Volume 8, Issue 6, June 2023 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Image Restoration using Deep Learning

Dr. Hansaraj Wankhede, Abhishikth Thul, Deep Malviya, Pratik Pathe, Shivam Kuite
GHRCE Nagpur, India

Abstract:- Image restoration using deep learning more accurate and visually appealing image restoration
attempts to create an image recovery system that can outcomes. Deep learning-based image restoration utilising
restore outdated and corrupted images regardless of the the U-Net architecture has yielded promising results in a
type of noise present. Photos play an important role in variety of applications such as denoising, deblurring, super-
capturing and preserving cherished moments in today's resolution, and inpainting. The U-Net learns the underlying
digital age. However, due to a variety of environmental mapping between the corrupted input and the clean output by
conditions, these images can get distorted over time. training it on a large dataset of matched clean and corrupted
Manual recovery is time-consuming and labor-intensive, images, allowing it to generalise well to unseen data.
emphasizing the need for an automated alternative. Deep
learning techniques, notably convolutional neural II. LITERATURE REVIEW
networks, are used in the proposed system, which has
demonstrated promising results in image processing The study's authors, Olaf Ronneberger, Philip Fischer,
tasks. The essay goes over these approaches in detail, and others [1], suggested a network and training strategy that
focusing on image noise reduction, deblurring, dehazing, relies significantly on data augmentation to make better use
and super-resolution. Different network topologies are of the annotated samples provided. Offers. The layout is
investigated, including those with residual or merge-skip made up of contracting routes that collect context and
connections, as well as their receptive fields and the expanding paths that allow for exact localisation. They
discovered that this sort of network could be trained
usage of unsupervised autoencoder processes. The study
consistently from a very small number of photos in his ISBI
also looks at image quality criteria to see how helpful
they are in image recovery. An effective deblurring task of segmenting neural structures within an electron
network and adaptable training algorithms for high- microscopy stack, and the best answer to date (sliding
resolution recovery tasks are suggested to handle the window convolutional networks) has been proved to
special difficulty of deblurring. The proposed method's outperform. Jiwon Kim, Jung Kwon Lee, and Kyung Mu Lee
performance is compared to state-of-the-art approaches [2] offer a highly accurate single-frame super-resolution (SR)
utilizing both quantitative and qualitative assessments. technique. Our method use a deep convolutional network
The study finishes with a discussion of prospective future inspired by the Image Net classification tool VGG-net. They
research topics and outstanding challenges in image discovered that raising the network's depth considerably
recovery. The ultimate goal of this research is to create a increased accuracy. The final model employs 20 weight
robust and efficient image recovery system capable of layers. Contextual information for vast picture areas may be
restoring photos to their original quality regardless of successfully obtained by cascading tiny filters numerous
the type or severity of corruption or noise present. times in deep network topologies.

Keywords:- Image recovery, Noise reduction, Neural Kai Zhang, Wangmeng Zuo, and colleagues [3] present
networks, Image restoration, Deep learning, Image FFDNet, a rapid and adaptable convolutional noise reduction
processing, Super-resolution, Deblurring. neural network that uses a noise level map as input. When
applied to downsampled sub-images, his suggested FFDNet
I. INTRODUCTION performs well in terms of both inference time and noise
reduction performance. FFDNet has several advantages over
Image restoration is a key activity in image processing existing discrimination noise suppressors, including: B. (i)
that aims to recover the original quality of images that have the ability to efficiently manage a wide range of noise levels
been distorted by causes such as noise, blur, or other flaws. (i.e. [0, 75]) in a single network, (ii) spatially diverse noise
Image restoration approaches have seen considerable gains in by specifying a non-uniform noise level map, and (iii) faster
terms of accuracy and efficiency as a result of recent than benchmark BM3D even on CPU without sacrificing
advances in deep learning, notably convolutional neural noise reduction efficiency.
networks (CNNs). The U-Net architecture is differentiated by
its U-shaped network design, which incorporates encoder and Pengju Liu, Hongzhi Zhang, and colleagues [4]
decoder paths. The encoder path uses down-sampling to introduce a novel multilevel wavelet CNN (MWCNN) model
collect contextual information from the input picture, to achieve a better balance between received field size and
whereas the decoder path uses up-sampling and skip processing efficiency. To minimise the size of the feature
connections to restore spatial features and provide a maps in the reduced subnetwork, a wavelet transform is
reconstructed image. The skip connections of the U-Net introduced to the modified U-Net design. To lower the
architecture enable the integration of The high-resolution number of feature map channels, a second convolutional
feature maps from the encoder path are mixed with the up- layer is also employed. An inverse wavelet transform is then
sampled feature maps from the decoder path. This integration used to build a high-resolution feature map on the expanding
preserves tiny features while allowing the model to exploit subnetwork.
contextual knowledge learned at different sizes, resulting in

