International Journal of Information Management Data Insights 1 (2021) 100004

Contents lists available at ScienceDirect

International Journal of Information Management Data

Insights
journal homepage: www.elsevier.com/locate/jjimei

Review

Generative adversarial network: An overview of theory and applications

Alankrita Aggarwal a, Mamta Mittal b, Gopi Battineni c,∗
a
Department of Computer Science and Engineering, Panipat Institute of Engineering and Technology, Samalkha 132101, India
b
Department of Computer Science and Engineering, G.B. Pant Government Engineering College, Okhla, New Delhi, India
c
Medical Informatics centre, School of Medicinal and Health Products Sciences, University of Camerino, Camerino 62032, Italy

a r t i c l e i n f o a b s t r a c t

Keywords: In recent times, image segmentation has been involving everywhere including disease diagnosis to autonomous
GAN vehicle driving. In computer vision, this image segmentation is one of the vital works and it is relatively compli-
Deep learning cated than other vision undertakings as it needs low-level spatial data. Especially, Deep Learning has impacted
Image mining
the field of segmentation incredibly and gave us today different successful models. The deep learning associ-
Big data
ated Generated Adversarial Networks (GAN) has presenting remarkable outcomes on image segmentation. In this
Literature review
Neural networks study, the authors have presented a systematic review analysis on recent publications of GAN models and their
applications. Three libraries such as Embase (Scopus), WoS, and PubMed have been considered for searching the
relevant papers available in this area. Search outcomes have identified 2084 documents, after two-phase screen-
ing 52 potential records are included for final review. The following applications of GAN have been emerged: 3D
object generation, medicine, pandemics, image processing, face detection, texture transfer, and traffic controlling.
Before 2016, research in this field was limited and thereafter its practical usage came into existence worldwide.
The present study also envisions the challenges associated with GAN and paves the path for future research in
this realm.

1. Introduction network architectures (Lecun, Bengio & Hinton, 2015). It can also be
considered as a secondary field of ML algorithms inspired by the brain
A Generative Adversarial Network (GAN) emanates in the category structure and functionality. In the applications of image identification,
of Machine Learning (ML) frameworks. These networks have acquired speech synthesis, text mining applications by receiving a distinct kind
their inspiration from Ian Goodfellow and his colleagues based on noise of data that hierarchical models can be built by representing probability
contrastive estimation and used loss function used in present GAN distributions. Deep learning dependant on an end to end wireless com-
(Grnarova et al., 2019). Actual working using GAN started in 2017 with munication system with conditional GANs using Deep Neural Networks
human faces to adopt image enhancement that produces better illus- (DNNs) do function of message passing like encoding, decoding, mod-
tration at high intensity. Adversarial networks were fundamentally in- ulation, and demodulation. For this, the right judgement of immediate
spired by the blog that has written by Olli Niemitalo in 2010 but the channel transfer state is required to transfer DNN (Ye, Liang, Li & Juang,
same idea is known as Conditional GAN. 2020).
In the examination of the GAN rigorous impact of 2D to 3D im- The most important feature of deep learning is discriminative models
age conversation, initially, the corresponding dataset has to do live that can relate high dimensional sensory input sent to a class of labels.
data fetching and create the benchmark with key features (Wu, Zhang, These generative models based on deep learning impact are lesser be-
Xue, Freeman & Tenenbaum, 2016). Thereafter, for calculating thresh- cause approximation of obstinate probabilistic computation is difficult
old and suitability score, image merging has to be done. Image data and leads to the utmost chances of judgement (He, Zhang, Ren & Sun,
pre-processing steps involve image segmentation and cleansing which 2016; Lecun et al., 2015). If deep learning models are applied on geni-
follows the GAN training. Outcomes are expected pattern analysis and tive networks then the advantage will be that deep learning models are
exactness of the image generation. Fig. 1 presents the example of a 3D work on big datasets. These datasets are largely dependant on high-end
generative adversarial network with five volumetric conventional neu- machines and took a long time to do model training and less time for
ral layers of 2 strides and kernel sizes 4 × 4 × 4. testing. Applications of GAN networks are exploring contemporary ad-
Deep learning techniques could be used as generative models. Deep vancements and accomplishing our daily life needs.
learning is an idea neural networks with many layers in one of the

∗
Corresponding author.
E-mail address: [email protected] (G. Battineni).

https://doi.org/10.1016/j.jjimei.2020.100004
Received 9 November 2020; Received in revised form 5 December 2020; Accepted 5 December 2020
2667-0968/© 2020 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/)
A. Aggarwal, M. Mittal and G. Battineni International Journal of Information Management Data Insights 1 (2021) 100004

Fig. 1. Image generation /learning with the

help of 3D GAN modelling (Wu et al., 2016).

Fig. 2. Block diagram of the Generative Adversarial Network (GAN).

The GAN working based on three principles, firstly to make the gen- The implementation of the 3D patterns of the image with GAN fol-
erative model learn, and the data can be generated employing some lows the initiation of the random generator and discriminator and the
probabilistic representation. Secondly, the training of a model is done generator helps to understand the image type. The 2D image labelling
can be done in any conflicting situation. Lastly by using the deep learn- has been trained by discriminator with a label as y = 1 and the 3D im-
ing neural networks and using the artificial intelligence algorithms for ages produced in return labelled as y = 0. Thereafter, the discriminator
training the complete system (Liu & Tuzel, 2016). The basic idea of checks the image weights by the discriminator and quantity them to the
GAN network deployment is for unsupervised ML techniques but also generator. This quantification network with images has been produced
proved to be better solutions for semi-supervised and reinforcement by label enforcement as y = 1, this process repeats till complete extrac-
learning. These factors all together enable GAN networks as compre- tion of 3D image features.
hensive solutions in many fields such as healthcare, mechanics, banking, As mentioned, the functionality of GAN is based on similar principles
etc. of neural networks as a training set has given as input to learning gener-
GAN is an analogous type of idea generated to model animal be- ate novel data that similar to the training set. Especially, the image data
haviour by researchers around 2013 (Bryant, 2013). It is a relative in- training by GAN can result in new images that are similar characteristics
novation in the field of deep learning that uses two different networks of human behaviour.
one that generates images. For instance, during fake image classifica- The step by step functionality of GAN has been explained as follows
tion, one network called a generator creates fake images after an image
by another network called a discriminator (Hsu, Zhuang & Lee, 2020). • The users have produced using a generator by the discriminative
These networks are a category of deep learning models in particular network from the true data distribution.
convolutional neural network (CNN) frameworks. If at any time the dis- • The system has trained so that the liability rate of the network can
criminator is not able to notify the distinction between the two gener- be increased and the discriminator network can be fooled by pro-
ate images and actual images representation is considered as converged. ducing such candidates that are not synthesized i.e. still part of data
The training set trains to learn to produce novel information similar to distribution.
the training set. Images generated from GAN are also the same images • A dataset acts as initial training data for the discriminator.
that give the impression of the seemingly genuine to the individual ob- • For training samples datasets are presented till accuracy is achieved.
server which may have real features (Marra, Gragnaniello, Cozzolino • The generator is trained to produce candidates when the discrimi-
& Verdoliva, 2018). GAN can work on the unsupervised, supervised as nator is fooled when it is fed random input it processes them.
well as for reinforcement. This generative network produces the image • Lastly, backpropagation has been applied to generators as well as
candidate and the discriminator used for evaluation. Fig. 2 is the block discriminator where the former produces better images and the latter
diagram representation of GAN. is skilled at fading artificial images.

