1 Introduction
Image denoising is an important research topic in computer vision, aiming at recovering the underlying clean image from an observed noisy one. The noise contained in a real noisy image is generally accumulated from multiple different sources, e.g., capturing instruments, data transmission media, image quantization, etc. [36]. Such complicated generation process makes it fairly difficult to access the noise information accurately and recover the underlying clean image from the noisy one. This constitutes the main aim of blind image denoising.
There are two main categories of image denoising methods. Most classical methods belong to the first category, mainly focusing on constructing a rational maximum a posteriori (MAP) model, involving the fidelity (loss) and regularization terms, from a Bayesian perspective [6]. An understanding for data generation mechanism is required for designing a rational MAP objective, especially better image priors like sparsity [3], lowrankness [16, 44, 38], and nonlocal similarity [9, 26]. These methods are superior mainly in their interpretability naturally led by the Bayesian framework. They, however, still exist critical limitations due to their assumptions on both image prior and noise (generally i.i.d. Gaussian), possibly deviating from real spatially variant (i.e.,noni.i.d.) noise, and their relatively low implementation speed since the algorithm needs to be reimplemented for any new coming image. Recently, deep learning approaches represent a new trend along this research line. The main idea is to firstly collect large amount of noisyclean image pairs and then train a deep neural network denoiser on these training data in an endtoend learning manner. This approach is especially superior in its effective accumulation of knowledge from large datasets and fast denoising speed for test images. They, however, are easy to overfit to the training data with certain noisy types, and still could not be generalized well on test images with unknown but complicated noises.
Thus, blind image denoising especially for real images is still a challenging task, since the real noise distribution is difficult to be preknown (for modeldriven MAP approaches) and hard to be comprehensively simulated by training data (for datadriven deep learning approaches).
Against this issue, this paper proposes a new variational inference method, aiming at directly inferring both the underlying clean image and the noise distribution from an observed noisy image in a unique Bayesian framework. Specifically, an approximate posterior is presented by taking the intrinsic clean image and noise variances as latent variables conditioned on the input noisy image. This posterior provides explicit parametric forms for all its involved hyperparameters, and thus can be efficiently implemented for blind image denoising with automatic noise estimation for test noisy images.
In summary, this paper mainly makes following contributions: 1) The proposed method is capable of simultaneously implementing both noise estimation and blind image denoising tasks in a unique Bayesian framework. The noise distribution is modeled as a general noni.i.d. configurations with spatial relevance across the image, which evidently better complies with the heterogeneous real noise beyond the conventional i.i.d. noise assumption. 2) Succeeded from the fine generalization capability of the generative model, the proposed method is verified to be able to effectively estimate and remove complicated noni.i.d. noises in test images even though such noise types have never appeared in training data, as clearly shown in Fig. 3
. 3) The proposed method is a generative approach outputted a complete distribution revealing how the noisy image is generated. This not only makes the result with more comprehensive interpretability beyond traditional methods purely aiming at obtaining a clean image, but also naturally leads to a learnable likelihood (fidelity) term according to the dataself. 4) The most commonly utilized deep learning paradigm, i.e., taking MSE as loss function and training on large noisyclean image pairs, can be understood as a degenerated form of the proposed generative approach. Their overfitting issue can then be easily explained under this variational inference perspective: these methods intrinsically put dominant emphasis on fitting the priors of the latent clean image, while almost neglects the effect of noise variations. This makes them incline to overfit noise bias on training data and sensitive to the distinct noises in test noisy images.
The paper is organized as follows: Section 2 introduces related work. Sections 3 presents the proposed full Bayesion model, the deep variational inference algorithm, the network architecture and some discussions. Section 4 demonstrates experimental results and the paper is finally concluded.