IJISRT23JUN621 www.ijisrt.com 682

Volume 8, Issue 6, June 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
Yulun Zhang, Yapeng Tian, and colleagues [5] provide dataset, as well as visual and metric analogies of operation
a novel and effective residual dense network (RDN) to utilising the trained models.
address this IR problem by improving the trade-off between
efficiency and effectiveness in utilising hierarchical Chunwei Tian and Lunke Fei et al. [10] performed a
information from all convolution layers. We recommend preliminary analysis of deep image denoising techniques in
employing tightly coupled convolutional layers to extract this paper. First, identify an additive white noise image using
many local characteristics using residual dense block (RDB). a deep convolutional neural network (CNN). Deep CNN for
RDB also allows for direct connections between the state of real-world noise. A deep CNN for blind denoising, and a
one RDB and each layer of the current RDB, resulting in a deep CNN for hybrid noise pictures that represent a mix of
contiguous memory system. They proposed using local noisy, fuzzy, and low-resolution images. The motives and
feature fusion in RDB to stabilise bigger network training concepts of various deep learning approaches are then
and adaptively learn more effective features from earlier and examined.
current local features.
III. METHODOLOGY
Saeed Anwar and Nick Barnes [6] present a unique
single-stage blind real image denoising network (RIDNet) We devise an approach to address the damage in the
based on a modular design in their paper. To simplify the image. These flaws include blurring and noise. The most
flow of low-frequency information and to take advantage of prevalent problems that can occur in most settings are blur
channel dependencies, we employ a residual on the residual and noise. Several ways have been developed to fix any
structure. A comparison of our approach to 19 cutting-edge damage that may occur in the image. One of the most well-
algorithms on three synthetic and four noisy datasets in terms known ways is to employ deep learning techniques learned to
of quantitative measurements and visual quality gives more remove blur, noise in photos, and add colour to black-and-
proof of our RIDNet's superiority. In this study, Yiyun Zhao, white images; the technology has evolved through time.
Zhuqing Jiang, and colleagues [7] introduce a unique three- Image quality has steadily increased over time. The layers of
stage pyramid real image denoising network (PRIDNet). The a convolutional neural network (CNN) are convolutional,
noise estimation step begins by adjusting the channel activation, and pooling. To increase neural network
significance of the input noise using a channel attention performance and generalisation, these layers extract image
mechanism. Second, pyramid pooling is used to extract properties, introduce nonlinearities, and reduce feature
multiscale features during the multiscale denoising stage. dimensionality. CNNs use convolutional layers for feature
Finally, at the feature fusion step, a kernel-selecting extraction, which keeps the spatial relationships between
technique is used to fuse multi-scale features adaptively. The pixels in the original image. To obtain the result, a
technique can compete with state-of-the-art denoisers in convolution kernel or filter moves a picture pixel by pixel,
terms of both quantitative metrics and visual perception computes element-wise multiplication, and adds the pixels. A
quality, according to testing on two datasets of genuine noisy feature map is the resulting matrix. A convolution kernel,
photographs. often known as a deep learning filter, is a set of two-
dimensional weight arrays. Each 2D weight array is applied
Chong Mou, Qian Wang, and colleagues [8] present a to all of the input channels of the previous level, resulting in
Deep Generalised Unfolding Network (DGUNet) for picture numerous channels that are summed to produce one. This
restoration in this study. We specifically integrate a gradient technique is used to create multiple output level feature maps
estimation mechanism into the gradient descent phase to for all 2D weight arrays. As a result, given the same input
allow the Proximal Gradient Descent (PGD) approach to image, various weighted convolution kernels create different
manage complicated and real-world picture degeneration feature maps. As a consequence, different convolution
without sacrificing interpretability. We also develop multi- kernels extract different information. Lower convolutional
scale and spatially adaptive inter-stage information channels layers collect a high amount of structural detail and location
spanning proximal mapping in several PGD cycles to information from images, but upper layer feature maps
overcome the inherent information loss in the majority of incorporate more semantic meaning and less location
deep unfolding networks (DUN). In this study, Manish information. CNNs are a deep learning network design that is
Yadav and Namita Tiwari [9] will focus on cutting-edge commonly used for image recognition and pixel data
CNN-based denoising algorithms and give a comparative processing.We utilise a very deep convolutional network
evaluation of numerous approaches. Use it as a basic network inspired by Simonyan and Zisserman [2] for SR picture
and for fingerprint improvement now that you've learnt all of reconstruction. We utilise d layers, with the first and last
the standard image-denoising approaches. Experiments were layers of the same type: 64 filters of 3 x 3 x 64 size, with
undertaken to illustrate the efficiency and correctness of the each filter acting on a 3 x 3 spatial region over 64 channels.
model learnt only based on the description of the picture The first layer is responsible for working with the incoming
picture. The final layer is a single 3 x 3 x 64 filter that is used
for picture reconstruction.

