Cut2Self: Single Image based Self-supervised

Denoising.

Jubyrea
MCE 077 05513
Supervisor: Md. Tauhid Bin Iqbal
August 30, 2023
Department of Computer Science and Engineering
Stamford University Bangladesh

1
Table of contents
1. Introduction
2. Background Studies
Traditional Denoiser
Learning-based Denoising Methods
3. Limitations of Existing Works
4. Proposed Solutions
Cut2Self
Experimental Result

Cut2Self-DDM
Experimental Results

5. Comparison Between Methods on Higher Noise Level

6. Conclusion

2
Introduction
Image Denoising?

y x n

Figure 1: Visual representation of a noisy image (y), a clean counterpart (x), and noise (n). Noise
(n) can be signal independent, signal dependent, or both.

y=x+n (1)

◦ Fundamental computer vision task.

◦ Involves reconstructing or recovaring a clean image from a noisy counterpart.
3
Image Denoising

Why does noise appear in image?

• Environmental factor
• Lighting Conditions, i.e. low light
• Atmospheric Disturbances, i.e haze,
fog, or dust
• Temperature, Humidity, and Motion Application areas:
• Magnetic Field, Electromagnetic
• Artificial Intelligence and Robotics
Interference (EMI), and Electrical
Noise • Biomedical Imaging
• Surveillance and Security
• Device related issues
• Remote Sensing
• High ISO Settings
• Small Sensor Seize, and Sensor • Astronomy
Noise • Forensic Analysis, etc.

• Pre-processing and Post-processing

• i.e. Compression, Aggressive
sharpening

4
Background Studies
Denoising Methods

Denoising Methods

Traditional Denoising Learning-base Denoising

Algorithms Methods

Un/Semi-
Supervised Self-supervised
supervised

5
Traditional Denoiser

Non-learning based traditional denoising algorithms: NLM [1],

K-SVD [2], BM3D [3], WNLM [4], etc.

Algorithms are based on: Limitations of these methods:

• non-local matching • cannot handle complex noise pattern
• K-singular value decomposition • design for specific noise, fail to generalize
• self-similar patches and block matching on noise patterns
with 3D filtering • manual parameter tuning
• weighted nuclear norm minimization • computationally demanding

6
Learning-based Denoising Methods

Supervised Learning
DnCNN [5], U-Net [6], FFDNet [7], RIDNet [8], are some supervised learning models
with impressive performance.

• use paired dataset for training (xi , yi )

here xi = clean image, yi = noisy image
• performance depends on the volume of training dataset [9] [10] [11]
• paired datasets are inadequate for every real scenario
• the issue handled by adding synthetic noise to clean image
• impressive performance on additive noise, while relatively poor on real world noise

7
Learning-based Denoising Methods (contd.)

Un-supervised and Semi-supervised Learning

Noise2Noise [12], UIDNet [13], IDR [14], C2N [15], GCBD [16], UIDNet [13] are some
unsupervised and generative models.

• does not need labeled dataset

(xi ′ , yi ′ ) → (noisy, noisy) [12]
• require noisy images from same underlying distribution
• expensive to capture in real world

• generate noise model and paired it to clean image

noise model + clean image → (xi , yi ) [17] [18]
• noise models fails to prototype real noise in many cases

Self-supervised Learning
• denoiser train from the body of input image
• no external data other than input image
• applicable in areas only noisy image available

Noise2Void [19], Noise2Self [20], Self2Self [21], line-19 [22], Noise2Same[23],

Neighbor2Neighbor [24], CVF-SID [25]

8
Limitations of Existing Works
Challenges and Strategies of Self-supervised Learning

Some Strategies Basic challenges

• blind spot network in receptive • identity mapping
fields→Noise2Void • network overfitting
• blind spot strategy→Noise2Self • increased variance
• mask based denoising • desired performance gain
methods→Self2Self, line-19, Noise2Same
• random neighbor sub-sampler
→Neighbor2Neighbor
Self2Self Strategies

