GMR Institute of Technology, Rajam

GMR Institute of Technology

An Extensive Review of Generative Adversarial

Network Based Image Denoising Methods

Dec 31, 2024

1
emb
Dec
201

1
er
7

Dec 31, 2024 1

 An Extensive Review of Generative Adversarial
Network Based Image Denoising Methods

(23345A4205)
GMR Institute of Technology

Under the guidance of

Mr.G.Dharma Raju
Assistant Professor
GMR Institute of Technology
 Abstract
One of the main challenges in image processing is image denoising, which
aims to eliminate noise while keeping crucial details. This Paper has benefited
greatly from recent developments in Generative Adversarial Networks
(GANs), which provide fresh approaches to age-old denoising issues. This
GMR Institute of Technology

paper looks at several approaches and their efficacy in providing a thorough

overview of GAN-based picture denoising techniques. We examine the
development of GAN architectures used for denoising tasks, starting with
simple models and moving up to more complex versions like self-supervised
Noise2Noise methods and Wasserstein GANs (WGANs). Important topics
covered include single picture denoising and the modification of GANs to
handle blurriness. In addition to offering suggestions for future research paths
in the field, our assessment aims to shed light on the advantages and
disadvantages of the GAN-based denoising techniques that are already in use

Keywords: Generative Adversarial Networks (GANs), Image Denoising,

Noise2Noise, Wasserstein GAN (WGAN), Image Quality Enhancement,
 Base Paper
Asavaron Limsuebschuea (2024), Self-Augmented Noisy Image for
Noise2Noise Image Denoising, IEEE Access, 71076-71087 .
GMR Institute of Technology

Objective
1. Enhance Image Quality

2. Preserve Image Features

3. Examine the Role of Generative Models in Enhancing Denoising

Performance

4. Evaluate the Robustness of Denoising Models to Adversarial Models

5. Evaluate the Effectiveness of Pre-trained GAN Models for Denoising

 Introduction
In several fields, such as satellite imaging, seismic exploration, and general image processing, image
denoising and enhancement are essential duties. Low resolution and noise distortions can seriously
impair the quality of photographs in these domains, making it more difficult to analyze and
comprehend the data. Conventional approaches to these problems often depend on supervised
learning strategies that need large-scale noisy-clean picture pairings, which are hard to come by in
practical applications.
GMR Institute of Technology

Innovative methods to address these issues have been brought about by recent developments in deep
learning. Without requiring paired datasets, Generative Adversarial Networks (GANs) have
demonstrated potential in improving picture quality and eliminating noise. Notably, versions like
CycleGAN and Noise2Noise are examples of these. These techniques efficiently denoise and
reconstruct pictures from unpaired data by utilizing adversarial loss and cycle-consistency loss,
which makes them more applicable to real-world scenarios where labeled data is hard to come by.
GAN-based techniques have been modified for multispectral satellite imagery to handle the unique
problems
. caused by resolution mismatches and sensor noise. Wavelet transformations are used in
methods such as the wavelet directional cycle-consistent adversarial network (WavCycleGAN) to
retain high-frequency information while eliminating noise, producing satellite pictures that are
crisper and more detailed.
Similar to this, the denoising of seismic profiles has been enhanced in seismic exploration through
the use of GANs in conjunction with residual learning and data augmentation techniques. This
method not only improves noise reduction but also makes sure that the priceless seismic signals are
not damaged too much, which makes analysis more precise.
5
 Literature survey
Reference 1
Qu, Z., Zhang, Y., Sun, Y., & Lin, X(2018), A New Generative Adversarial Network
for Texture Preserving Image Denoising ,IEEE,5356-5361
• Deep learning techniques for picture denoising are reviewed.
• Different strategies were considered, including residual networks, CNNs, and
GMR Institute of Technology

autoencoders.
• Adversarial training used for de-raining and denoising grayscale images.
• Adversarial training using innovative network topology for color picture
denoising.
Reference 2
Samik Banerjee, Sukhendu Das(2020), SD-GAN: Structural and Denoising GAN
reveals facial parts under occlusion,arxiv,1-24

• uses generative adversarial models to address occlusion in face recognition.

• suggests the SD-GAN model for improving face recognition and
completion.
• Provides structural loss and denoising for improving and reconstructing
images.
 Literature survey
Reference 3
Xin Cheng , Jingmei Zhou , Jiachun Song, and Xiangmo Zhao,(2023), A Highway
Traffic Image Enhancement Algorithm Based on Improved GAN in Complex
Weather Conditions,IEEE,8716-8726
GMR Institute of Technology

• Traditional methods include filter transformation, and atmospheric scattering

models.
• suggested a combined super-resolution and sharpness-enhancement technique.
• proposed a defogging algorithm combining morphological reconstruction.