2 Related Work
We present a brief review for the two main categories of image denoising methods, i.e., modeldriven MAP based methods and datadriven deep learning based methods.
Modeldriven MAP based Methods: Most classical image denoising methods belong to this category, through designing a MAP model with a fidelity/loss term and a regularization one delivering the preknown image prior. Along this line, total variation denoising [34], anisotropic diffusion [29] and wavelet coring [35] use the statistical regularities of images to remove the image noise. Later, the nonlocal similarity prior, meaning many small patches in a nonlocal image area possess similar configurations, was widely used in image denoising. Typical ones include CBM3D [11] and nonlocal means [9]. Some dictionary learning methods [16, 13, 38] and FieldofExperts (FoE) [33], also revealing certain prior knowledge of image patches, had also been attempted for the task. Several other approaches focusing on the fidelity term, which are mainly determined by the noise assumption on data. E.g., Mulitscale [23]
assumed the noise of each patch and its similar patches in the same image to be correlated Gaussian distribution, and LRMoG
[44], DPGMM [39] and DDPT [43] fitted the image noise by using Mixture of Gaussian (MoG) as an approximator for noises.Datadriven Deep Learning based Methods: Instead of presetting image prior, deep learning methods directly learn a denoiser (formed as a deep neural network) from noisy to clean ones on a large collection of noisyclean image pairs. Jain and Seung [19]
firstly adopted a five layer convolution neural network (CNN) for the task. Then some autoencoder based methods
[37, 2] were applied. Meantime, Burger et al. [10]achieved the comparable performance with BM3D using plain multilayer perceptron (MLP). Zhang et al.
[40] further proposed the denoising convolution network (DnCNN) and achieved stateoftheart performance on Gaussian denoising tasks. Mao et al. [28] proposed a deep fully convolution encodingdecoding network with symmetric skip connection. In order to boost the flexibility against spatial variant noise, FFDNet [41] was proposed by preevaluating the noise level and inputting it to the network together with the noisy image. Guo et al. [17] and Brooks et al. [8] both attempted to simulate the generation process of the images in camera.3 Variational Denoising Network for Blind Noise Modeling
Given training set , where denote the training pair of noisy and the expected clean images, represents the number of training images, our aim is to construct a variational parametric approximation to the posterior of the latent variables, including the latent clean image and the noise variances, conditioned on the noisy image. Note that for the noisy image , its training pair is generally a simulated “clean” one obtained as the average of many noisy ones taken under similar camera conditions [4, 1], and thus is always not the exact latent clean image . This explicit parametric posterior can then be used to directly infer the clean image and noise distribution from any test noisy image. To this aim, we first need to formulate a rational full Bayesian model of the problem based on the knowledge delivered by the training image pairs.
3.1 Constructing Full Bayesian Model Based on Training Data
Denote and as any training pair in , where (width*height) is the size of a training image^{1}^{1}1We use and to express the indexes of training data and data dimension, respectively, throughout the entire paper.. We can then construct the following model to express the generation process of the noisy image :
(1) 
where is the latent clean image underlying , denotes the Gaussian distribution with mean and variance . Instead of assuming i.i.d. distribution for the noise as conventional [27, 13, 16, 38], which largely deviates the spatial variant and signaldepend characteristics of the real noise [41, 8], we models the noise as a noni.i.d. and pixelwise Gaussian distribution in Eq. (1).
The simulated “clean” image evidently provides a strong prior to the latent variable . Accordingly we impose the following conjugate Gaussian prior on :
(2) 
where is a hyperparameter and can be easily set as a small value.
Besides, for
, we also introduce a rational conjugate prior as follows:
(3) 
where
is the inverse Gamma distribution with parameter
and , represents the filtering output of the variance map by a Gaussian filter with window, and , are the matrix (image) forms of , , respectively. Note that the mode of above IG distribution is , which is a rational approximate evaluation of in window.3.2 Variational Form of Posterior