2
A. Aggarwal, M. Mittal and G. Battineni International Journal of Information Management Data Insights 1 (2021) 100004

• A deconvolutional neural network is a generative network and CNN 1) Does the selected paper present the application-based review GAN
acts as a discriminator. works?
• Sometimes GANs deal with mode collapse when the network fails to 2) Does the publication is discussing GAN progress in real-time indus-
generalize in case missing entire modes from input data. tries?
• Many solutions for one problem are proposed by the researchers. 3) Does the paper accurately address the research question and objec-
tives?
In this work, the authors have presented a review analysis of the
GAN functionality and its applications in real-time industries. Adver- The exclusion criteria were a) records of different characteristics
sarial principle approaches with deep learning to produce generative managed with inclusion criteria, b) language, and c) studies with no
models and simulation of other network theories have also been dis- precise GAN objectives. The objective here is to understand how these
cussed. Besides, possible future developments in GAN models have also models are evaluated, which industries are connected to GAN so that
been explored. sophisticated applications can be spoken for the beneficiary of society.
The further sections of the presenting paper are as follows: Section 2,
includes the methods conducting in the literature review and study se- 3. Results
lection process. In Section 3, the results of the review analysis including
key findings have been discussed. In Section 4, the authors have pre- In this section, the authors have presented the stages involved in the
sented the primary observations that evaluate the importance of GAN literature search and the evolution of GAN models in different applica-
models, and finally, the conclusion section summarizes the findings and tions. Broad areas of GAN applications are increasing quickly with time
future works. and its functionality including some of the major applications in each
thematic view has been discussed below.
2. Methods
3.1. Preliminary records outcome
2.1. Search strategy
Following the initial screening of the abstracts, 1783 articles that
Authors have adopted the systematic literature review approaches were not appropriate to the goals of this investigation were wiped out
that aligned with previous studies ((Agarwal, Chauhan, Kar & Goyal, for the accompanying reasons: 1757 articles were identified as dupli-
2017; Singh, Grover, Kar & Ilavarasan, 2020)). Three databases PubMed, cates, and 26 articles are not in English. In the second phase of screen-
EMBASE, and Web of Science (WoS) have been involved to extract the ing, the remaining 301 papers distributed equally to authors for inde-
relevant works. Search strategies have been developed to identify the pendent evaluation, and among only 61 articles are successfully satisfy
key literature amongst GAN applications and functionality. The possible the above-mentioned research objectives. Of these, following the pe-
synonyms, alternative words, and substitutes for the key terminologies rusing of the full-text form, nine were rejected because they did not
and Boolean operators like “GAN models”, OR “GAN applications” OR examine the relationship between GAN modelling and image diagnosis
“GAN in image processing”, OR “GAN AND cognitive computing’ OR which didn’t fall inside the recently indicated consideration standards.
‘face detection’ OR ‘3D classification with GAN’ OR ‘GAN in healthcare’ Ultimately, 52 papers are considered for further analysis. The literature
AND ‘deep learning transformation with GAN’ have been included. process model consists of screening steps that are further has presented
The authors worked the primer appraisal freely by perusing the ab- in Fig. 4.
stracts and drafting a rundown of the articles that they thought about
qualified. At that point, the previously mentioned search words were 3.2. Survey on GAN applications
looked at for disparities, and when one was distinguishing from others
of a similar kind, the various thoughts were talked about before reaching This section presents the explanation of the involvement of gener-
the final selection. When the authors chose which articles were qualified ative adversarial networks in major domains and Table 1 presents the
to be remembered for the survey, they read all the articles to assem- overview of GAN studies involved in different domains.
ble information helpful with the end goal of the exploration exertion
(Grover, Kar & Vigneswara Ilavarasan, 2018). The data assessment was 3.2.1. GAN in 3D object generation
directed freely and all opinions were compared to define an agreement. Investigating powerful 3D image generation techniques is a basic
The review analysis of research articles included in the previous viewpoint in the area of computer vision and computer graphics. To
five years (2016–2020) and in all applications where GAN has been in- identify 3D objects and enhance computer vision, some studies were em-
volved. The following applications have emerged: 3D object generation, ployed GAN networks. For instance, Yu et al. proposed a network that
medicine, pandemics, image processing, face detection, texture transfer, processes unclear data with no labelling, and the idea of 3D point en-
and traffic controlling. Prior Studies are considered to be less relevant coder cloud GAN Point encoder has been used in painting and uses max-
as their practical applications increasingly starting from 2016. pooling layer to resolve points for the learning process. Two networks
are worked as input encoder and decoder pipelines which results in a
2.2. Selection criteria better characteristic representation of the input point cloud (Yu, Huang,
Li, Zhang & Le, 2020).
The GAN literature search outcome provided 2084 records that are An architecture based on 3D-CNN lightweight multi-level architec-
categorised by publications associated with the objective of this study. ture connected super-resolution network and if generative adversarial
In particular, 1141 records in Scopus (Embase), 537 in WoS, and 406 in network-based training is provided to the network it creates sharp im-
PubMed were found. The yearly availability and library search distribu- ages with better quality (Chen et al., 2018). A GAN focusing enhance-
tion has presented in Fig. 3. ment of 2D monochromatic images in the creation of realistic 3D imag-
The following three types of articles have been considered: Original ing (Ye, Zhang, Ding, Li & Zhu, 2020). A generative 3D model which
papers, reviews, and analytical studies. The inclusion criteria involved is a group of people wearing clothes and doing 3D scans with different
during the selection process a) study type: original research, preliminary pose and outfits and is trained with a conditional Mesh-VAE-GAN so
analysis, literature works published after only 2016; b) articles with a that deformation of clothing can be learned from the SMPL body model
major focus on GAN applications and the latest advancements of GAN so that analysis of human motions and poses can be taken (Ma et al.,
models. More specifically, the records were screened for the three crite- 2020). Generalization for complex dressed people in common images
ria in the following sequence before the inclusion of final analysis: and videos is not done but learned by minimizing clothes 3D scans. A