• trains the denoising network by creating training pairs from input noisy image
• pairs are generated by randomly dropping pixels from the input image using Bernoulli
distribution
• try to address: identity mapping, overfitting, variance
• assumption → noise is pixel independent
• real world noise is pixel dependent [11] [26] [25]

9
Self2Self-Limitations

• single pixel dropping schema go against

the argument of Self2Self
• we are motivated by the simple
regularization technique proposed by
(Terrance DeVries et al., 2017) [28]a
• randomly masking out square regions of
input during training improve the
Figure 2: Self-similarity of real world noisy robustness and overall performance of
image; image is taken from SIDD+ validation CNN
dataset [27]. • we used this regularization technique in
our denoising framework to mitigate the
common self-supervised denoising issues
• images possess real contextual
a Improved Regularization of
information when corrupted by real world
noise Convolutional Neural Networks with Cutout
• dropping a pixel from one area do not
totally dispossess the contextual
information
• network learns similar information from
neighboring pixels
10
Proposed Solutions
Proposed Solutions

■ We present efficient self supervised single image denoising methods, in order to

combat the limitations of existing methods.
■ First, we propose Cut2Self.
• we masked out square regions from an area instead of pixel
• masking out a bigger square region instead of a singleton pixel, guides the network to
overlook same information of a location
• the network is forced to learn from a holistic perspective
• thus, avoid identity mapping, overfitting, and improve generalization

■ Second, we propose Cut2Self-DDM.

• this work present data dependent masking during training pair generation
• mask an entire column as an area
• sampling the dropout probability from a Gaussian distribution
• introduce data dependency by incorporating the standard deviation of the particular
column during sampling dropout probability
• introduce randomness in dropout rate
• ensure consistent dropout, faster convergence, and improved generalization
■ We have tested our methods against synthetic noise and real world noise with
BSD68, Set-14, Set-9, PolyU, and SIDD+ validation datasets and achieved superior
performance compared to other traditional, supervised, unsupervised, and self
supervised methods.

11
Cut2Self Discussion

Figure 3: The proposed framework for denoising. Our input to the network is a noisy image (a),
(b) is the noisy input to the network after masking out square regions as depicted in the figure, (c)
represents the denoising network depicting the encoder and decoder blocks. Final output is the
denoised image (d); for ease, area under red box is shown after zooming in for a clear view.

12
Cut2Self (Training Schema)

◦ We create training pairs from the noisy input by masking out square regions.
First sample of the training pair: ŷ = c ⊙ y
Second sample of the training pair: ȳ = (1 − c) ⊙ y

y c ŷ ȳ

Figure 4: Noisy input (y), proposed cutout mask (c), first and sample of training pair (ŷ & ȳ).

For N training iteration our pairs are: {yˆn , ȳn }N

n=1 . Our loss function L is defined as:

N
∑
min ∥Fθ (ŷn ) − ȳn ∥ℓcn , (2)
n=1

13
Cut2Self (Training Schema) (contd.)
We rewrite the loss function Eq. 2 for ℓ2 loss,
N
∑ 2
∥Fθ (ŷn ) − ȳn ∥c =
n
∑N
2
∥Fθ (ŷn ) − (xn + nn )∥c
n
Proposition
n=1 n=1 Training a denoiser with loss function (Eq. 2) is
N N
equivalent to training with ground truth loss and
∑ ∑
=
2
∥Fθ (ŷn ) − xn ∥c +
2
∥nn ∥c noise variance. The expectation of the loss
n=1
n
n=1
n function with respect to noise is similar to the
expectation of training with ground truth loss
N
−2
∑
(Fθ (ŷn ) − xn )((1 − cn ) ⊙ n)
T (3) and noise variance. Proof is shown in Eq. 4.
n=1 N N
∑ ℓ ∑ ℓ
N
∑ ∑N ∥Fθ (ŷ) − x∥c + ∥σ∥c (5)
2 2 n n
= ∥Fθ (ŷ) − x∥c + ∥n∥c n=1 n=1
n n
n=1 n=1
N Here σ represents the standard deviation of
T ∑
−2n ( (1 − cn ) ⊙ (Fθ (ŷn ) − xn )
n=1 noise. The proof above is shown using ℓ = 2. The
proof can however be generalized for other
Now if we take En on Eq. 3:
values of ℓ as well.
N
∑ 2
En [ ∥Fθ (ŷn ) − ȳn ∥c ]
n
n=1
N
∑ N
∑
2 2
= En [ ∥Fθ (ŷ) − x∥c ] + En [ ∥n∥c ]
n n
n=1 n=1
(4)
N
T ∑
−2En [n ( (1 − cn ) ⊙ (Fθ (ŷn ) − xn )]
n=1
N
∑ ∑N
2 2
≡ ∥Fθ (ŷ) − x∥c + ∥σ∥c
n n
n=1 n=1
14
Cut2Self (Denoising Schema)