We first construct a variational distribution to approximate the posterior led by Eqs. (1)(3). Similar to the commonly used meanfield variation inference techniques, we assume conditional independence between variables and , i.e.,
(4) 
Based on the conjugate priors in Eqs. (2) and (3), it is natural to formulate variational posterior forms of and as follows:
(5) 
where and are designed as the prediction functions for getting posterior parameters of latent variable directly from . The function is represented as a network, called denoising network or DNet, with parameters . Similarly, and denote the prediction functions for evaluating posterior parameters of from , where represents the parameters of the network, called Sigma network or SNet. The aforementioned is illustrated in Fig. 1. Our aim is then to optimize these network parameters and so as to get the explicit functions for predicting clean image as well as noise knowledge from any test noisy image . A rational objective function with respect to and is thus necessary to train both the networks.
Note that the network parameters and are shared by posteriors calculated on all training data, and thus if we train them on the entire training set, the method is expected to induce the general statistical inference insight from noisy image to its underlying clean image and noise level.
3.3 Variational Lower Bound of Marginal Data Likelihood
For notation convenience, we simply write , , , as , , , in the following calculations. For any noisy image and its simulated “clean” image in the training set, we can decompose its marginal likelihood as the following form [7]:
(6) 
where
(7) 
Here represents the exception of w.r.t. stochastic variable
with probability density function
. The second term of Eq. (6) is a KL divergence between the variational approximate posterior and the true posterior with nonnegative value. Thus the first term constitutes a variational lower bound on the marginal likelihood of , i.e.,(8) 
According to Eqs. (4), (5) and (7), the lower bound can then be rewritten as:
(9) 
It’s pleased that all the three terms in Eq (9) can be integrated analytically as follows:
(10)  
(11)  
(12) 
where denotes the digamma function. Calculation details are listed in supplementary material.
We can then easily get the expected objective function (i.e., a negtive lower bound of the marginal likelihood on entire training set) for optimizing the network parameters of DNet and SNet as follows:
(13) 
3.4 Network Learning
As aforementioned, we use DNet and SNet together to infer the variational parameters , and , from the input noisy image , respectively, as shown in Fig. 1. It is critical to consider how to calculate derivatives of this objective with respect to involved in , , and to facilitate an easy use of stochastic gradient varitional inference. Fortunately, different from other related variational inference techniques like VAE [22], all three terms of Eqs. (10)(12) in the lower bound Eq. (9) are differentiable and their derivatives can be calculated analytically without the need of any reparameterization trick, largely reducing the difficulty of network training.
At the training stage of our method, the network parameters can be easily updated with backpropagation (BP) algorithm
[15] through Eq. (13). The function of each term in this objective can be intuitively explained: the first term represents the likelihood of the observed noisy images in training set, and the last two terms control the discrepancy between the variational posterior and the corresponding prior. During the BP training process, the gradient information from the likelihood term of Eq. (10) is used for updating both the parameters of DNet and SNet simultaneously, implying that the inference for the latent clean image and is guided to be learned from each other.At the test stage, for any test noisy image, through feeding it into DNet, the final denoising result can be directly obtained by . Additionally, through inputting the noisy image to the SNet, the noise distribution knowledge (i.e., ) is easily inferred. Specifically, the noise variance in each pixel can be directly obtained by using the mode of the inferred inverse Gamma distribution: .
3.5 Network Architecture