3
A. Aggarwal, M. Mittal and G. Battineni International Journal of Information Management Data Insights 1 (2021) 100004

Fig. 3. Library Search outcomes: Library distribution (left) and Yearly distribution (right).

Fig. 4. Steps involved in records screening.

4
A. Aggarwal, M. Mittal and G. Battineni International Journal of Information Management Data Insights 1 (2021) 100004

Table 1
Key studies that define different GAN applications.

Type Authors [Ref] Year Model Application

3 D object Yu Y.et al. (Yu et al., 2020) 2020 GAN Point encoder Processes unstructured data with no
generation labelling
Y Chen et al. (Chen et al., 2018) 2018 3D-CNN Create sharp images of good quality
G Ye et al. (Ye et al., 2020) 2020 Deep learning-based GAN Improving 2D monochromatic images
Q Ma et al. (Ma et al., 2020) 2020 Generative 3D model Human motion capturing
Y Jin et al. (Jin et al., 2020) 2020 GAN model with three-tier adversarial Production of high-quality 3D objects
principle
Medicine S Baek et al. (Baek et al., 2020) 2020 GAN and Mesh Model Production of MR Images in sealed
pixels
Jain D K et al. (Jain et al., 2020) 2020 GAN poser Detection of human motion
A Teramoto et al. (Teramoto et al., 2020) 2020 Deep convolutional neural network Classify cytological images
(DCCN) with GAN
M D Cirillo et al. (Cirillo et al., 2020) 2020 Vox2Vox: 3D-GAN Brain tumour segmentation
H C Shin et al. (Shin et al., 2018) 2018 Conventional GAN Identify medical images
J. Islam et al. (Islam & Zhang, 2020) 2020 Conventional GAN Brain image generation
H Lan et al. (Lan & Toga, 2020) 2020 SC-GAN NeuroImage synthesis
G Zhaoa (Zhaoa, 2020) 2020 Bayesian Conditional GAN MRI Brain Image Synthesis
R Oulbacha et al. (Oulbacha & Kadoury, 2020) 2020 Pseudo-3D Cycle GAN MRI to CT Synthesis of the Lumbar
Spine
X Zhang et al. (X. Zhang et al., 2020) 2020 Deform-GAN Noise reduction in 3D medical images
D Yang et al. (Yang et al., 2019) 2019 Adversarial image-to-image networks Medical image synthesis and semantic
segmentation
Pandemics Loey M et al. (Loey et al., 2020) 2020 GAN and deep transfer learning COVID-19 detection with chest images
S Albahli (Albahli, 2020) 2020 GAN with the deep neural network Diagnose coronavirus disease
model pneumonia
Image C Li et al. (Li & Wand, 2016) 2016 Markovian GAN Generate 3D image from 2D image
processing H Zhou et al. (Zhou et al., 2020) 2020 Dual GAN Recovering of high-resolution images
T Go et al. (Go et al., 2020) 2020 Deep neural network-based GAN Perform image transformation
S Zhang et al. (S. Zhang et al., 2020) 2020 Conventional GAN Image denoising
H Tang et al. (Tang et al., 2020) 2020 Conventional GAN Semantic guided scene generation
Face detection F Mokhayeri et al. (Mokhayeri et al., 2020) 2020 A new Controllable GAN (C-GAN Cross-domain face synthesis
J Zhao et al. (Zhao et al., 2019) 2019 Dual-Agent Generative Adversarial Unconstrained Face Recognition
Network (DA-GAN)
M Kowalski et al. (Kowalski et al., 2020) 2020 Deep learning-based GAN Face Image Generation
D P Jaiswal et al. (Jaiswal et al., 2020) 2020 Conventional GAN Face animation
Text L Sixt et al. (Sixt et al., 2019) 2019 Conventional GAN Generating realistic labelled data
transferring R Spick et al. (Spick et al., 2020) 2020 3D-GAN Generate high-quality texture by
adding colour
Traffic control D Xu et al. (Xu et al., 2020) 2020 GE-GAN Road traffic estimation
Fathi-Kazerooni S et al. (Beery et al., 2020) 2020 GAN Tunnel Detection of traffic images