◦ In N iterations we create Fθ1 , Fθ2 , ..., FθN denoisers.

◦ We have x1 , x2 , x3 , ...xn estimations from the denoisers.
Final denoised image is the average of all predictions:
N
∑
x = 1/N xn . (6)
n=1

Important takeaways:

• in the final output, we ensemble all the network predictions

• we create more training sample using data augmentation
• these techniques reduce the variance in the final prediction

15
Network Architecture

Figure 5: Encoder-Decoder denoising network.

• encoder decoder network • first four DBs consecutively connects an

• input to the encoder is a masked image up-sampling layer, a concatenation, and
two convolutional layers with LReLUs
• first five EBs successively connects a PConv
layer, a LReLU, and a max pooling layer • last DB returns the feature cube back to
the image size
• the last EB connecs to a PConv layer and a
LReLU

16
Parameter Setting: Cut2Self
◦ Cut2Self is tested for best hyper ◦ Best parameters setting:
parameter setting on Set-9 dataset. • Square shape of mask worked best
◦ Some of the key parameters are: • Dropping 0.3% area yields best denoising
• Shape of mask results
• Dropping percentage • Size of mask is set to 4 × 4
• Size of mask • After 1000 training iterations we averaged
• Number of training iteration add 100 images to generate a prediction
prediction • We have experimented with ℓ1 loss
• Choice of loss function function

Circle shaped mask Triangle shaped mask Ellipse shaped mask Square shaped mask

Figure 6: Visual representation of a noisy image with different mask shapes.

17
Parameter Setting: Cut2Self (contd.)

Area dropping: 0.1% Area dropping: 0.3% Area dropping: 1.0% Area dropping: 1.5% Area dropping: 3.0%

Figure 7: Visual representation of masked area for different dropping percentages, e.g., 0.1%,
0.3%,1.0%, 1.5%, and 3.0% respectively.

30.8
30.8 30.71 30.71
30.71 30.71
30.7 30.63 30.7
30.6

PSNR(db)
30.6 30.5
PSNR(db)

PSNR(db)
30.62
PSNR(db)
PSNR(db)

30.6 30.58 30.6

30.4
30.4 30.53
30.31 30
30.26 30.5
30.2 30.18 30.5
30.24 30.45
30.44
30.2
30.4 1000-100 10000-1000 200000-10000
△ □ ⃝ ◦ 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 2×2 4×4 8 × 8 16 × 16 32 × 32 Number of Itr-Pred ℓ1 ℓ2

(a) Shape of Mask (b) Dropping Percentage (c)Size of Mask (d) Iteration-Prediction (e) Loss function

Figure 8: PSNR results of Cut2Self on different parameter settings.

18
Cut2Self (Quantitative Results - Additive noise)

Table 1: Quantitative comparison of BSD68 using (PSNR(db)/SSIM) matrices. Highest PSNR is

marked in bold. Results marked in † are taken from [14] and ‡ are raken from [21].

