络卷积处理视野,实现一个端到端(end-to-end)的非局部处理网络,从而进一步提高现有DCNN网络的降噪能力。ENB模块具体的设计与实现将是一个未来可以积极探索的研究方向。
1.2 降低网络参数$\mathit{\pmb{\Theta}}$ 与训练集之间的依赖关系
基于数据驱动的DCNN降噪模型充分利用了从训练集大量数据中学习到的先验知识,其降噪能力相比现有主流降噪算法而言提升非常显著。然而,通过训练获得的先验知识泛化能力较弱,只有当用于训练降噪模型的图像集合与待降噪图像在降质程度(高斯噪声模型时即为噪声水平值)近似时才能获得最好的降噪效果。换句话说,现有的DCNN网络中的网络参数
因此,在噪声图像的基础上通过联结噪声水平映射图
1.3 充分利用DCNN网络的建模能力
现有基于数据驱动的DCNN降噪模型都需要在相当庞大的数据集上完成训练才可以投入使用。在具体实践中,常常仅有待降噪图像一幅图像可以利用,而且其噪声分布是未知的,并不属于常用且简化的高斯模型。这意味着对于这种待降噪图像,无法事先生成合适的噪声图像训练集,训练降噪模型也无从谈起,故在这种情况下基于训练策略的DCNN降噪模型将失效。
为解决训练数据缺乏的问题,关键需要研究如何仅依赖单幅的噪声图像获得它的噪声分布模型,有了噪声分布模型就可以实现在选定的无失真图像集合上加入根据噪声分布模型生成的噪声信号,从而生成用于训练的噪声图像训练集。然而仅依赖单张的噪声图像获得它的噪声分布模型是非常困难的,这是个典型的欠约束问题。幸运的是自然图像像素点亮度值的分布是非常有规律的,而且噪声图像的分布模型显著异于正常的无失真图像。目前,利用解析的方法获得噪声图像的分布模型效果并不理想。考虑到神经网络强大的非线性逼近能力,已有一些研究学者基于神经网络技术开展了这方面的工作。例如,Chen等人[39]所提出的GCBD(GAN-CNN based blind denoiser)算法试图利用具有复杂分布估计能力的生成对抗网络(GAN)隐式地学习给定噪声图像的噪声分布模型。然后,基于噪声模型构造噪声图像训练数据集。最后,再利用文献[24]提出的DnCNN网络结构训练降噪模型实现盲降噪任务。Chen等人提出的GCBD算法可以直接基于给定的单幅噪声图像完成降噪任务,克服了显式定义未知噪声模型的困难,避免了模型对图像训练集的依赖,其所采用的技术路线对未来的研究工作具有很好的借鉴价值。然而,该方法仅局限于对加性噪声有效,在噪声模型方面仍然存在一定的限制。未来需要将GCBD模型扩展到能够处理更为真实的噪声(real-world noise),以增强降噪模型的鲁棒性和普适性。
鉴于DCNN强大的非线性生成(映射)能力,研究者未来还可以尝试另外一种完全摆脱训练数据集的降噪策略[40]。也就是说,基于随机初始化网络参数的DCNN网络和给定的单张噪声图像,通过搜索DCNN的参数空间
2 结论
综上所述,基于DCNN的降噪模型近几年已经在图像降噪领域获得了巨大的成功,然而这种基于数据驱动的降噪模型在发展过程中也遇到了其本身技术特点所带来的结构性瓶颈问题亟需解决。
对于DCNN的小感受野问题,可以将基于NSS特性实现的降噪算法处理过程网络化后集成到DCNN来解决;对于降噪模型的网络参数
参考文献
-
[1] Chen L, Yang G, Chen Z Y, et al. Linearized Bregman iteration algorithm for matrix completion with structural noise[J]. Chinese Journal of Computers, 2015, 38(7): 1357–1371. [陈蕾, 杨庚, 陈正宇, 等. 基于线性Bregman迭代的结构化噪声矩阵补全算法[J]. 计算机学报, 2015, 38(7): 1357–1371. ] [DOI:10.11897/SP.J.1016.2015.01357]
-
[2] Shen J M, Li J F, Dai W Z. No-reference image quality assessment based on structure information and luminance statistics[J]. Acta Electronica Sinica, 2016, 44(4): 804–812. [沈军民, 李俊峰, 戴文战. 结合结构信息和亮度统计的无参考图像质量评价[J]. 电子学报, 2016, 44(4): 804–812. ] [DOI:10.3969/j.issn.0372-2112.2016.04.008]
-
[3] Xu S P, Zeng X X, Tang Y L, et al. A noise level estimation algorithm using prior knowledge of similar images[J]. Journal of Computer Research and Development, 2018, 55(12): 2741–2752. [徐少平, 曾小霞, 唐祎玲, 等. 基于多图像先验知识的噪声水平评估算法[J]. 计算机研究与发展, 2018, 55(12): 2741–2752. ] [DOI:10.7544/issn1000-1239.2018.20170336]
-
[4] Xu S P, Zeng X X, Tang Y L. Fast noise level estimation algorithm based on two-stage support vector regression[J]. Journal of Computer-Aided Design & Computer Graphics, 2018, 30(3): 447–458. [徐少平, 曾小霞, 唐祎玲. 基于两阶段支持向量回归的快速噪声水平估计算法[J]. 计算机辅助设计与图形学学报, 2018, 30(3): 447–458. ] [DOI:10.3724/SP.J.1089.2018.16422]
-
[5] Gong X C, Li Z M. An image denoising method using deep asymmetrical skip connection[J]. Journal of Computer-Aided Design & Computer Graphics, 2019, 31(2): 295–302. [公绪超, 李宗民. 深度非对称跳跃连接的图像降噪方法[J]. 计算机辅助设计与图形学学报, 2019, 31(2): 295–302. ] [DOI:10.3724/SP.J.1089.2019.17167]
-
[6] Smith S M, Brady J M. SUSAN——a new approach to low level image processing[J]. International Journal of Computer Vision, 1997, 23(1): 45–78. [DOI:10.1023/a:1007963824710]
-
[7] Tomasi C, Manduchi R. Bilateral filtering for gray and color images[C]//Proceedings of the 6th International Conference on Computer Vision. Bombay, India: IEEE, 1998: 839-846.[DOI: 10.1109/iccv.1998.710815]
-
[8] Aharon M, Elad M, Bruckstein A. K-SVD:an algorithm for designing overcomplete dictionaries for sparse representation[J]. IEEE Transactions on Signal Processing, 2006, 54(11): 4311–4322. [DOI:10.1109/tsp.2006.881199]