3D GAN developed for the creation of 3D objects probabilistic space the best output can be achieved when generator loss is weighted 5
in volume convolutional networks and generative adversarial networks. times compared to the discriminator loss (Cirillo, Abramian & Eklund,
The model uses a three-tier adversarial principle in place of heuristic 2020). A clinical application by adopting GAN that helps in the unsuper-
enable the generator to detain object structure absolutely and produce vised image to image conversation and identification of medical images
high-quality 3D objects along with generator maps from a small dimen- (Shin et al., 2018). It is also proven that artificial medical metaphors
sional to 3D objects (Jin, Zhang, Li, Tian & Zhu, 2020). these networks classify brain PET images for the identification of three-
stage (i.e., normal, mild, severe) Alzheimer’s disease (AD) (Islam &
Zhang, 2020). It is stated that any system is designed on 3D conditional
3.2.2. GAN in medicine
GAN and uses a normalization of spectral stabilizes criteria of feature
Machine learning and deep learning concepts are highly involv-
matching for achieving convergence optimization. An autonomous con-
ing in the medical domain especially to identify chronic diseases
ditional GAN considerably surpasses traditional 2D conditional GAN and
(Battineni, Sagaro, Chinatalapudi & Amenta, 2020), (Kaur et al., 2018).
3D functioning so that a dynamic 3D deep learning-based neuroimaging
It is reported that GAN can produce the MR Images successfully in
synthesis can be done (Lan & Toga, 2020).
sealed pixels (Baek, Kim & Kim, 2020). Jain et al. proposed a novel
An enhanced deep Super-Resolution Generative Adversarial Network
technique based on a generator discriminator called GAN-Poser for the
which creates images for three diverse stages of brain normal control,
prediction of human motion with 3D input of human skeleton image
mild cognitive impairment, and disease are image stages of Alzheimer’s
(Jain, Zareapoor, Jain, Kathuria & Bachhety, 2020). A bidirectional GAN
(Islam & Zhang, 2020). Likewise, a Bayesian conditional GAN with un-
framework iterative prediction scheme so that form crumple can be
necessary feature dropouts to get better image mixture exactness. Un-
predicted to normalize the training rather than using traditional Eu-
certainty in feature interpretation has been generated by the Bayesian
clidean loss. The model gives better performance over conventional deep
network done on the samples of brain tumour datasets of more than
learning methods evaluated on the datasets called NTU-RGB-D and Hu-
100 examples when compared to traditional Bayesian neural networks
man3.6 M. Besides, tumour classification by MR images also possible by
(Zhaoa, 2020).
applying the deep learning associated GAN’s models, and it results in a
An unsupervised approach Pseudo-3D Cycle GAN architecture in
comprehensive classification of glioblastoma patients (Teramoto et al.,
which neighbouring parts in mixture along with cyclic loss function en-
2020).
suring consistency from the fusion of CT images of the lumbar spine for
Brain tumour images showing that 3D volume using GAN for seg-
guided surgical images from T2 weighted MRI acquired for diagnostics.
mentation called Vox2Vox works on multi-channel 3D MR images and

5
A. Aggarwal, M. Mittal and G. Battineni International Journal of Information Management Data Insights 1 (2021) 100004

This approach makes use of a trainable pre-processing pipeline using separately constructs and learns sub generators concentrating on the
low capacity fully convolutional networks to normalize MRI data to cas- generation of different screens (Tang, Xu, Yan, Torr & Sebe, 2020).
cade for the segmentation of vertebral bodies and pedicles (Oulbacha &
Kadoury, 2020). 3.2.5. Face detection applications by GAN
The alignment of the sequences and modalities when there is a reg- Fake face identification is crucial for intelligent frameworks since
istration of the 3D medical images by use of unsupervised learning and generative models becoming famous by day-to-day. Because of enhance-
gradient descent. This approach provides a decrement in the noise and ment in quality in the fake face, the prepared models become increas-
blurriness in 3D medical images by the GAN model (Zhang, Jian, Chen ingly not efficient to identify the fake faces and corresponding training
& Yang, 2020). Yang et al. (2020) adopted a method of deep learn- data has been mentioned as outdated. In such cases, if the performance
ing network has created as a baseline for medical image segmentation of faces has to be recognized in the video surveillance system, and can
using GAN with machine learning, computer vision, and medical im- be improved by imitating face generation. Therefore, scholars proposed
age analysis. It is also able to get a better segmentation presentation a cross-area face imitation combined with the GAN named as controlled
with global shape constraints and applied to domain adaptation for GAN (CGAN) (Mokhayeri, Kamali & Granger, 2020). In CGAN refine-
both image synthesis and semantic segmentation (Yang, Xiong, Xu & ment, model simulation has been conducted by the face images for noise
Kevin Zhou, 2019). Medical images and ML and deep learning create a removal.
cochlea signed distance map depending on four input parameters that A Dual-Attention Generative Adversarial Network (DA-GAN) in
leads to a 60-fold improvement in the time of computation compared which a photo-realistic face frontal by capturing both contextual depen-
to more classical SDM generation methods. The complexity of SDM for dency and local consistency during GAN training for highlighting the
3D parametric shapes is overcome by SDM generation neural network is required pose and illumination discrepancy in the image (Zhao et al.,
implemented with dimensional cochlea shape model parameterized by 2019). Also, Kowalski et al. proposed a model called CONFIG-Net which
4 shape parameters (Zhang et al., 2018). is an attribute detection network. The neural face model permits the
dominant person image trained on the real face and fake face detec-
3.2.3. Pandemics tion by synthetic data, and it is a separating aspect such as pose, facial
In the pandemics like novel coronavirus (COVD-19) (Chawla, Mit- expression, hairstyle, etc. (Kowalski et al., 2020). The automatic gener-
tal, Chawla & Goyal, 2020), deep learning models are highly involved ation of facial images using GAN in solving the problem of facial images
to mitigate the virus severity in the form of bots (Battineni, Chintalapudi of animated works (Jaiswal, Kumar & Badr, 2020).
& Amenta, 2020). These models coupled to GAN can be used to iden-
3.2.6. Texture transferring
tify the infection severity. The studies like Loey et al. have proved that
A GAN based texture interpretation in the need of computations, by
GAN couple with deep learning models is the best approach for COVID-
that high-resolution texture sampling and shading in million times fold
19 detection by involving chest X-ray images. The authors collected set
can be performed. Recently, deep learning has been used by researchers
307 chest X-ray images and identify high accuracy images that are help-
on content combinations and style representations from a different type
ing to detect novel coronavirus (Loey, Smarandache & Khalifa, 2020).
of image analysis. The content and style extraction has called texture
It is reported that a model called COVID-GAN that is associated with
transferring. In Sixt, Wild and Landgraf (2019), the authors trained GAN
synthetic chest X-ray images has successfully done image classification
with 3D extracted image with no texturing results. The output image is
with a combination of shaping the synthetic images and achieved 89%
another 3D rendered image with texture. Finally, the generative deep
accuracy (Albahli, 2020).
learning model produces an output image trying to match the target
image. Given this, Spick et al. proposed a model initial voxel-based 3D
3.2.4. Image processing by GAN
GAN learning model that includes colour to produce generated sam-
GAN models can solve the ultrasound image resolution by inte-
ples by adapting channels of voxel inputs. If unsupervised learning is
grating deep learning models. These are employed to end encoding
used to generate high-quality texture then there is an improvement in
and decoding for making high-resolution images from regularly capture
turnaround time and these are tested on a thin collection of inputs from
prostate ultrasound images (Van Sloun et al., 2019). GAN can produce
a set of open access textured models (Spick, Demediuk & Alfred Walker,
different lesion classes from a small sample size of each lesion, and af-
2020).
ter application of deep convolutional GAN to generate a 3D image from
2D image resulting in faster analysis of images (Li & Wand, 2016). For 3.2.7. Traffic control
that optical clearing of images had been involved for the high-resolution Different investigations are been highlighted that precise road traffic
image volumes that match low-resolution volume images (Zhou et al., details by detectors are more accurate than different model predictions.
2020). Xu et al. developed a traffic road estimation framework using deep learn-
Holography defines the creation of a unique photography image with ing called GE-GAN where dual street systems of two cities are been used
no involvement of lens. Some studies applied GAN identification of holo- as a case study. This is done by using the data from neighbouring links
grams. Digital holography microscopic which records hologram con- to guess the road traffic of the states by application of graphs for the il-
tains 3D data using neural network and GAN had applied (Go, Lee, You lustration of the road network and using GAN’s (Xu, Wei, Peng, Xuan &
& Lee, 2020). Image to image conversion is not only focused on a sim- Guo, 2020). The traffic images can detect and classify the traffic occur-
ple white light source to holographic picture by measuring the network rences, this helps to control self-driving cars on roads (Fathi-Kazerooni
comparison of generated and true holograms of microspheres in a 3D & Rojas-Cessa, 2020).
image.
A process of achieving clean images from a hard threshold with 4. Discussion
wiener filtering for solving stained images. These images are disinte-
grating to get coefficients of a 3D block-matching algorithm to get clean In this study, the authors have presented GAN architecture and func-
images by training latent clean images through GAN. For that peak tionality along with its major domain applications. A survey of im-
signal-to-noise ratio, structural resemblance, and edge preserve index age segmentation by deep learning approaches like GAN can produce
are taken as criteria for noise removal in visual effects (Zhang et al., rigorous literature review and studying the range of works on seg-
2020). Besides, Tang et al. addressed that the image generation of small mentation of semantic and illustration level, networks covering recur-
objects and images of local texture based on the guided scene is quite rent networks, encoder-decoder architectures, convolutional pixel la-
difficult. For that, a scene is generated with local background and a lo- belling networks, and visual attention generative with adversarial set-
cal class with semantic maps has been designed for the guidance which tings. The advantages, challenges, strengths, similarity of other models,