Reference 4
SONGKUI CHEN 1 , DAMING SHI(2020), Image Denoising With Generative
Adversarial Networks and Its Application to Cell Image Enhancement ,IEEE, 82819-
82831

• CNN-based techniques are widely used for denoising because they take use
of statistical features.
• CNN-based discriminative learning techniques, in particular, prevent
overfitting and provide superior results.
 Literature survey
Reference 5
HAIBO ZHANG 1 AND KOUICHI SAKURA(2022), Conditional Generative
Adversarial Network-Based Image Denoising for Defending Against Adversarial
Attack,IEEE Access, 169031-169043
• Image denoising and model training are two examples of protection strategies
GMR Institute of Technology

against adversarial assaults.

• The suggested defense methods include adversarial loss, defense-GAN, and
reconstruction
• Techniques like Universal Adversarial Perturbations and FGSM(Fast Gradient
Signed Method) are covered.
Reference 6
Shaobo Zhao, Sheng Lin , Xi Cheng(2024), Dual-GAN Complementary Learning for
Real-World Image Denoising,IEEE, 355-366
• Current approaches rely on a solitary GAN, intricate architecture, and training
difficulties.
• The suggested DGCL approach performs better than comparable denoising
methods.
• For modular complementary learning in image denoising, DGCL uses dual GAN.
 Literature survey
Reference 7
HAIJUN HU 1 , BO GAO3 , ZHIYUAN SHEN(2020), Image Smear Removal via
Improved Conditional GAN and Semantic Network,IEEE Access, 113104-113111

• Pay attention to GAN, Semantic Network, and mesh smear removal techniques.
GMR Institute of Technology

• Sets itself apart from jobs involving super-resolution and denoising.

Reference 8
Xin Jin1 , Ying Hu1 , Chu-Yue Zhang1(2020), Image restoration method based on GAN
and multi-scale feature fusion ,IEEE Xplore, 2305-2310

• focuses on GAN and multi-scale feature fusion while comparing several image
restoration techniques.
• Assesses the precision and aesthetic coherence of current picture restoration
techniques.
• focuses on GAN and multi-scale feature fusion while comparing several image
restoration techniques.
.
 Literature survey
Reference 9
LUAN THANH TRINH AND TOMOKI HAMAGAMI(2024), Latent Denoising
Diffusion GAN: Faster Sampling, Higher Image Quality,IEEE Access, 78161-78172

• Diffusion models are superior in literature in terms of picture quality and diversity.
GMR Institute of Technology

• Real-time applications of diffusion models are hampered by slow inference speed.

• makes use of weighted learning to improve image quality and diversity.

Reference 10
IBRAHIM H. EL-SHAL 1 , OMAR M. FAHM(2022), License Plate Image Analysis
Empowered by Generative Adversarial Neural Networks (GANs) ,IEEE Access,
30846-30857

• LP detection with SRGAN for high-resolution augmentation in naturalistic

photos.
• The accuracy of license plate recognition is enhanced by the proposed framework
by using realistic super-resolution photos.
• GAN-based end-to-end deep learning model for LP recognition.
 Literature survey
Reference 11
Huihui Li , Cang Gu , Dongqing Wu , Gong Cheng/(2022), Multiscale Generative
Adversarial Network Based on Wavelet Feature Learning for SAR-to-Optical Image
Translation,IEEEE, 5236115-5236125
GMR Institute of Technology

• compares S2OIT approaches, focusing on realism and detail in generated

images.
• Previous methods aren't realistic or clear in terms of style or details.
• WFLM-GAN produces images that are more meaningful and have more
distinct details.
Reference 12
WENDA LI 1,2,3 AND JIAN WANG(2021), Residual Learning of Cycle-GAN
for Seismic Data Denoising,IEEE Access, 11585-11597

• In denoising, transform domain algorithms are popular and well developed.

• The F-x deconvolution algorithm is commonly used for denoising seismic
signals.
• On the basis of dictionary learning, adaptive transformation algorithms are
created.
 Literature survey
Reference 13
Qi‑Feng Sun1 · Jia‑Yue Xu1 · Han‑Xiao Zhang1(2021), Random noise suppression
and super‑resolution reconstruction algorithm of seismic profle based
on GAN,Springer, 2107–2119
GMR Institute of Technology

• Put your attention on GAN for noise reduction and improved seismic profiles.
• Reconstruction was done using a back-projection unit and residual learning.
• Convolutional neural networks were used to implement discriminators for data
features.