Methods σ = 25 σ = 50
PSNR SSIM PSNR SSIM
†BM3D 27.96 7.74 28.71 0.670
‡DIP 28.55 0.782 25.04 6.45
‡N2V 25.34 6.81 23.85 6.18
‡K-SVD 28.42 7.96 25.08 6.53
†N2N 29.15 0.831 26.23 0.725
‡N2S 27.19 7.69 24.53 6.42
†IDR 29.20 0.835 26.25 0.726
‡Self2Self 28.70 8.03 25.92 6.99
Cut2Self 31.70 0.835 28.95 0.525

19
Cut2Self (Quantitative Results - Additive noise) (contd.)

Table 2: Quantitative comparison of different methods for Gaussian noise on Set-14.

Comparisons are done using (PSNR(db)/SSIM) matrices, and the highest PSNR is marked in bold.
Results except ours are taken from [24].

Methods σ = 25 σ ∈ [5, 50]

PSNR SSIM PSNR SSIM
C-BM3D 30.88 0.854 30.94 0.866
DIP 27.16 0.758 26.61 0.738
Self2Self 30.08 0.839 29.97 0.849
N2V 28.84 0.802 29.01 0.792
Laine19-mu 29.93 0.830 29.71 0.822
Laine19-pme 31.36 0.866 31.21 0.855
DBSN 30.63 0.846 29.49 0.814
Neighbor2Neighbor 31.09 0.864 31.05 0.858
Cut2Self 31.38 0.817 31.71 0.815

20
Self2Self-Cutout (Quantitative Results - Real-World noise)

Table 3: Quantitative comparison of real world noise for different methods on PolyU dataset.
Comparisons are done using (PSNR(db)/SSIM) matrices, and the highest PSNR is marked in bold.
Results except ours are taken from [21].

Metric TWSC DIP CBM3D N2V N2S CDnCNN Self2Self Cut2Self

PSNR 36.10 36.95 36.98 34.08 35.46 37.55 37.52 37.57
SSIM 0.963 0.975 0.977 0.954 0.965 0.983 0.980 0.942

Table 4: Quantitative comparison of real world noise for different methods for SSID+ validation
dataset. Comparisons are done using (PSNR(db)/SSIM) matrices, and the highest PSNR is marked
in bold. Results except ours are taken from [25].

Metric BM3D WNNM NC MCWNNM C2N R2R N2V N2S CVF-SID Cut2Self
PSNR 25.65 26.20 31.31 33.40 34.08 35.04 29.35 30.72 34.81 31.95
SSIM 0.475 0.693 0.725 0.815 - 0.844 0.651 0.787 0.944 0.642

21
Qualitative Results

GT Noisy Cut2Self

Figure 9: Denoised output of SIDD+ validation dataset.

22
Qualitative Results (contd.)

GT Noisy Cut2Self

Figure 10: Denoised output of PolyU dataset.

23
Qualitative Results (contd.)

GT Noisy BM3D

KSVD DIP N2V

N2S Self2Self Cut2Self

Figure 11: Visual comparisons of Set-14 with different models for varied σ ∈ [5, 50]. 24
Improvement Schopes of Cut2Self

Although Cut2Self achieves prominent result, there are scope for improvement.

GT Noisy Cut2Self

Figure 12: Some under performing denoising outputs of Cut2Self.

◦ The dropout rate is fixed which should be varied for better generalization.
◦ Trivial reconstruction errors (i.e., noise in output, blurriness) can be further improved.
◦ Method can be improved for faster convergence. 25
Cut2Self-DDM Discussion

◦ We propose a denoising schema with varied dropout rate.

◦ Dropout probability is sampled from Gaussian distribution.
◦ We incorporated the statistical properties of the input image during probability sampling.

Dropout: 49.37% 52.19% 48.75% 49.37% 50.31%

Iteration: 1 2 3 4. . . . ...N

Figure 13: Sample demonstration of masked input with a varied dropout rate during training
iterations. Moderately, the average dropout rate is half of the input image.

1
1 [29], [30], [31], [32], [33].