6
A. Aggarwal, M. Mittal and G. Battineni International Journal of Information Management Data Insights 1 (2021) 100004

their datasets, and prospects in the above area can be well explained inal level of details and colours. The other wide range of GAN applica-
(Sultana, Sufian & Dutta, 2020). Because of recent advancements in deep tions including Speech to image construction, visualize climate changes,
learning and the presentation of huge 3D CAD datasets like Shape Net face ageing, photo blending, motion video capturing, video prediction,
(Chang et al., 2020), there have been some motivating works in deep etc.
learning object notation. Unique about part-based techniques, a large
number of these generative methodologies don’t expressly show the idea 4.1. GAN limitations
of parts or recover them from an object repository; as an alternative,
they arrange fresh objects based on learned object presentation. However, GAN architecture has some limitations. The images cre-
A pro-signed distance map approach generates a cochlea signed dis- ated by GAN look misleadingly like a photograph of a real person based
tance map depending on four input parameters and demonstrated with on the analysis of portraits. Different concern by the people has been
help of deep learning resulted in sixty times enhancement as compared raised for using the human image synthesis by GAN potentially by frauds
to traditional generation methods (Wang et al., 2020). This is a difficult thereby producing the fake and photographs and videos without per-
issue because compared with the space of 2D images, it is harder to show mission. On social media, fake profiles can be prevented using GANs for
the space of 3D shapes due to its higher dimensionality. Their present generating the unique or pragmatic pictures of persons that do not exist.
outcomes are empowering, however frequently there still exist gaps in DARPA’s Media Forensics programs help in countering such fake media
the created objects. To overcome those, generative adversarial networks profiles produced using GANs and along with that many laws are passed
have been introduced in this study. The common datasets for generation and they are implemented by the year 2020.
of 3D objects using GAN are 2D-to-3D deformable sketches (Zorah Läh-
ner, Rodolà, Schmidt & Bronstein, 2020), 3D deformable objects in clut- 4.2. Future works
ter (Cosmo, Rodola, Masci, Torsello & Bronstein, 2016), ANN_SIFT1M
(sift 1 M dataset & ANN - Frankie Yan’s Blog, 2020), CIFAR-10 (CIFAR- GANs are representing a new concept in deep learning with the
10 & CIFAR-100 datasets, 2020), and CLEF-IP 2011 evaluation on patent fast pace continuation of the AI research society and bringing about
images. The data created by GANs are casual vectors under the cate- many ongoing publications pushing the technologies beyond its pri-
gory of concentrated casual vectors and behave as Gaussian mixtures mary limits. The deficiency of the GAN essential hypothesis is an ob-
by using deep learning algorithms (El, Seddik, Louart, Tamaazousti & struction for GAN models to develop high-quality generative models.
Couillet, 2020). Accordingly, the most significant implementation for future works is
The library-like Py-Torch is the most famous GANs packages that to have breakthroughs in hypothetical aspects to tackle issues, for
can implement and provide a comprehensive approach of GAN training example, difficulties in training, non-union, and model breakdown
with different image datasets. This package helps to control the issues (Salimans et al., 2016). Despite some generally improved strategies, for
during GAN model implementation using different frameworks and at example, weights pruning and regularization (Arjovsky, Chintala & Bot-
the same time evaluating when the same metric has been used (Lee & tou, 2017; Gulrajani, Ahmed, Arjovsky, Dumoulin & Courville, 2017),
Town, 2020). A deep learning fault detection process based on unbal- Nash Equilibrium (Kodali, Abernethy, Hays & Kira, 2017), and new loss
anced data with global optimization GAN leads to high misclassification. functions (Mao et al., 2017), future improvements are still in need. Be-
Such a method of novel generator and discriminator are planned to using sides, GAN can address the new theories and research outcomes in ML
an auto encode reaching to global optimization and refine unqualified models, for example, attention mechanisms can be incorporated for cap-
produced samples from qualified samples for error analysis (Zhou, Yang, turing global features. GAN research with policy gradient procedures
Fujita, Chen & Wen, 2020). in reinforcement learning can overcome the weakness in dealing with
In other applications like space sciences, the GAN network to build discrete variables, therefore GAN can work in different conditions to
better astronomical images to predict as well as simulate gravitational increase the scope of its application (Kurakin, Goodfellow & Bengio,
sensing for dark matter research to model distribution in any direction 2019).
in space (Mahdizadehaghdam, Panahi & Krim, 2019). Scholars men-
tioned that radar-based map deals with the difficulty of signal loss and 5. Conclusions
a map has been created to interpret the climatic changes and light-
ing compatible with sensor nature, and the promise of vehicle localiza- GANs established in a way to the comprehensive domain of inde-
tion has done when a picture from FMCW radar is placed on a land pendent data expansion and solve problems that require a generative
vehicle (Cornick, Koechling, Stanley & Zhang, 2016). Besides, GANs solution like the image to image transformation. In this work, many ap-
projected a better method of modelling high energy jet formation, ap- plications of GAN have been analysed and after going through in-depth
proximate hurdles in costly simulations of particle physics experiments revision of GAN and deep learning and its applications in preceding
((de Oliveira, Paganini & Nachman, 2017; Lin, Bhimji & Nachman, years it can be seen as there are many cutting edge learning models lay
2019)). To classify the images using GAN the discriminator is changed the category of supervised, unsupervised, and reinforcement learning.
to predict the label of the class of any image despite accepting as input. Furthermore, many deep learning datasets and frameworks are used to
To stabilize the training and generation of large excellence images can present the performance of deep learning problems.
be allowed. In arts like fashion, Fs-GANs have been implemented for
imaginary pictures of the models without hiring a photographer hiring, Declaration of Competing Interest
makeup artists, also cut down the studio costs (Singh, Bajpai, Vijayara-
jan & Prasath, 2019). Fashion advertising companies using GANs having The authors had no conflicts during the publication.
various groups of models increasing people who resemble models. Land-
scapes, portraits, album covers can be created using GAN. Funding
GAN networks can also be used to creating games by a technique
of scaling the 2D texture resolution of video games and recreating in This work is not receiving any means of funding.
larger resolutions (Rodriguez Torrado et al., 2020). Process of training
and after that down sampled so that it can be fitted in-game native res- Authors’ contributions
olution. The final outputs are similar to the super sampling method of
anti-aliasing. If such networks are trained properly can provide a clear as A.A. selected the studies, performed a literature review, and partici-
well as a sharper image with high magnitudes improves quality if com- pated in drafting the manuscript. M.M. guided the study’s overall stages,
pared with the original. The images were developed to retain the orig- revisited the manuscript critically for important intellectual content, and