Reference 14
ASAVARON LIMSUEBCHUEA 1 , RAKKRIT DUANGSOITHONG(2024), Self-
Augmented Noisy Image for Noise2Noise Image Denoising,IEEE Access, 71076-
71087
• uses variance in loss function for self-augmented network training.
• Addresses issues in blind noise settings with self-augmented noisy images.
• Compares Noise2Noise and Noise2Clean techniques in image denoising.
 Literature survey
Reference 15
Joonyoung Song , Jae-Heon Jeong , Dae-Soon Park (2021), Unsupervised
Denoising for Satellite Imagery Using Wavelet Directional CycleGAN,IEEE
Xplore, 6823-6839
GMR Institute of Technology

• Conventional methods use total variation and low-rank matrix recovery for
denoising.
• Proposed WavCycleGAN for multispectral denoising due to structured noise
patterns.\
• Removes noise, preserves high-frequency features in multispectral satellite imagery
effectively.
 GMR Institute of Technology

Graphical Representation

14
 Comparison Table
Title year Objectives Limitations Advantages Performance metrics

Reference 1 A New Generative Adversarial 2018 • Propose GAN for image denoising with texture • Preservation of • Superiority in image • PSNR and
Network for Texture Preserving preservation details, texture denoising over other SSIM(Structural Similarity
Image Denoising • Compare new method with existing denoising restoration, and methods. Index) used for
techniques for superiority. special noise • Unique generator quantitative evaluation of
removal. network preserves denoising methods.
texture features and • PSNR and SSIM values
surface details compared for gray and
color denoised images

Reference 2 SD-GAN: Structural and 2020 • Develop a generative model for face • Occlusion hinders • SD-GAN • Evaluation metrics include
Denoising GAN reveals facial reconstruction under occlusion. face recognition reconstructs faces PSNR and SSIM for
GMR Institute of Technology

parts under occlusion • Propose a novel bimodal training algorithm for accuracy. under occlusion for quantitative assessment.
GAN convergence. • GANs generate efficient face • PSNR captures pixel value
noise due to missing recognition. differences between
pixels in output • Outperforms other generated and reference
images methods in face images
recognition
performance

Reference 3 A Highway Traffic Image 2023 • Enhance highway images in complex weather • Low illumination • Improved GAN • PSNR and SSIM used for
Enhancement Algorithm Based conditions using improved GAN. affects image enhances images in quantitative evaluation of
on Improved GAN in Complex • Combine attention mechanism and multiscale quality and feature complex weather image enhancement.
Weather Conditions, fusion for image feature enhancement recognition. conditions • PSNR measures image
• Challenges in effectively. distortion, while SSIM
processing images • Algorithm achieves evaluates image similarity.
due to insufficient higher PSNR and
light at night SSIM values
compared to others

Reference 4 Image Denoising With 2020 • Analyze blurriness issue in CNN-based • Blurriness issue in • WGAN-based • Average PSNR compared
Generative Adversarial Networks denoising methods. denoised images framework resolves between ID-MSE, ID-
and Its Application to Cell Image • Propose WGAN-based training framework for due to MSE-based blurriness issue in MSE-WGAN(WF), and
Enhancement image denoising methods. denoising images ID-MSE-WGAN.
• Difficulty in effectively. • ID-MSE-WGAN
recovering feature • Achieves better outperformed state-of-the-
details with current denoising art methods in denoising
denoising methods performance
compared to state-
of-the-art methods
 Title year Objectives Limitations Advantages Performance metrics

Reference 5 Conditional Generative 2022 • Propose image denoising method to defend • Long model • Proposed • Recovery accuracy
Adversarial Network-Based against adversarial image attacks. training time for method 20.22% higher than
Image Denoising for Defending • Enhance image recovery efficiency using traditional enhances image state-of-the-art.
Against Adversarial Attack Conditional Generative Adversarial image denoising recovery with • Classification results
Networks methods. perceptual loss significantly improved
• Limited addition. with image recovery
verification with • Defends against
only FGSM and adversarial
PGD attacks attacks by pre-
denoising input
images
Reference 6 Dual-GAN Complementary 2024 • Propose DGCL strategy for real-world • Complex • DGCL integrates • Evaluation metrics:
Learning for Real-World Image image denoising based on modular learning. network denoised image PSNR and SSIM used
GMR Institute of Technology