IJISRT23JUN621 www.ijisrt.com 683

Volume 8, Issue 6, June 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165

Fig. 1: Workflow of Model

The procedure begins with the installation of the a reduction in the concentration of blur visible in the image.
essential libraries and packages for data processing, In the actual world, images deteriorate quickly, and noise is
visualisation, and machine learning. The data is then complicated. As a result, strategies for blind denoising are
preprocessed to account for missing values, outliers, and needed. An FFDNet used noise level and noise as CNN input
categorical factors before being divided into training and to train a denoiser for unknown noisy pictures. To address
testing sets. Following that, data visualisation methods are the issue of blind denoising, many techniques were proposed.
used to acquire insights into the dataset and find patterns or Kenzo et al. [10] suggested an image device approach for
anomalies. To suit the model's needs, the data is further blind denoising that employed soft shrinkage to adjust the
altered via slicing and reshaping. The architecture, layers, noise level. It was efficient to use CNNs to estimate noise in
and activation functions of the model are chosen. After that, unpaired noisy pictures.
it is assembled with a loss function, optimizer, and
evaluation metrics. Backpropagation is used to update the C. Dealing with black and white images:
weights of the model as it is trained on the training data. When working with ancient photographs, the images
Once trained, the model's performance on validation or test obtained were primarily black and white; no colors were
data is assessed using measures such as accuracy or discovered. This is a subset of the process of constructing the
precision. The model's predictions are visualised to assess matrices so that the neural network model may be trained to
their accuracy and indicate possible areas for improvement. deliver meaningful colours to black and white images. Few
Finally, the model is evaluated on real, unseen photos to methods have been created that can reliably allocate colour
determine how well it performs in real-world circumstances. to a picture. This necessitates a massive amount of memory
for training the model. The core method was captured using a
A. Dealing with Blurry images: model trained with a simple neural network that can partially
The collection includes both fuzzy and clear photos. For colorize a black and white image.
this challenge, we will be working with crisp photos. Now
that we have sharp photographs, we will develop an D. Deep Learning Models:
algorithm called Gaussian Blur, which will apply blur to the U-Net is a deep learning architecture that was originally
input images and store the resulting images in a new folder designed to segment medical images. It has, however, been
named 'Gaussian Blur'. These photos are now ready for usage effectively used to a variety of other image processing
by the deep learning model. applications, including picture restoration. The process of
restoring a high-quality image from a damaged or noisy
The model chooses two photos from a collection of ten, image is known as image restoration. This may be
and the two images are blended. These blur photos are fed accomplished by training the network on pairs of degraded
into the model, and at each iteration (epoch), a decreased blur and high-quality photos using U-Net. The network learns to
concentration image is recorded in the saved output folder. map the degraded picture to the high-quality image during
As a result, the concentration of the blur is shown to decrease training.
as the number of epochs is increased. We can map the
difference in the amount of blur that is decreased from the An encoder and a decoder comprise the U-Net
previous iteration thanks to the saved image at each iteration. architecture. The encoder collects information from the input
picture and decreases the image's spatial resolution. The
B. Dealing with the noisy images: encoded characteristics are then sent to the decoder, which
The deep learning model that has been efficiently trained gradually raises the spatial resolution to generate the final
on high resolution photos is employed for this purpose. output image. Skip connections connect the encoder and
picture denoising methods based on neural networks have decoder, allowing the network to absorb fine-grained features
revolutionized picture analysis. The model is built on a from the input picture.
neural network that has been trained to minimize the amount
of blur in the image. As a result, the picture is intensively
processed by the neural network at each iteration, resulting in