The DNet in Fig. 1 takes the noisy image as input to infer the variational parameters and in of Eq. (5), and performs the denoising task in the proposed variational inference algorithm. In order to capture multiscale information of the image, we use a UNet [32] with depth 4 as the DNet, which contains 4 encoder blocks ([Conv+ReLU]2+Average pooling), 3 decoder blocks (Transpose Conv+[Conv+ReLU]2) and symmetric skip connection under each scale. For parameter , the residual learning strategy is adopted as in [40], i.e., , where denotes the DNet with parameters . As for the SNet, which takes the noisy image as input and outputs the predicted variational parameters and in of Eq (5), we use the DnCNN [40] architecture with five layers, and the feature channels of each layer is set as 64. It should be noted that our proposed method is a general framework, most of the commonly used network architectures [41, 31, 24, 42] in image restoration can also be easily substituted.
3.6 Some Discussions
It can be seen that the proposed method succeeds advantages of both modeldriven MAP and datadriven deep learning methods. On one hand, our method is a generative approach and possesses fine interpretability to the data generation mechanism; and on the other hand it conducts an explicit prediction function, facilitating efficient image denoising as well as noise estimation directly through an input noisy image. Furthermore, beyond current methods, our method can finely evaluate and remove noni.i.d. noises embedded in images, and has a good generalization capability to images with complicated noises, as evaluated in our experiments. This complies with the main requirement of the blind image denoising task.
If we set the hyperparameter in Eq.(2) as an extremely small value close to , it is easy to see that the objective of the proposed method is dominated by the second term of Eq. (10), which makes the objective degenerate as the MSE loss generally used in traditional deep learning methods (i.e., minimizing . This provides a new understanding to explain why they incline to overfit noise bias in training data. The posterior inference process puts dominant emphasis on fitting priors imposed on the latent clean image, while almost neglects the effect of noise variations. This naturally leads to its sensitiveness to unseen complicated noises contained in test images.
Very recently, both CBDNet [17] and FFDNet [41]
are presented for the denoising task by feeding the noisy image integrated with the preestimated noise level into the deep network to make it better generalize to distinct noise types in training stage. Albeit more or less improving the generalization capability of network, such strategy is still too heuristic and is not easy to interpret how the input noise level intrinsically influence the final denoising result. Comparatively, our method is constructed in a sound Bayesian manner to estimate clean image and noise distribution together from the input noisy image, and its generalization can be easily explained from the perspective of generative model.
4 Experimental Results
We evaluate the performance of our method on synthetic and real datasets in this section. All experiments are evaluated in the sRGB space. We briefly denote our method as VDN in the following.
4.1 Experimental Setting
Network training and parameter setting: The weights of DNet and SNet in our variational algorithm were initialized according to [18]
. In each epoch, we randomly crop
patches with size from the images for training. The Adam algorithm [21] is adopted to optimize the network parameters through minimizing the proposed negative lower bound objective. The initial learning rate is set as and linearly decayed in half every 5 epochs until to . The window size in Eq. (3) is set as 11.Comparison methods: Several stateoftheart denoising methods are adopted for performance comparison, including CBM3D [11], WNNM [16], NCSR [14], MLP [10], DnCNNB [40], FFDNet [41], UDNet [24] and CBDNet [17]. Note that CBDNet is mainly designed for blind denoising task, and thus we only compared CBDNet on the real noise removal experiments.
4.2 Experiments on Synthetic NonI.I.D. Gaussian Noise Cases
Similar to [41], we collected a set of source images to train the network, including 432 images from BSD [5]
, 400 images from the validation set of ImageNet
[12] and 4744 images from the Waterloo Exploration Database [25]. Three commonly used datasets in image restoration (Set5, LIVE1 and BSD68 in [20]) were adopted as test datasets to evaluate the performance of different methods. In order to evaluate the effectiveness and robustness of VDN under the noni.i.d. noise configuration, we simulated the noni.i.d. Gaussian noise as following,(14) 
where is a spatially variant map with the same size as the source image. We totally generated four kinds of s as shown in Fig. 2. The first (Fig. 2 (a)) is used for generating noisy images of training data and the others (Fig. 2 (b)(d)) generating three groups of testing data (denotes as Cases 13). Under this noise generation manner, the noises in training data and testing data are with evident difference, suitable to verify the robustness and generalization capability of competing methods.