Denoising • Verify DGCL network on PolyU and SIDD structure and and noise for image quality
datasets for performance evaluation training learning for assessment.
difficulty in optimal results. • DGCL superior in
single GAN • DGCL uses PSNR and SSIM
methods modular compared to other
• Improvement networks for denoising methods
potential in the easier training
dual-GAN and performance
complementary improvement
learning
strategy
Reference 7 Image Smear Removal via 2020 • Develop method for mesh stains removal • Weak • Removes mesh • Peak Signal to Noise
Improved Conditional GAN and using GAN and Semantic Network. discriminator stains effectively Ratio (PSNR) and
Semantic Network, • Achieve high PSNR and SSIM on 1000 led to fast loss using Improved Structural Similarity
testing images convergence in Conditional Index (SSIM)
early training GAN. • Gain Ratio (GR)
stages. • Utilizes assesses the gain ratio
• No specific Semantic of each score
limitations Network for
mentioned in fine-tuning
the provided image details
contexts. post
decontamination

Reference 8 ), Image restoration method based 2020 • Traditional • Improved image • Peak signal-to-noise
on GAN and multi-scale feature • Propose image restoration method using autoencoder has accuracy and ratio (PSNR) for
fusion GAN and multi-scale feature fusion. poor restoration visual quantitative evaluation
• Enhance accuracy and visual consistency of due to high consistency in of data sets.
restored images. reconstruction restoration • Mean square error
loss method. (MSE) between
• DCGAN lacks • Enhanced restored and original
image feature images
continuity and extraction and
detail expression
restoration in quality
Dec
ber
em

16
17
20

image compared to
4

restoration traditional
 Title year Objectives Limitations Advantages Performance metrics
Reference 9 Latent Denoising Diffusion 2024 • Introduce Latent • Slow inference • LDDGAN reduces • Inference time
GAN: Faster Sampling, Denoising speed due to input image size measured by
Higher Image Quality Diffusion GAN for Gaussian significantly, number of function
image quality and assumption in enhancing model's evaluations and
speed denoising step. speed. average time
enhancement. • Diffusion models • Low-dimensional • Sample fidelity
• Utilize pre-trained require hundreds or latent space evaluated using
autoencoders for thousands of improves quality Frechet Inception
compact latent denoising steps and diversity of Distance (FID)
space in diffusion generated samples metric
models
Reference 10 ), License Plate Image 2022 • Enhance LP • Limited to ideal • Enhances license • PSNR and SSIM
GMR Institute of Technology

Analysis Empowered by detection in training data with plate recognition used for
Generative Adversarial naturalistic restricted scenarios accuracy through performance
Neural Networks (GANs) environments using • Relies on low- super-resolution evaluation.
GAN-based SR resolution images techniques. • SRGAN
techniques in surveillance • Improves image outperformed other
• Improve LP image systems quality by methods in visual
quality through generating realistic analysis
SRGAN network super-resolution
modifications and images
TV regularization

Reference 11 Multiscale Generative 2022 • Develop WFLM- • Previous methods • SAR has all-day, • SSIM, PSNR, FID,
Adversarial Network Based GAN for SAR-to- lack realism and all-weather imaging and KID used for
on Wavelet Feature optical image clarity in generated capabilities, translation effect
Learning for SAR-to- translation with images. valuable in remote evaluation.
Optical Image Translation improved • Optimization for sensing • SSIM measures
interpretation. the model is still • SAR penetrates image similarity,
• Address key issues limited in multiple masks, reveals higher value
by refining SAR aspects hidden target info, indicates better
image content and crucial in military quality
detail information apps

Reference 12 Residual Learning of Cycle- 2021 • Propose RCGAN • No explicit mention • GANs consider • SNR used for
GAN for Seismic Data for seismic data of computational seismic data denoising quality
Denoising denoising with limitations in the structure, assessment.
residual learning. paper enhancing • Adversarial losses
• Improve training denoising and cycle
efficiency, diversity efficiency. consistency loss in
of samples, and • RCGAN method the objective
denoising improves denoising function
parameters performance while
reducing manual
intervention

17
 Title year Objectives Limitations Advantages Performance metrics
Reference 13 Random noise suppression 2021 • Develop noise • No specific • Improved seismic • Average SNR,
and super‑resolution suppression and limitations were profile quality, PSNR, and SSIM
reconstruction algorithm super-resolution mentioned in the SNR, and values evaluated
of seismic profle based algorithm for research paper resolution. the reconstruction
on GAN seismic profiles • Clearer performance.
• Improve signal-to- observations of • MSE loss function
noise ratio and geological ensured accuracy of
resolution of structures like noise prediction
seismic cross- fractures and image integrity
sections
GMR Institute of Technology