IJISRT23JUN621 www.ijisrt.com 684

Volume 8, Issue 6, June 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
The deteriorated picture is supplied into the encoder in Two main level architecture that were considered in
the case of image restoration, and the network learns to map this work are
it to the high-quality image. The network uses the degraded  U-Net Architecture.
picture as input and provides the repaired image as output  Reformed U-Net
during inference. U-Net has been used effectively for image
restoration tasks such as denoising, deblurring, and super- The designing details of the above-said two
resolution. It has been demonstrated that it outperforms architectures are explained as below.
standard picture restoration algorithms as well as alternative
deep learning architectures.  U-Net Architecture
In our investigation, we are cautiously delving into the
U-Net architecture as the basis generator model. The U-net
architecture, which has piqued our interest. The network's
architecture model, which creates a U shape, is shown below.
On the right side, the network basically encodes to compress
and decodes to rebuild the denoisy image.

Fig. 2: U-net Architecture [11]

This U-Net is utilised for more than just picture DCGAN (Deep Convolutional Generative Adversarial
segmentation. Image restoration applications are also Networks) is a picture generating architecture. It is made up
common. Segmentation aids in reducing picture size while of two networks: a generator network that learns to generate
capturing characteristics. It reduces vast amounts of data to a realistic pictures from random noise and a discriminator
few pixels by extracting features. Based on global and local network that attempts to discern between created and actual
parameter information, the threshold may be used to rebuild images. By combining U-Net with DCGAN, an architecture
a picture. Local feature extractions can be used to recover capable of producing high-quality pictures while maintaining
lost photos. Convolution generates a vast number of fine-grained features may be created. This hybrid
parameters for training the model. Convolution layers read architecture is commonly referred to as U-Net GAN or U-
the real image size input and output dependent on the number GAN.
of neurons. Kernel size, padding size (if used), and stride the
kernel takes when convoluting are all specified in those The U-GAN architecture combines the U-Net encoder
layers. We may use numerous layers of convolution to help and decoder structure with the DCGAN generation network.
the model learn better. The generator network receives a random noise vector as
input and creates an image, which is then processed by U-
 LM U-NET DCGAN: Net's encoder and decoder structures to produce the final
Deep Conditional GANs' performance is improved by output. In U-GAN, the discriminator network is often
replacing pooling layers with fractional-stride convolutions designed to differentiate between produced and actual
and employing Batch Normalisation (Batch Normalisation pictures while simultaneously giving input to the generator
normalises layer input and allows for a greater learning rate network to enhance the quality of the generated images. U-
to be applied to each layer's output in the model). GAN has been utilised for image generating tasks like as
super-resolution and inpainting. It has been demonstrated
that it can deliver high-quality outcomes while keeping fine-
grained features.