Cases  Datasets  Methods  
11´  
CBM3D  WNNM  NCSR  MLP  DnCNNB  FFDNet  UDNet  VDN  


Case 1  Set5  27.76  26.53  26.62  27.26  29.87  30.16  30.15  28.13  30.39 
11´  
LIVE1  26.58  25.27  24.96  25.71  28.81  28.99  28.96  27.19  29.22  
11´  
BSD68  26.51  25.13  24.96  25.58  28.72  28.78  28.77  27.13  29.02  


Case 2  Set5  26.34  24.61  25.76  25.73  29.05  29.60  29.56  26.01  29.80 
11´  
LIVE1  25.18  23.52  24.08  24.31  28.18  28.58  28.56  25.25  28.82  
11´  
BSD68  25.28  23.52  24.27  24.30  28.14  28.43  28.42  25.13  28.67  


Case 3  Set5  27.88  26.07  26.84  26.88  29.17  29.54  29.49  27.54  29.74 
11´  
LIVE1  26.50  24.67  24.96  25.26  28.15  28.39  28.38  26.48  28.65  
11´  
BSD68  26.44  24.60  24.95  25.10  28.10  28.22  28.20  26.44  28.46  

Comparson with the Stateoftheart: Table 1 lists the average PSNR results of all competing methods on three groups of testing data. From Table 1, it can be easily observed that: 1) The VDN outperforms other competing methods in all cases, indicating that VDN is able to handle such complicated noni.i.d. noise; 2) VDN surpasses FFDNet about 0.25dB averagely even though FFDNet depends on the true noise level information instead of automatically inferring noise distribution as our method. 3) the discriminative methods MLP, DnCNNB and UDNet seem to evidently overfit on training noise bias; 4) the classical modeldriven method CBM3D performs more stably than WNNM and NCSR, possibly due to the latter’s improper i.i.d. Gaussian noise assumption. Fig. 3 shows the denoising results of different competing methods on one typical image in testing set of Case 2, and more denoising results can be found in the supplementary material. Note that we only display the top four best results from all due to page limitation. It can be seen that the denoised images by CBM3D and DnCNNB still contain obvious noise, and FFDNet oversmoothes the image and loses some edge information, while our proposed VDN removes most of the noise and preserves more details.


Sigma  Datasets  Methods  
11´  
CBM3D  WNNM  NCSR  MLP  DnCNNB  FFDNet  UDNet  VDN  


Set5  33.42  32.92  32.57    34.04  34.30  34.31  34.19  34.34  
11´  
LIVE1  32.85  31.70  31.46    33.72  33.96  33.96  33.74  33.94  
11´  
BSD68  32.67  31.27  30.84    33.87  33.85  33.68  33.76  33.90  


Set5  30.92  30.61  30.33  30.55  31.88  32.10  32.09  31.82  32.24  
11´  
LIVE1  30.05  29.15  29.05  29.16  31.23  31.37  31.37  31.09  31.50  
11´  
BSD68  29.83  28.62  28.35  28.93  31.22  31.21  31.20  31.02  31.35  


Set5  28.16  27.58  27.20  27.59  28.95  29.25  29.25  28.87  29.47  
11´  
LIVE1  26.98  26.07  26.06  26.12  27.95  28.10  28.10  27.82  28.36  
11´  
BSD68  26.81  25.86  25.75  26.01  27.91  27.95  27.95  27.76  28.19  

Even though our VDN is designed based on the noni.i.d. noise assumption and trained on the noni.i.d. noise data, it also performs well on additive white Gaussian noise (AWGN) removal task. Table 2 lists the average PSNR results under three noise levels () of AWGN. It is easy to see that our method obtains the best or at least comparable performance with the stateoftheart method FFDNet. Combining Table 1 and Table 2, it should be rational to say that our VDN is robust and able to handle a wide range of noise types, due to its better noise modeling manner.
Noise Level Prediction: The SNet plays the role of noise modeling and is able to infer the noise distribution from the noisy image. To verify the fitting capability of SNet, we provided the predicted by SNet as the input of FFDNet, and the denoising results are listed in Table 1 (denoted as ). It is obvious that FFDNet under the real noise level and almost have the same performance, indicating that the SNet effectively captures proper noise information. The predicted noise level Maps on three groups of testing data are shown in Fig. 2 (b2d2) for easy observation.