7
A. Aggarwal, M. Mittal and G. Battineni International Journal of Information Management Data Insights 1 (2021) 100004

approved the study. G.B. designed the study, selected the studies, per- Kodali,.N., Abernethy,.J., Hays,.J., & Kira, en Z. (2017). “On convergence and stability of
formed literature review, methods, data extraction, analysed data, in- gans”, arXiv.
Kowalski, M., Garbin, S. J., Estellers, V., Baltrušaitis, T., Johnson, M., & Shotton, en
terpreted the result, and drafted the manuscript. All authors approved J. (2020). CONFIG: Controllable neural face image generation. Me. Toegang verkry:
the final article. Nov 02, 2020. [Online]. Available at: http://arxiv.org/abs/2005.02671 .
Kurakin,.A., Goodfellow, I.J., & Bengio, en S. (2019). “Adversarial examples in the physical
world”,
References Lan, H., & Toga, en A. W. (2020). SC-GAN: 3D self-attention conditional GAN with spectral
normalization 2 for multi-modal neuroimaging synthesis 3 4 authors 5 Haoyu Lan a,
Agarwal, N., Chauhan, S., Kar, A. K., & Goyal, en S. (2017). Role of human behaviour the Alzheimer disease neuroimaging initiative. bioRxiv.
attributes in mobile crowd sensing: A systematic literature review. Digital Policy, Reg- Lecun, Y., Bengio, Y., & Hinton, en G. (2015). Deep learning. Nature. 10.1038/na-
ulation and Governance. 10.1108/DPRG-05-2016-0023. ture14539.
Albahli, S. (2020). Efficient GAN-based chest radiographs (CXR) augmentation to di- Lee, K. S., & Town, en C. (2020). Mimicry: Towards the Reproducibility of
agnose coronavirus disease pneumonia. International Journal of Medical Sciences. GAN Research. Mei 2020. Toegang verkry. Nov 03[Online]. Available at:
10.7150/ijms.46684. http://arxiv.org/abs/2005.02494 .
Arjovsky,.M., Chintala,.S., & Bottou, en L. (2017). “Wasserstein GaN”, arXiv. Li,.C., & Wand, en M. (2016). “Precomputed real-time texture synthesis with markovian
Baek,.S., Kim, K.I., & Kim, en T.-K. (2020). “Weakly-supervised domain adaptation generative adversarial networks”, doi: 10.1007/978-3-319-46487-9_43
via GAN and mesh model for estimating 3D hand poses interacting objects”, doi: Lin, J., Bhimji, W., & Nachman, en B. (2019). Machine learning templates for QCD factor-
10.1109/cvpr42600.2020.00616 ization in the search for physics beyond the standard model. Journal of High Energy
Battineni, G., Chintalapudi, N., & Amenta, en F. (2020a). AI Chatbot design dur- Physics. 10.1007/JHEP05(2019)181.
ing an epidemic like the novel coronavirus. Healthcare. 10.3390/healthcare Liu, M.Y., & Tuzel, en O. (2016). “Coupled generative adversarial networks”,
8020154. Loey, M., Smarandache, F., & Khalifa, en N. E. M. (2020). Within the lack of chest COVID-
Battineni, G., Sagaro, G. G., Chinatalapudi, N., & Amenta, en F. (2020b). Applications 19 X-ray dataset: A novel detection model based on GAN and deep transfer learning.
of machine learning predictive models in the chronic disease diagnosis. Journal of Symmetry (Basel).. 10.3390/SYM12040651.
Personalized Medicine. 10.3390/jpm10020021. Ma,.Q. et al., (2020). “Learning to Dress 3D people in generative clothing”, doi:
Beery,.S. et al., (2020). “Synthetic examples improve generalization for rare classes” doi: 10.1109/cvpr42600.2020.00650
10.1109/WACV45572.2020.9093570 Mahdizadehaghdam,.S., Panahi,.A., & Krim, en H. (2019). “Sparse generative adversarial
Bryant, G. A. (2013). Animal signals and emotion in music: Coordinating affect across network”, doi: 10.1109/ICCVW.2019.00369
groups. Frontiers in Psychology. 10.3389/fpsyg.2013.00990. Mao,.X., Li,.Q., Xie,.H., Lau, R.Y.K., Wang,.Z., & Smolley, en S.P. (2017). “Least squares
Chang, A. X., et al. (2020). ShapeNet: An information-rich 3D model repository. Des generative adversarial networks”, doi: 10.1109/ICCV.2017.304
2015. Toegang Verkry. Nov 03[Online]. Available at http://arxiv.org/abs/1512. Marra,.F., Gragnaniello,.D., Cozzolino,.D., & Verdoliva, en L. (2018). “Detection of GAN-
03012. generated fake images over social networks”, doi: 10.1109/MIPR.2018.00084
Chawla, S., Mittal, M., Chawla, M., & Goyal, en L. (2020). Corona Virus - SARS-CoV-2: Mokhayeri,.F., Kamali,.K., & Granger, en E. (2020). “Cross-domain face synthesis using a
An insight to another way of natural disaster. EAI Endorsed Transactions on Pervasive controllable GAN”, doi: 10.1109/WACV45572.2020.9093275
Health and Technology. 10.4108/eai.28-5-2020.164823. Oulbacha,.R., & Kadoury, en S. (2020). “MRI to CT synthesis of the lumbar spine from a
Chen,.Y., Shi,.F., Christodoulou, A.G., Xie,.Y., Zhou,.Z., & Li, en D. (2018). “Efficient and pseudo-3D cycle GAN”, doi: 10.1109/ISBI45749.2020.9098421
accurate MRI super-resolution using a generative adversarial network and 3D multi- Rodriguez Torrado, R., Khalifa, A., Cerny Green, M., Justesen, N., Risi, S., & To-
level densely connected network”, doi: 10.1007/978-3-030-00928-1_11 gelius, en J. (2020). Bootstrapping conditional GANs for video game level gener-
“CIFAR-10 and CIFAR-100 datasets”. https://www.cs.toronto.edu/~kriz/cifar.html (toe- ation. In Proceedings of the IEEE Conference on Games (CoG) (pp. 41–48). Aubll.
gang verkry Nov 03, (2020)). 10.1109/CoG47356.2020.9231576.
Cirillo, M. D., Abramian, D., & Eklund, E. A. (2020). Vox2Vox: 3D-GAN for brain Salimans,.T., Goodfellow,.I., Zaremba,.W., Cheung,.V., Radford,.A., & Chen, en X. (2016).
tumour segmentation. Mrt 2020, Toegang verkry. Nov 02[Online]. Available at: “Improved techniques for training GANs”,
Http://arxiv.org/abs/2003.13653 . Shin, H.C. et al., (2018). “Medical image synthesis for data augmentation and anonymiza-
Cornick, M., Koechling, J., Stanley, B., & Zhang, en B. (2016). Localizing ground penetrat- tion using generative adversarial networks”, doi: 10.1007/978-3-030-00536-8_1
ing RADAR: A step toward robust autonomous ground vehicle localization. Journal of “sift 1M dataset and ANN - Frankie Yan’s Blog”. https://frankie-yanfeng.github.io/2019/