Reference 14 ), Self-Augmented Noisy 2024 • Improve Noise2Noise • Challenges in • Self-augmented • PSNR and SSIM
Image for Noise2Noise for single image and single image and strategy improves metrics used for
Image Denoising blind noise scenarios. blind noise Noise2Noise performance
• Introduce self- scenarios for denoising for single evaluation.
augmented noisy denoising. image scenarios. • Comparison based
images for single image • Limitations in • Achieves on PSNR and SSIM
denoising generating noisy comparable metrics in
image datasets, performance to experimental
especially with low unsupervised results
noise levels denoising methods
without extra
augmentation
Reference 15 Unsupervised Denoising for 2022 • Develop • Supervised learning • Wavelet transform • PSNR and SSIM
Satellite Imagery Using unsupervised requires paired detects fine details used for
Wavelet Directional denoising method noisy and clean with minimal quantitative
CycleGAN for satellite images for training. support length. evaluation of
imagery using • Difficulty in • Daubechies denoising methods.
WavCycleGAN. collecting real wavelets provide • Identity loss and
• Remove noise and paired satellite sparse adversarial loss
preserve high- images for training representation for used in the
frequency features models clean and noisy generator-
in satellite images signals discriminator
effectively framework

18
 References
1. Qu, Z., Zhang, Y., Sun, Y., & Lin, X(2018), A New Generative Adversarial Network for Texture Preserving Image
Denoising ,IEEE,5356-5361
2. Samik Banerjee, Sukhendu Das(2020), SD-GAN: Structural and Denoising GAN reveals facial parts under
occlusion,arxiv,1-24
3. Xin Cheng , Jingmei Zhou , Jiachun Song, and Xiangmo Zhao,(2023), A Highway Traffic Image Enhancement
Algorithm Based on Improved GAN in Complex Weather Conditions,IEEE,8716-8726
4. SONGKUI CHEN 1 , DAMING SHI(2020), Image Denoising With Generative Adversarial Networks and Its
GMR Institute of Technology

Application to Cell Image Enhancement ,IEEE, 82819-82831

5. HAIBO ZHANG 1 AND KOUICHI SAKURA(2022), Conditional Generative Adversarial Network-Based Image
Denoising for Defending Against Adversarial Attack,IEEE Access, 169031-169043
6. Shaobo Zhao, Sheng Lin , Xi Cheng(2024), Dual-GAN Complementary Learning for Real-World Image
Denoising,IEEE, 355-366
7. HAIJUN HU 1 , BO GAO3 , ZHIYUAN SHEN(2020), Image Smear Removal via Improved Conditional GAN and
Semantic Network,IEEE Access, 113104-113111
8. Xin Jin1 , Ying Hu1 , Chu-Yue Zhang1(2020), Image restoration method based on GAN and multi-scale feature
fusion ,IEEE Xplore, 2305-2310
9. LUAN THANH TRINH AND TOMOKI HAMAGAMI(2024), Latent Denoising Diffusion GAN: Faster Sampling,
Higher Image Quality,IEEE Access, 78161-78172
10. IBRAHIM H. EL-SHAL 1 , OMAR M. FAHM(2022), License Plate Image Analysis Empowered by Generative
Adversarial Neural Networks (GANs) ,IEEE Access, 30846-30857
11. Huihui Li , Cang Gu , Dongqing Wu , Gong Cheng/(2022), Multiscale Generative Adversarial Network Based on
Wavelet Feature Learning for SAR-to-Optical Image Translation,IEEEE, 5236115-5236125
12. WENDA LI 1,2,3 AND JIAN WANG(2021), Residual Learning of Cycle-GAN for Seismic Data Denoising,IEEE
Access, 11585-11597
13. Qi‑Feng Sun1 · Jia‑Yue Xu1 · Han‑Xiao Zhang1(2021), Random noise suppression and super‑resolution reconstruction
algorithm of seismic profle based on GAN,Springer, 2107–2119
14. ASAVARON LIMSUEBCHUEA 1 , RAKKRIT DUANGSOITHONG(2024), Self-Augmented Noisy Image for
Noise2Noise Image Denoising,IEEE Access, 71076-71087
15. Joonyoung Song , Jae-Heon Jeong , Dae-Soon Park (2021), Unsupervised Denoising for Satellite Imagery Using
Wavelet Directional CycleGAN,IEEE Xplore, 6823-6839
 GMR Institute of Technology

20