-
[9] Zhang L, Dong W S, Zhang D, et al. Two-stage image denoising by principal component analysis with local pixel grouping[J]. Pattern Recognition, 2010, 43(4): 1531–1549. [DOI:10.1016/j.patcog.2009.09.023]
-
[10] Tang C, Xu J L, Zhou Z G. Strong noise image-denoising algorithm based on improved curvature filters[J]. Journal of Image and Graphics, 2019, 24(3): 346–356. [汤成, 许建龙, 周志光. 改进的曲率滤波强噪声图像去噪方法[J]. 中国图象图形学报, 2019, 24(3): 346–356. ] [DOI:10.11834/jig.180302]
-
[11] Xu S P, Zhang X Q, Jiang Y N, et al. Noise level estimation based on local means and its application to the blind BM3D denoising algorithm[J]. Journal of Image and Graphics, 2017, 22(4): 422–434. [徐少平, 张兴强, 姜尹楠, 等. 局部均值噪声估计的盲3维滤波降噪算法[J]. 中国图象图形学报, 2017, 22(4): 422–434. ] [DOI:10.11834/jig.20170402]
-
[12] Remez T, Litany O, Giryes R, et al. Class-aware fully convolutional Gaussian and Poisson denoising[J]. IEEE Transactions on Image Processing, 2018, 27(11): 5707–5722. [DOI:10.1109/TIP.2018.2859044]
-
[13] Buades A, Coll B, Morel J M. A non-local algorithm for image denoising[C]//Proceedings of 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. San Diego, CA, USA: IEEE, 2005: 60-65.[DOI: 10.1109/CVPR.2005.38]
-
[14] 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:10.1109/TIP.2007.901238]
-
[15] Mairal J, Bach F, Ponce J, et al. Non-local sparse models for image restoration[C]//Proceedings of the 12th International Conference on Computer Vision. Kyoto, Japan: IEEE, 2009: 2272-2279.[DOI: 10.1109/ICCV.2009.5459452]
-
[16] Dong W S, Zhang L, Shi G M, et al. Nonlocally centralized sparse representation for image restoration[J]. IEEE Transactions on Image Processing, 2013, 22(4): 1620–1630. [DOI:10.1109/TIP.2012.2235847]
-
[17] Zhang J, Zhao D B, Gao W. Group-based sparse representation for image restoration[J]. IEEE Transactions on Image Processing, 2014, 23(8): 3336–3351. [DOI:10.1109/TIP.2014.2323127]
-
[18] Dong W S, Shi G M, Li X. Nonlocal image restoration with bilateral variance estimation:a low-rank approach[J]. IEEE Transactions on Image Processing, 2013, 22(2): 700–711. [DOI:10.1109/TIP.2012.2221729]
-
[19] Gu S H, Zhang L, Zuo W M, et al. Weighted nuclear norm minimization with application to image denoising[C]//Proceedings of 2014 IEEE Conference on Computer Vision and Pattern Recognition. Columbus, OH, USA: IEEE, 2014: 2862-2869.[DOI: 10.1109/CVPR.2014.366]
-
[20] Xie J Y, Xu L L, Chen E H. Image denoising and inpainting with deep neural networks[C]//Proceedings of the 25th International Conference on Neural Information Processing Systems. Lake Tahoe, Nevada: ACM, 2012: 341-349. http://cn.bing.com/academic/profile?id=8008a6f168a5fee75d5682204f849790&encoded=0&v=paper_preview&mkt=zh-cn
-
[21] Burger H C, Schuler C J, Harmeling S. Image denoising: can plain neural networks compete with BM3D?[C]//Proceedings of 2012 IEEE Conference on Computer Vision and Pattern Recognition. Providence, RI, USA: IEEE, 2012: 2392-2399.[DOI: 10.1109/CVPR.2012.6247952]
-
[22] Santhanam V, Morariu V I, Davis L S. Generalized deep image to image regression[C]//Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition. Honolulu, HI, USA: IEEE, 2017: 5395-5405.[DOI: 10.1109/CVPR.2017.573]
-