IJISRT23JUN621 www.ijisrt.com 685

Volume 8, Issue 6, June 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
IV. DATA COLLECTION & PLATFORMS USED

V. DATASET VI. DESIGN, IMPLEMENTATION AND

MODELLING
The DIV2K dataset is broken into the following
sections:  Step-1: import all the necessary libraries and download
the dataset
 Train data: We extract similar low resolution images
from 700 high definition high resolution photographs and  Step-2: Data Preparation
give both high and low resolution images for downscaling
factors of 2, 3, and 4.  We construct a function named sorted alphanumeric in the
given code to sort a list of files in alphanumeric order.
 Data for validation: The low quality photographs are
This function will come in handy during our further
made available at the start of the challenge in order for
processing.
participants to obtain online feedback from the validation
server; the high resolution images will be made available  Next, we define the required picture size as 256x256
when the challenge's final phase begins. pixels. To hold the produced high-quality photos, we
build an empty list named high-img.
 Test data: When the final evaluation phase begins, the
participants will receive the low quality photographs, and  We have a directory location where we keep the high-
the results will be revealed once the challenge is quality photos. We get a list of files in that directory by
concluded and the winners are determined. using the os.listdir function, which we then sort in
alphabetic order with the sorted alphanumeric function.
Context The dataset was produced to test the blur  We begin a loop that will run through each file in the
detection method. The dataset might be used to evaluate sorted list. We exit the loop if the current file is '891.png'
picture deblurring methods. As a result, while CNN and U- since we want to exclude it. Otherwise, we read the
net cannot distinguish between fuzzy and clear images, picture with OpenCV, which reads BGR images. To deal
high-quality photographs may be used for visual with the RGB format, we use cv2.cvtColor to transform
comparison. the image from BGR to RGB.
 Next, we use cv2.resize to resize the picture to the
requested size of 256x256 pixels. After resizing, we
divide the pixel values by 255.0 to normalise them and
transform the image to a floating-point format. Finally, we
use img_to_array to convert the picture to an array and
attach it to the high_img list.

IJISRT23JUN621 www.ijisrt.com 686

Volume 8, Issue 6, June 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
 Next, we do identical actions on low-quality photos. To the high and low image loops. This signifies that the loop
hold the produced low-quality photos, we build another will end after processing all photos except '891.png'.
empty list named low_img. We change the directory path Remove the break statements if you wish to include
to the location of the low-quality photographs and repeat '891.png' in the processed pictures.
the process of reading, converting, resizing, normalising,  In general, this code allows us to load and preprocess a set
and converting the images to arrays. of high-quality and low-quality photos for subsequent
 It's worth noticing that the code contains a break analysis or image-processing activities.
statement when the image '891.png' is encountered in both

 Step-3: The Input Image include images ranging from index 700 to 829. The test
sets ('test_high_image' and 'test_low_image') include
 Step-4: Generating Patches photos beginning at index 830.
 The processed high-quality and low-quality photos are  All sets are reshaped to the same size: '(number of
divided into training, validation, and test sets. The photos, SIZE, SIZE, 3)'. The code then prints the
training sets ('train_high_image' and 'train_low_image') training, test, and validation set shapes.
include the first 700 photos, whereas the validation sets
('validation_high_image' and 'validation_low_image')