CBM3D  WNNM  MLP  DnCNNB  CBDNet  VDN 


25.65  25.78  24.71  23.66  33.28  39.27 



DnCNNB  CBDNet  VDN 


38.65  38.68  39.28 



CBM3D  WNNM  NCSR  MLP  DnCNNB  FFDNet  CBDNet  VDN 


34.51  34.67  34.05  34.23  37.90  37.61  38.06  39.38 

4.3 Experiments on RealWorld Noise
In this part, we evaluate the performance of VDN on real blind denoising task, including two banchmark datasets: DND [30] and SIDD [1]. DND consists of 50 highresolution images with realistic noise from 50 scenes taken by 4 consumer cameras. However, it does not provide any other additional noisy and clean image pairs to train the network. SIDD [1] is a new realworld denoising benchmark released at 2018, containing real noisy images captured by 5 cameras under 10 scenes. For each noisy image, it estimates one simulated “clean” image through some statistical methods [1]. And 320 image pairs selected from all are packaged together as a medium version of SIDD, called SIDD Medium Dataset^{2}^{2}2https://www.eecs.yorku.ca/ kamel/sidd/index.php, for fast training of a denoiser.
Unfortunately, the whole SIDD Dataset only contains 10 scenes even though large number of image pairs in total. In order to train a robust blind denoiser under possibly more real scenarios, we combined the SIDD Medium Dataset and Renoir Dataset [4]
as our training data. The Renoir Dataset contains 117 noisy and relatively lownoise image pairs under different scenes, and the lownoise image in each pair is obtained under low light sensitivity and long exposure time compared with the noisy one. In the training precess, the variational hyperparameter
is set as and for image pairs from Renoir and SIDD Datasets, respectively, meaning that the simulated “clean” images in SIDD Dataset provides stronger prior than Renoir in our model (see Eq. (2)).Table 4 lists PSNR results of different methods on SIDD benchmark2. Note that we only list the results of the competing methods that are available on the official benchmark website. It is evident that VDN outperforms other methods. However, note that neither DnCNNB nor CBDNet performs well, possibly because they were trained on the other datasets, whose noise type is different from SIDD. For fair comparison, we retrained DnCNNB and CBDNet based on the SIDD dataset. The performance on the SIDD validation set ^{3}^{3}3https://competitions.codalab.org/competitions/21266 is listed in Table 4. Under same training conditions, VDN still outperforms DnCNNB 0.67 PSNR and CBDNet 0.64dB PSNR, indicating the effectiveness and significance of our noni.i.d. noise modeling manner. For easy visualization, on one typical denoising example, results of the best four competing methods are displayed in Fig. 4
Table 5 lists the performance of all competing methods on the DND benchmark^{4}^{4}4https://noise.visinf.tudarmstadt.de/. From the table, it is easy to be seen that discriminative deep learning based methods surpass traditional optimized methods. The proposed VDN outperforms the stateoftheart CBDNet [17] since it takes more intrinsic characteristics of noises into consideration.
5 Conclusion
We have proposed a new variational inference algorithm, namely varitional denoising network (VDN), for blind image denoising. The main idea is to learn an approximate posterior to the true posterior with the latent variables (including clean image and noise variances) conditioned on the input noisy image. Using this variational posterior expression, both tasks of blind image denoising and noise estimation can be naturally attained in a unique Bayesian framework. The proposed VDN is a generative method, which can easily estimate the noise distribution from the input data. Comprehensive experiments have demonstrated the superiority of VDN to previous works on blind image denoising. Our method can also facilitate the study of other lowlevel vision tasks, such as superresolution and deblurring. Specifically, the fidelity term in these tasks can be more faithfully set under the estimated noni.i.d. noise distribution by VDN, instead of the traditional i.i.d. Gaussian noise assumption.
References