[23] Zhang K, Zuo W M, Gu S H, et al. Learning deep CNN denoiser prior for image restoration[C]//Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition. Honolulu, HI, USA: IEEE, 2017: 2808-2817.[DOI: 10.1109/CVPR.2017.300]
-
[24] 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(7): 3142–3155. [DOI:10.1109/TIP.2017.2662206]
-
[25] Du B, Xiong W, Wu J, et al. Stacked convolutional denoising auto-encoders for feature representation[J]. IEEE Transactions on Cybernetics, 2017, 47(4): 1017–1027. [DOI:10.1109/TCYB.2016.2536638]
-
[26] Wu C Z, Chen X, Ji D, et al. Image denoising via residual network based on perceptual loss[J]. Journal of Image and Graphics, 2018, 23(10): 1483–1491. [吴从中, 陈曦, 季栋, 等. 结合深度残差学习和感知损失的图像去噪[J]. 中国图象图形学报, 2018, 23(10): 1483–1491. ] [DOI:10.11834/jig.180069]
-
[27] Zhang J S, Zhang Y, Bai L F, et al. Lossless-constraint denoising based auto-encoders[J]. Signal Processing:Image Communication, 2018, 63: 92–99. [DOI:10.1016/j.image.2018.02.002]
-
[28] 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. [DOI:10.1109/TIP.2018.2839891]
-
[29] Plötz T, Roth S. Neural nearest neighbors networks[EB/OL]. 2018-10-30[2019-04-04]. http://arxiv.org/pdf/1810.12575.pdf.
-
[30] Liu H F, Xiong R Q, Ma S W, et al. Non-local extension of total variation regularization for image restoration[C]//Proceedings of 2014 IEEE International Symposium on Circuits and Systems. Melbourne VIC, Australia: IEEE, 2014: 1102-1105.[DOI: 10.1109/ISCAS.2014.6865332]
-
[31] Xu J, Zhang L, Zuo W M, et al. Patch group based nonlocal self-similarity prior learning for image denoising[C]//Proceedings of 2015 IEEE International Conference on Computer Vision. Santiago, Chile: IEEE, 2015: 244-252.[DOI: 10.1109/ICCV.2015.36]
-
[32] Xu S P, Yang X H, Jiang S L. A fast nonlocally centralized sparse representation algorithm for image denoising[J]. Signal Processing, 2017, 131: 99–112. [DOI:10.1016/j.sigpro.2016.08.006]
-
[33] Jin K H, Ye J C. Sparse and low-rank decomposition of a Hankel structured matrix for impulse noise removal[J]. IEEE Transactions on Image Processing, 2018, 27(3): 1448–1461. [DOI:10.1109/TIP.2017.2771471]
-
[34] Yu F, Koltun V. Multi-scale context aggregation by dilated convolutions[EB/OL]. 2016-04-30[2019-04-04]. http://arxiv.org/pdf/1511.07122.pdf.
-
[35] Shelhamer E, Long J, Darrell T. Fully convolutional networks for semantic segmentation[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(4): 640–651. [DOI:10.1109/TPAMI.2016.2572683]
-
[36] Ahn B, Cho N I. Block-matching convolutional neural network for image denoising[EB/OL]. 2017-04-03[2019-04-04]. http://arxiv.org/pdf/1704.00524.pdf.
-
[37] Cruz C, Foi A, Katkovnik V, et al. Nonlocality-reinforced convolutional neural networks for image denoising[J]. IEEE Signal Processing Letters, 2018, 25(8): 1216–1220. [DOI:10.1109/LSP.2018.2850222]
-
[38] Yang D, Sun J. BM3D-Net:a convolutional neural network for transform-domain collaborative filtering[J]. IEEE Signal Processing Letters, 2018, 25(1): 55–59. [DOI:10.1109/lsp.2017.2768660]
-
[39] Chen J W, Chen J W, Chao H Y, et al. Image blind denoising with generative adversarial network based noise modeling[C]//Proceedings of 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City, UT, USA: IEEE, 2018: 3155-3164.[DOI: 10.1109/CVPR.2018.00333]
-
[40] Ulyanov D, Vedaldi A, Lempitsky V. Deep image prior[EB/OL]. 2018-04-05[2019-04-04]. http://arxiv.org/pdf/1711.10925.pdf.