IJISRT23JUN621 www.ijisrt.com 687

Volume 8, Issue 6, June 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
 Step-5: model configuration phase activation function. It returns the model of the upsampling
Using a U-Net architecture, we created a TensorFlow block
model for an image translation challenge. Here's an  The build_model function is used to construct the whole
explanation of what the code does: U-Net model. It accepts an optional input shape argument
 It imports the TensorFlow modules required to create the and sets the default value for RGB pictures to (256, 256,
model, including the layers module. 3). It defines the input layer based on the input shape. It
 The down function is used to define the U-Net model's then employs a succession of down-sampling blocks
down-sampling (encoder) blocks. It accepts parameters (down function) with varying filter widths. Following
for the number of filters, kernel size, and batch that, it applies a symmetric series of up-sampling blocks
normalisation. It stacks layers using a Sequential model, (up function), concatenating the appropriate down-
which includes a convolutional layer, optional batch sampling block outputs along the way. Finally, a
normalisation, and a LeakyReLU activation function. It convolutional layer is used to create the output image. The
returns a model of the downsampling block. model is built with the tf.keras.Model class and the
 The up function is developed to generate the U-Net supplied inputs and outputs.
model's up-sampling (decoder) blocks. It accepts  The build_model function is called to create the model
parameters for the number of filters, kernel size, and a instance.
dropout regularisation option. It stacks layers using a  The model is invoked using the summary method to
Sequential model, which includes a transposed provide a summary of its architecture, including the
convolutional layer, optional dropout, and a LeakyReLU number of parameters and the forms of the intermediate
outputs.

 Step-6: Training the mode

IJISRT23JUN621 www.ijisrt.com 688

Volume 8, Issue 6, June 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
 Step-7: Modal Accuracy and Validation loss plot

Fig. 4: Modal Accuracy

Fig. 5: Modal Loss

 Step-8: Prediction

IJISRT23JUN621 www.ijisrt.com 689

Volume 8, Issue 6, June 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165

Fig. 6: Visualization of Original, Input and Predicted image.

VII. OUTPUT

Fig. 7: Predicted Image

VIII. TESTING & SUMMARY RESULTS

In the testing phase, we have used an image which contains blur images. Example

Fig. 8: Low quality image

IJISRT23JUN621 www.ijisrt.com 690

Volume 8, Issue 6, June 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
 Importing the necessary libraries:

 Loading the saved model:

 Uploading and opening an image:

 Resizing and preprocessing the input image:

 Making predictions with the model:

IJISRT23JUN621 www.ijisrt.com 691

Volume 8, Issue 6, June 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
 Converting the predicted array to an image:

 Displaying the input image and predicted image:

 Evaluation Metrics.

IX. CONCLUSION
X. FUTURE SCOPE
We employed deep learning techniques, notably the U-
Net model, to fix picture faults such as noise, blurring, and Investigate innovative network topologies (for
low resolution. The restoration accuracy varied from 85% to example, Attention U-Net, Dense U-Net) to increase picture
90% after intensive training. The U-Net design effectively restoration performance and handle complex problems.
gathered both low-level and high-level information while Investigate the use of Generative Adversarial Networks
preserving critical spatial features via skip connections. (GANs) to generate aesthetically attractive and realistic
Further improvements are feasible, however, due to issues image restorations. Extend multi-modal image restoration
such as picture complexity, dataset limits, and trade-offs technologies, allowing pictures to be restored in many
between noise reduction, sharpness, and resolution. To modalities at the same time. Focus on video restoration,
enhance restoration outcomes, future research might focus on including temporal consistency and motion estimation for
extending the training dataset, experimenting with other denoising, deblurring, and resolution enhancement. Using
network topologies, and tweaking hyperparameters. unlabeled data, use self-supervised learning algorithms to
Nonetheless, the attained accuracy shows the promise of improve picture restoration accuracy. Improve current
deep learning in image restoration, namely the U-Net model, models and methods for real-time implementation, allowing
opening the way for applications in medical imaging, for resource-constrained devices and realistic applications.
surveillance systems, and visual content improvement across Create picture restoration algorithms that are domain-specific
sectors. for applications such as medical imaging, satellite images,