26
Cut2Self-DDM: Mask Generation

Mask generation: Seudocode:

• sample the dropout probability from 1. Calculate Stdev along axis [0] of input image,
Gaussian distribution 2. prob array ← 0 (size: axis [0] of input image )
• drop column as an area according to 3. For j to len(prob):
the probability sample 3.1 probs[j] = sample from Gaussian
• mean of the distribution set to 0.5, distribution(0.5, scale=stdev[j], size=1)
and scale is the standard deviation of
3.1.1 if prob[j] > 1
the particular column
Clip prob[j] = 1 prob[j] < 0
• incorporating standard deviation in Clip prob[j] = 0
probability sampling add randomness 3.2 probs = tile(probs, (same size as input
in dropout rate image))
• fluctuation of dropout rate is not 3.3 mask = create binary mask (sample
abrupt, on average nearly half of the from Bernoulli distribution according
pixels of input image drop in each to prob)
iteration 4. input to the network = mask * input image

27
Cut2Self-DDM (Training Schema)

First sample of training pair ŷ = y ⊙ g. We minimizing the loss between:

The second sample of the training pair
M
∑
ȳ = y ⊙ (1 − g).
min ∥Fθ (ŷm ) − ȳm ∥ℓgm (7)
m=1
Final denoising:

x = 1/M(Fθ1 + Fθ2 + .. + FθM )

x = 1/M(x1 + x2 + x3 + ... + xM )
(8)
N
∑
x = 1/M xm .
ŷ ȳ m=1

◦ we ensemble all the network prediction

Figure 14: Training Samples (ŷ & ȳ).
◦ in Cut2Self-DDM there is a balance between the
covariance and independence of predictions
◦ we have further data augmentation i.e.,
◦ g is the proposed data driven mask
horizontal and vertical flipping, and dropout in fc
◦ all the pixels p in column c is set to zero,
layer
∀p∈c = 0.
◦ these techniques lower the variance of
◦ training pairs can be denoted as {yˆn , ȳn }M
m=1 prediction
◦ in M training iteration we create M numbers of
training pairs, m = {1, 2, 3, ..., M}
28
Parameter Setting: Cut2Self-DDM

Table 5: Quantitative comparison between

horizontal and vertical masking.

Mask
Set-9 Set-14
Orientation
PSNR SSIM PSNR SSIM
Vertical 32.605 0.79 32.15 0.855
Horizontal 32.295 0.754 31.832 0.812

(a) (b)
Table 6: Quantitative comparison between L1
and L2 regularization.

Figure 15: A typical input image after

Dataset L1 L2
horizontal (a), and vertical (b) masking.
PSNR SSIM PSNR SSIM
Set-14 σ = 25 32.52 0.861 32.15 0.855
Parameter setting for Cut2Self-DDM is Set-14 σ ∈ [5, 50] 32.56 0.897 32.32 0.896
conducted on Set-9, and Set-14. Set-9 σ = 25 32.70 0.795 32.61 0.790
◦ Orientation of mask is set to vertical. Set-9 σ = 50 32.01 0.749 30.90 0.786
29
◦ L1 loss uses throughout the experiments.
Cut2Self-DDM (Quantitative Results)

Table 7: Quantitative comparison on BSD68 dataset using (PSNR(db)/SSIM) matrices, and the
highest PSNR is marked in bold.

Methods σ = 25 σ = 50
PSNR SSIM PSNR SSIM
BM3D 27.96 7.74 28.71 0.670
DIP 28.55 0.782 25.04 6.45
N2V 25.34 6.81 23.85 6.18
K-SVD 28.42 7.96 25.08 6.53
DBSN 28.81 0.818 25.95 0.703
N2N 29.15 0.831 26.23 0.725
N2S 27.19 7.69 24.53 6.42
IDR 29.20 0.835 26.25 0.726
Self2Self 28.70 8.03 25.92 6.99
Cut2Self 31.70 0.835 28.95 0.525
Cut2Self-DDM 33.25 0.871 31.96 0.794

30
Cut2Self-DDM (Quantitative Results) (contd.)

Table 8: Quantitative result on Set-14, the highest PSNR is marked in bold.