Field Robotics. 10.1002/rob.21605. 02/18/sift-1M-dataset-and-ANN-2019/ (toegang verkry Nov 03, (2020)).
Cosmo, L., Rodola, E., Masci, J., Torsello, A., & Bronstein, en M. M. (2016). Matching Singh, H., Grover, P., Kar, A. K., & Ilavarasan, en P. V. (2020). Review of performance
deformable objects in clutter. In Proceedings - 2016 4th international conference on 3D assessment frameworks of e-government projects. Transforming Government: People,
vision, 3DV 2016, De (pp. 1–10). bll. 10.1109/3DV.2016.10. Process and Policy. 10.1108/TG-02-2019-0011.
de Oliveira, L., Paganini, M., & Nachman, en B. (2017). Learning particle physics by exam- Singh, M., Bajpai, U., Vijayarajan, V., & Prasath, en S. (2019). Generation of fashionable
ple: location-aware generative adversarial networks for physics synthesis. Computer clothes using generative adversarial networks: A preliminary feasibility study. Inter-
and Software for Big Science. 10.1007/s41781-017-0004-6. national Journal of Clothing Science and Technology. 10.1108/IJCST-12-2018-0148.
El,.M., Seddik,.A., Louart,.C., Tamaazousti,.M., & Couillet, en R. (2020). “Random ma- Sixt,.L., Wild,.B., & Landgraf, en T. (2019). “Rendergan: Generating realistic labeled data”,
trix theory proves that deep learning representations of gan-data behave as Gaussian Spick,.R., Demediuk,.S., & Alfred Walker, en J. (2020). “Naive mesh-to-mesh coloured
mixtures”, model generation using 3D GANs”, doi: 10.1145/3373017.3373067
Fathi-Kazerooni, S., & Rojas-Cessa, en R. (2020). GAN tunnel: Network traffic steganog- Sultana, F., Sufian, A., & Dutta, en P. (2020). Evolution of image segmentation
raphy by using GANs to counter internet traffic classifiers. IEEE Access : Practical In- using deep convolutional neural network: a survey. Knowledge-Based System.
novations, Open Solutions. 10.1109/ACCESS.2020.3007577. 10.1016/j.knosys.2020.106062.
Go, T., Lee, S., You, D., & Lee, en S. J. (2020). Deep learning-based hologram generation Tang,.H., Xu,.D., Yan,.Y., Torr, P.H.S., & Sebe, en N. (2020). “Local class-specific and
using a white light source. Scientific Reports. 10.1038/s41598-020-65716-4. global image-level generative adversarial networks for semantic-guided scene gener-
Grnarova,.P., Zurich Kfir Levy, E.Y., Lucchi,.A., Zurich Nathanaël Perraudin, E.T.H., Good- ation”, doi: 10.1109/cvpr42600.2020.00789
fellow Thomas Hofmann, en I., Zurich Andreas Krause, E.T.H. et al. (2019). “A Domain Teramoto, A., et al. (2020). Deep learning approach to classification of lung cytological
agnostic measure for monitoring and evaluating GANs”, images: Two-step training using actual and synthesized images by progressive growing
Grover,.P., Kar, A.K., & Vigneswara Ilavarasan, en P. (2018). “Blockchain for businesses: of generative adversarial networks. Plos One. 10.1371/journal.pone.0229951.
A systematic literature review”, doi: 10.1007/978-3-030-02131-3_29 Van Sloun, R. J. G., Solomon, O., Bruce, M., Khaing, Z. Z., Eldar, Y. C., & Mischi, en
Gulrajani,.I., Ahmed,.F., Arjovsky,.M., Dumoulin,.V., & Courville, en A. (2017). “Improved M. (2019). Deep learning for super-resolution vascular ultrasound imaging. In Pro-
training of wasserstein GANs”, ceedings of the IEEE international conference on acoustics, speech and signal processing
He,.K., Zhang,.X., Ren,.S., & Sun, en J. (2016). “Deep residual learning for image recog- (ICASSP) (pp. 1055–1059). 10.1109/ICASSP.2019.8683813.
nition”, doi: 10.1109/CVPR.2016.90 Wang,.Z. et al., (2020). “A deep learning based fast signed distance map generation”,
Hsu, C. C., Zhuang, Y. X., & Lee, en C. Y. (2020). Deep fake image detection based on Wu, J., Zhang, C., Xue, T., Freeman, W. T., & Tenenbaum, en J. B. (2016). Learning a
pairwise learning. Applied Science. 10.3390/app10010370. probabilistic latent space of object shapes via 3D generative-adversarial modeling.
Islam, J., & Zhang, en Y. (2020). GAN-based synthetic brain PET image generation. Brain Advances in Neural Information Processing Systems, 82–90. OkToegang verkry: Nov 03,
Informatics. 10.1186/s40708-020-00104-2. 2020. [Online]. Available at http://arxiv.org/abs/1610.07584 .
Jain, D. K., Zareapoor, M., Jain, R., Kathuria, A., & Bachhety, en S. (2020). GAN-Poser: An Xu, D., Wei, C., Peng, P., Xuan, Q., & Guo, en H. (2020). GE-GAN: A novel deep learning
improvised bidirectional GAN model for human motion prediction. Neural Computing framework for road traffic state estimation. Transportation Research Part C: Emerging
and Applications. 10.1007/s00521-020-04941-4. Technologies. 10.1016/j.trc.2020.102635.
Jaiswal, D. P., Kumar, S., & Badr, en Y. (2020). Towards an artificial intelligence aided Yang, D., Xiong, T., Xu, D., & Kevin Zhou, en S. (2019). Segmentation using adversar-
design approach: application to anime faces with generative adversarial networks. ial image-to-image networks. In Handbook of medical image computing and computer
Procedia Computer Science. 10.1016/j.procs.2020.02.257. assisted intervention (pp. 165–182). Elsevier. bll.
Jin, Y., Zhang, J., Li, M., Tian, Y., & Zhu, en H. (2020). Towards the high-quality anime Ye, G., Zhang, Z., Ding, L., Li, Y., & Zhu, en Y. (2020a). GAN-based focusing-
characters generation with generative adversarial networks. Toegang Verkry. Nov enhancement method for monochromatic synthetic aperture imaging. IEEE Sensors
03[Online]. Available at http://make.girls.moe . Journal. 10.1109/JSEN.2020.2996656.
Kaur, B., Sharma, M., Mittal, M., Verma, A., Goyal, L. M., & Hemanth, en D. J. (2018). Ye, H., Liang, L., Li, G. Y., & Juang, en B. H. (2020b). Deep learning-based end-to-end
An improved salient object detection algorithm combining background and fore- wireless communication systems with conditional GANs as unknown channels. IEEE
ground connectivity for brain image analysis. Computers and Electrical Engineering. Transactions on Wireless Communications. 10.1109/TWC.2020.2970707.
10.1016/j.compeleceng.2018.08.018.