[1]
Abdelrahman Abdelhamed, Stephen Lin, and Michael S. Brown.
A highquality denoising dataset for smartphone cameras.
In
The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)
, June 2018.  [2] Forest Agostinelli, Michael R Anderson, and Honglak Lee. Adaptive multicolumn deep neural networks with application to robust image denoising. In Advances in Neural Information Processing Systems, pages 1493–1501, 2013.
 [3] Michal Aharon, Michael Elad, Alfred Bruckstein, et al. Ksvd: An algorithm for designing overcomplete dictionaries for sparse representation. IEEE Transactions on signal processing, 54(11):4311, 2006.
 [4] Josue Anaya and Adrian Barbu. Renoir  a dataset for real lowlight noise image reduction. arXiv preprint arXiv:1409.8230, 2014.
 [5] Pablo Arbelaez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. IEEE Trans. Pattern Anal. Mach. Intell., 33(5):898–916, May 2011.

[6]
Christopher M Bishop.
Pattern recognition and machine learning
. springer, 2006.  [7] David M Blei, Michael I Jordan, et al. Variational inference for dirichlet process mixtures. Bayesian analysis, 1(1):121–143, 2006.
 [8] Tim Brooks, Ben Mildenhall, Tianfan Xue, Jiawen Chen, Dillon Sharlet, and Jonathan T Barron. Unprocessing images for learned raw denoising. arXiv preprint arXiv:1811.11127, 2018.
 [9] Antoni Buades, Bartomeu Coll, and JM Morel. A nonlocal algorithm for image denoising. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’05), volume 2, pages 60–65. IEEE, 2005.
 [10] Harold C Burger, Christian J Schuler, and Stefan Harmeling. Image denoising: Can plain neural networks compete with bm3d? In 2012 IEEE conference on computer vision and pattern recognition, pages 2392–2399. IEEE, 2012.
 [11] K. Dabov, A. Foi, V. Katkovnik, and K. Egiazarian. Image denoising by sparse 3d transformdomain collaborative filtering. IEEE Transactions on Image Processing, 16(8):2080–2095, Aug 2007.
 [12] Jia Deng, Olga Russakovsky, Jonathan Krause, Michael Bernstein, Alexander C. Berg, and Li FeiFei. Scalable multilabel annotation. In ACM Conference on Human Factors in Computing Systems (CHI), 2014.
 [13] Weisheng Dong, Guangming Shi, and Xin Li. Nonlocal image restoration with bilateral variance estimation: a lowrank approach. IEEE transactions on image processing, 22(2):700–711, 2013.
 [14] Weisheng Dong, Lei Zhang, Guangming Shi, and Xin Li. Nonlocally centralized sparse representation for image restoration. IEEE transactions on Image Processing, 22(4):1620–1630, 2012.
 [15] Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep learning. MIT press, 2016.
 [16] Shuhang Gu, Lei Zhang, Wangmeng Zuo, and Xiangchu Feng. Weighted nuclear norm minimization with application to image denoising. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2862–2869, 2014.
 [17] Shi Guo, Zifei Yan, Kai Zhang, Wangmeng Zuo, and Lei Zhang. Toward convolutional blind denoising of real photographs. arXiv preprint arXiv:1807.04686, 2018.
 [18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing humanlevel performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026–1034, 2015.
 [19] Viren Jain and Sebastian Seung. Natural image denoising with convolutional networks. In Advances in neural information processing systems, pages 769–776, 2009.
 [20] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image superresolution using very deep convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1646–1654, 2016.
 [21] Diederik P. Kingma and Jimmy Lei Ba. Adam: A method for stochastic optimization. international conference on learning representations, 2015.
 [22] Diederik P Kingma and Max Welling. Autoencoding variational bayes. arXiv preprint arXiv:1312.6114, 2013.
 [23] Marc Lebrun, Miguel Colom, and JeanMichel Morel. Multiscale image blind denoising. IEEE Transactions on Image Processing, 24(10):3149–3161, 2015.