Methods σ = 25 σ ∈ [5, 50]

PSNR SSIM PSNR SSIM
C-BM3D 30.88 0.854 30.94 0.866
DIP 27.16 0.758 26.61 0.738
N2V 28.84 0.802 29.01 0.792
Laine19-mu 29.93 0.830 29.71 0.822
Laine19-pme 31.36 0.866 31.21 0.855
DBSN 30.63 0.846 29.49 0.814
Neighbor2Neighbor 31.09 0.864 31.05 0.858
Self2Self 30.08 0.839 29.97 0.849
Cut2Self 31.38 0.817 31.71 0.815
Cut2Self-DDM 32.52 0.861 32.56 0.896

31
Cut2Self-DDM (Quantitative Results) (contd.)

Table 9: Quantitative comparison of different methods for Gaussian noise on Set-9.

Comparisons are done using (PSNR(db)) matrices, and the highest PSNR is marked in bold.

Set-9 K-SVD CBM3D CDnCNN N2V N2S DIP Self2Self Cut2Self Cut2Self-DDM
σ = 25 30.00 31.67 31.42 28.12 29.30 30.77 31.74 30.71 32.70
σ = 50 26.50 28.95 28.84 26.01 27.25 28.23 29.25 28.88 32.01

Table 10: Quantitative comparison of real world noise for different methods on PolyU dataset.
Comparisons are done using (PSNR(db)/SSIM) matrices, and the highest PSNR is marked in bold.

Metric DIP CBM3D N2V N2S CDnCNN Self2Self Cut2Self Cut2Self-DDM

PSNR 36.95 36.98 34.08 35.46 37.55 37.52 37.57 37.896
SSIM 0.975 0.977 0.954 0.965 0.983 0.980 0.942 0.948

2
2ℓ loss performed poorly on PolyU dataset; result is generated using ℓ2 loss.
1

32
Qualitative Results

GT Noisy

BM3D KSVD DIP

Noise2same N2V N2S

Self2Self Cut2Self Cut2Self-DDM

Figure 16: Visual comparisons of different models for σ = 25 on BSD68 dataset. 33

Qualitative Results (contd.)

GT Noisy

BM3D KSVD DIP N2V

N2S Self2Self Cut2Self Cut2Self-DDM

Figure 17: Visual comparisons of Set-14 with different models for varied σ ∈ [5, 50]. 34
Qualitative Results (contd.)

GT Noisy Cut2Self-DDM

Figure 18: Denoised output of PolyU dataset.

35
Comparison Between Methods on
Higher Noise Level
Comparison between methods

GT Noisy

Self2Self Cut2Self Cut2Self-DDM

Figure 19: Visual representation of our model for σ = 50 on BSD68 datasets. 36

Comparison between methods (contd.)

GT Noisy

Self2Self Cut2Self Cut2Self-DDM

Figure 20: Qualitative demonstration between Self2Self, Cut2Self, and Cut2Self-DDM for σ = 50
on Set-9 datasets.

37
Conclusion
Conclusion

• We present robust and efficient self supervised single image denoising

methods without the need of any external data or any prior information.
• In particular, we propose two different denoising approaches, Cut2Self and
Cut2Self with Data Driven Mask (Cut2Self-DDM).
• The presented approaches efficiently reduce overfitting, identity mapping,
network variance during prediction while ensuring credible denoising and
faster convergence.
• We have tested our methods in the presence of additive noise and real world
noise, and compared our results against other state-of-the-art methods. Our
method performed better compared to other existing traditional and learning
based methods.

• Future Works:
• The proposed work can be further used to domain-specific applications such as
medical imaging, satellite and surveillance imaging, and space imaging.
• The methods can be integrated with image superresolution tasks, as in many cases
denoising is a prior work of superresolution. The proposed work can be modified for
inpainting as well.