8
A. Aggarwal, M. Mittal and G. Battineni International Journal of Information Management Data Insights 1 (2021) 100004

Yu, Y., Huang, Z., Li, F., Zhang, H., & Le, en X. (2020). Point Encoder Zhaoa,.G. (2020)) “Bayesian conditional GAN for MRI brain Image synthesis | DeepAI”.
GAN: A deep learning model for 3D point cloud inpainting. Neurocomputing. https://deepai.org/publication/bayesian-conditional-gan-for-mri-brain-image-
10.1016/j.neucom.2019.12.032. synthesis (toegang verkry Nov 02.
Zhang, S., et al. (2020a). An Image Denoising Method Based on BM4D and GAN Zhou, F., Yang, S., Fujita, H., Chen, D., & Wen, en C. (2020a). Deep learning fault diagno-
in 3D Shearlet Domain. Mathematical Problems in Engineering. 10.1155/2020/ sis method based on global optimization GAN for unbalanced data. Knowledge-Based
1730321. System. 10.1016/j.knosys.2019.07.008.
Zhang, X., Jian, W., Chen, Y., & Yang, en S. (2020b). Deform-GAN:AN unsupervised Zhou, H., et al. (2020b). 3D high resolution generative deep-learning network for fluores-
learning model for deformable registration. Toegang verkry. [Online]. Available at: cence microscopy imaging. Optics Letters. 10.1364/ol.387486.
http://arxiv.org/abs/2002.11430 . Zorah Lähner, D.C., Rodolà, E., Schmidt, F. R., & Bronstein, M. M. “Efficient glob-
Zhang,.Y., Yin,.H., Huang,.Z., Du,.X., Yang,.G., & Lian, en D. (2018). “Discrete ally optimal 2d-to-3d deformable shape matching”. https://vision.in.tum.de/~
deep learning for fast content-aware recommendation”, doi: 10.1145/3159652. laehner/Elastic2D3D/ (toegang verkry Nov 03.
3159688
Zhao, J., Xiong, L., Li, J., Xing, J., Yan, S., & Feng, en J. (2019). 3D-aided dual-agent GANs
for unconstrained face recognition. IEEE Transactions on Pattern Analysis and Machine
Intelligence. 10.1109/TPAMI.2018.2858819.