 [24] Stamatios Lefkimmiatis. Universal denoising networks: a novel cnn architecture for image denoising. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3204–3213, 2018.
 [25] Kede Ma, Zhengfang Duanmu, Qingbo Wu, Zhou Wang, Hongwei Yong, Hongliang Li, and Lei Zhang. Waterloo Exploration Database: New challenges for image quality assessment models. IEEE Transactions on Image Processing, 26(2):1004–1016, Feb. 2017.
 [26] Matteo Maggioni, Vladimir Katkovnik, Karen Egiazarian, and Alessandro Foi. Nonlocal transformdomain filter for volumetric data denoising and reconstruction. IEEE transactions on image processing, 22(1):119–133, 2013.
 [27] Julien Mairal, Michael Elad, and Guillermo Sapiro. Sparse representation for color image restoration. IEEE Transactions on image processing, 17(1):53–69, 2008.
 [28] Xiaojiao Mao, Chunhua Shen, and YuBin Yang. Image restoration using very deep convolutional encoderdecoder networks with symmetric skip connections. In Advances in neural information processing systems, pages 2802–2810, 2016.
 [29] Pietro Perona and Jitendra Malik. Scalespace and edge detection using anisotropic diffusion. IEEE Transactions on pattern analysis and machine intelligence, 12(7):629–639, 1990.
 [30] Tobias Plotz and Stefan Roth. Benchmarking denoising algorithms with real photographs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1586–1595, 2017.
 [31] Tobias Plötz and Stefan Roth. Neural nearest neighbors networks. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. CesaBianchi, and R. Garnett, editors, Advances in Neural Information Processing Systems 31, pages 1087–1098. Curran Associates, Inc., 2018.
 [32] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. Unet: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computerassisted intervention, pages 234–241. Springer, 2015.
 [33] Stefan Roth and Michael J Black. Fields of experts. International Journal of Computer Vision, 82(2):205, 2009.
 [34] Leonid I Rudin, Stanley Osher, and Emad Fatemi. Nonlinear total variation based noise removal algorithms. Physica D: nonlinear phenomena, 60(14):259–268, 1992.
 [35] Eero P Simoncelli and Edward H Adelson. Noise removal via bayesian wavelet coring. In Proceedings of 3rd IEEE International Conference on Image Processing, volume 1, pages 379–382. IEEE, 1996.
 [36] Yanghai Tsin, Visvanathan Ramesh, and Takeo Kanade. Statistical calibration of ccd imaging process. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, volume 1, pages 480–487. IEEE, 2001.
 [37] Junyuan Xie, Linli Xu, and Enhong Chen. Image denoising and inpainting with deep neural networks. In Advances in neural information processing systems, pages 341–349, 2012.
 [38] Jun Xu, Lei Zhang, and David Zhang. A trilateral weighted sparse coding scheme for realworld image denoising. In The European Conference on Computer Vision (ECCV), September 2018.
 [39] Zongsheng Yue, Deyu Meng, Yongqing Sun, and Qian Zhao. Hyperspectral image restoration under complex multiband noises. Remote Sensing, 10(10):1631, 2018.
 [40] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE Transactions on Image Processing, 26(7):3142–3155, 2017.
 [41] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Ffdnet: Toward a fast and flexible solution for cnnbased image denoising. IEEE Transactions on Image Processing, 27(9):4608–4622, 2018.
 [42] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image superresolution. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2472–2481, 2018.
 [43] Fengyuan Zhu, Guangyong Chen, Jianye Hao, and PhengAnn Heng. Blind image denoising via dependent dirichlet process tree. IEEE transactions on pattern analysis and machine intelligence, 39(8):1518–1531, 2017.
 [44] Fengyuan Zhu, Guangyong Chen, and PhengAnn Heng. From noise modeling to blind image denoising. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 420–429, 2016.