38
Publication List

39
 Thank You
Questions Please

40
Appendix

Parameter Setting Cut2Self-DDM:

(a) (b)

Figure 21: Effect of mean and standard deviation (std) on data driven dropout.

• mean 0.5 aids to drop around half of the input image

• data driven standard deviation leads to smooth fluctuation in dropout rate
• smooth fluctuation reduce network variance during prediction
References

A. Buades, B. Coll, and J.-M. Morel, “A non-local algorithm for

image denoising,” in 2005 IEEE computer society conference on
computer vision and pattern recognition (CVPR’05), vol. 2,
pp. 60–65, Ieee, 2005.
M. Aharon, M. Elad, and A. Bruckstein, “K-svd: An algorithm for
designing overcomplete dictionaries for sparse representation,”
IEEE Transactions on signal processing, vol. 54, no. 11,
pp. 4311–4322, 2006.
K. Dabov, A. Foi, V. Katkovnik, and K. Egiazarian, “Image denoising
by sparse 3-d transform-domain collaborative filtering,” IEEE
Transactions on image processing, vol. 16, no. 8, pp. 2080–2095,
2007.
References (contd.)

S. Gu, L. Zhang, W. Zuo, and X. Feng, “Weighted nuclear norm

minimization with application to image denoising,” in
Proceedings of the IEEE conference on computer vision and
pattern recognition, pp. 2862–2869, 2014.
K. Zhang, W. Zuo, Y. Chen, D. Meng, and L. Zhang, “Beyond a
gaussian denoiser: Residual learning of deep cnn for image
denoising,” IEEE transactions on image processing, vol. 26, no. 7,
pp. 3142–3155, 2017.
O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional
networks for biomedical image segmentation,” in International
Conference on Medical image computing and computer-assisted
intervention, pp. 234–241, Springer, 2015.
References (contd.)

K. Zhang, W. Zuo, and L. Zhang, “Ffdnet: Toward a fast and flexible

solution for cnn-based image denoising,” IEEE Transactions on
Image Processing, vol. 27, no. 9, pp. 4608–4622, 2018.
S. Anwar and N. Barnes, “Real image denoising with feature
attention,” in Proceedings of the IEEE/CVF international
conference on computer vision, pp. 3155–3164, 2019.
J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet:
A large-scale hierarchical image database,” in 2009 IEEE
conference on computer vision and pattern recognition,
pp. 248–255, Ieee, 2009.
T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan,
P. Dollár, and C. L. Zitnick, “Microsoft coco: Common objects in
context,” in European conference on computer vision,
pp. 740–755, Springer, 2014.
References (contd.)

S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, M.-H. Yang, and

L. Shao, “Cycleisp: Real image restoration via improved data
synthesis,” in Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition, pp. 2696–2705, 2020.
J. Lehtinen, J. Munkberg, J. Hasselgren, S. Laine, T. Karras,
M. Aittala, and T. Aila, “Noise2Noise: Learning image restoration
without clean data,” in Proceedings of the 35th International
Conference on Machine Learning (J. Dy and A. Krause, eds.),
vol. 80 of Proceedings of Machine Learning Research,
pp. 2965–2974, PMLR, 10–15 Jul 2018.
Z. Hong, X. Fan, T. Jiang, and J. Feng, “End-to-end unpaired image
denoising with conditional adversarial networks,” in Proceedings
of the AAAI Conference on Artificial Intelligence, vol. 34,
pp. 4140–4149, 2020.
References (contd.)

Y. Zhang, D. Li, K. L. Law, X. Wang, H. Qin, and H. Li, “Idr:

Self-supervised image denoising via iterative data refinement,”
in Proceedings of the IEEE/CVF Conference on Computer Vision
and Pattern Recognition, pp. 2098–2107, 2022.
G. Jang, W. Lee, S. Son, and K. M. Lee, “C2n: Practical generative
noise modeling for real-world denoising,” in Proceedings of the
IEEE/CVF International Conference on Computer Vision,
pp. 2350–2359, 2021.
Y. Chen and T. Pock, “Trainable nonlinear reaction diffusion: A
flexible framework for fast and effective image restoration,” IEEE
transactions on pattern analysis and machine intelligence,
vol. 39, no. 6, pp. 1256–1272, 2016.
References (contd.)