IJISRT23JUN621 www.ijisrt.com 692

Volume 8, Issue 6, June 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
and underwater photography. Incorporate human perception- Fischer, and Thomas Brox Computer Science
driven assessment criteria to correctly quantify the quality of Department and BIOSS Centre for Biological
retrieved photographs. Investigate transfer learning and fine- Signalling Studies, University of Freiburg, Germany
tuning methods for leveraging pretrained models to improve [12.] Ignatov_2018_ECCV_Workshops, Ignatov, Andrey
restoration task performance. To solve specific challenges and Timofte, Radu and others, PIRM challenge on
and make significant contributions, stress real-world perceptual image enhancement on smartphones:
applications by collaborating with experts from various report, European Conference on Computer Vision
areas. Exploring these future avenues might lead to (ECCV) Workshops, January
advancements in deep learning-based picture restoration,
boosting restoration accuracy, establishing specialised
methodologies, and enabling practical applications in a wide
range of areas.

REFERENCES

[1.] Ronneberger, O., Fischer, P., Brox, T. (2015). U-Net:

Convolutional Networks for Biomedical Image
Segmentation. In: Navab, N., Hornegger, J., Wells,
W., Frangi, A. (eds) Medical Image Computing and
Computer-Assisted Intervention – MICCAI 2015.
MICCAI 2015. Lecture Notes in Computer Science(),
vol 9351. Springer, Cham.
[2.] J. Kim, J. K. Lee and K. M. Lee, "Accurate Image
Super-Resolution Using Very Deep Convolutional
Networks," 2016 IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), Las Vegas,
NV, USA, 2016.
[3.] Kai Zhang, Wangmeng Zuo, Lei Zhang, “FFDNet:
Toward a Fast and Flexible Solution for CNN based
Image Denoising” 22 May 2018.
[4.] P. Liu, H. Zhang, K. Zhang, L. Lin and W. Zuo,
"Multi-level Wavelet-CNN for Image Restoration," in
2018 IEEE/CVF Conference on Computer Vision and
Pattern Recognition Workshops (CVPRW), Salt Lake
City, UT, USA, 2018.
[5.] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng
Zhong, Yun Fu, “Residual Dense Network for Image
Restoration”, 23 Jan 2020, eprint1812.10477
[6.] S. Anwar and N. Barnes, "Real Image Denoising
With Feature Attention," 2019 IEEE/CVF
International Conference on Computer Vision
(ICCV), Seoul, Korea (South), 2019.
[7.] Y. Zhao, Z. Jiang, A. Men and G. Ju, "Pyramid Real
Image Denoising Network," 2019 IEEE Visual
Communications and Image Processing (VCIP),
Sydney, NSW, Australia, 2019.
[8.] C. Mou, Q. Wang, and J. Zhang, "Deep Generalized
Unfolding Networks for Image Restoration," 2022
IEEE/CVF Conference on Computer Vision and
Pattern Recognition (CVPR), New Orleans, LA,
USA, 2022.
[9.] M. Yadav and N. Tiwari, "Performance Comparison
of Image Restoration techniques using CNN and their
applications," 2021 5th International Conference on
Computing Methodologies and Communication
(ICCMC), Erode, India, 2021.
[10.] Chunwei Tian, Lunke Fei, Wenxian Zheng, Yong Xu,
Wangmeng Zuo, Chia-Wen Lin, Deep learning on
image denoising: An overview, Neural Networks,
Volume 131, 2020, Pages 251-275, ISSN 0893-6080.
[11.] U-Net: Convolutional Networks for Biomedical
Image Segmentation Olaf Ronneberger, Philipp

IJISRT23JUN621 www.ijisrt.com 693