L. D. Tran, S. M. Nguyen, and M. Arai, “Gan-based noise model for

denoising real images,” in Proceedings of the Asian Conference
on Computer Vision, 2020.
E. Park, Y.-J. Moon, D. Lim, and H. Lee, “De-noising sdo/hmi solar
magnetograms by image translation method based on deep
learning,” The Astrophysical Journal Letters, vol. 891, no. 1, p. L4,
2020.
A. Krull, T.-O. Buchholz, and F. Jug, “Noise2void-learning
denoising from single noisy images,” in Proceedings of the
IEEE/CVF conference on computer vision and pattern recognition,
pp. 2129–2137, 2019.
J. Batson and L. Royer, “Noise2self: Blind denoising by
self-supervision,” in International Conference on Machine
Learning, pp. 524–533, PMLR, 2019.
References (contd.)

Y. Quan, M. Chen, T. Pang, and H. Ji, “Self2self with dropout:

Learning self-supervised denoising from single image,” in
Proceedings of the IEEE/CVF conference on computer vision and
pattern recognition, pp. 1890–1898, 2020.
S. Laine, T. Karras, J. Lehtinen, and T. Aila, “High-quality
self-supervised deep image denoising,” Advances in Neural
Information Processing Systems, vol. 32, 2019.
Y. Xie, Z. Wang, and S. Ji, “Noise2same: Optimizing a
self-supervised bound for image denoising,” Advances in Neural
Information Processing Systems, vol. 33, pp. 20320–20330, 2020.
T. Huang, S. Li, X. Jia, H. Lu, and J. Liu, “Neighbor2neighbor:
Self-supervised denoising from single noisy images,” in
Proceedings of the IEEE/CVF conference on computer vision and
pattern recognition, pp. 14781–14790, 2021.
References (contd.)

R. Neshatavar, M. Yavartanoo, S. Son, and K. M. Lee, “Cvf-sid:

Cyclic multi-variate function for self-supervised image denoising
by disentangling noise from image,” in Proceedings of the
IEEE/CVF Conference on Computer Vision and Pattern
Recognition, pp. 17583–17591, 2022.
G. Torricelli, F. Argenti, and L. Alparone, “Modelling and
assessment of signal-dependent noise for image de-noising,” in
2002 11th European signal processing conference, pp. 1–4, IEEE,
2002.
A. Abdelhamed, S. Lin, and M. S. Brown, “A high-quality denoising
dataset for smartphone cameras,” in Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition,
pp. 1692–1700, 2018.
References (contd.)

T. DeVries and G. W. Taylor, “Improved regularization of

convolutional neural networks with cutout,” arXiv preprint
arXiv:1708.04552, 2017.
S. Wang and C. Manning, “Fast dropout training,” in international
conference on machine learning, pp. 118–126, PMLR, 2013.
J. Ba and B. Frey, “Adaptive dropout for training deep neural
networks,” Advances in neural information processing systems,
vol. 26, 2013.
X. Shen, X. Tian, T. Liu, F. Xu, and D. Tao, “Continuous dropout,”
IEEE transactions on neural networks and learning systems,
vol. 29, no. 9, pp. 3926–3937, 2017.
References (contd.)

S. Park and N. Kwak, “Analysis on the dropout effect in

convolutional neural networks,” in Computer Vision–ACCV 2016:
13th Asian Conference on Computer Vision, Taipei, Taiwan,
November 20-24, 2016, Revised Selected Papers, Part II 13,
pp. 189–204, 2017.
Z. Li, B. Gong, and T. Yang, “Improved dropout for shallow and
deep learning,” Advances in neural information processing
systems, vol. 29, 2016.