See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/357037859

Image Compression Using Neural Networks: A Review

Article  in  International Journal of Online and Biomedical Engineering (iJOE) · December 2021

DOI: 10.3991/ijoe.v17i14.26059

CITATIONS READS

0 159

4 authors:

Haval Sadeeq Thamer Hassan Hameed

Duhok Polytechnic University University of Duhok
8 PUBLICATIONS   38 CITATIONS    2 PUBLICATIONS   0 CITATIONS   

SEE PROFILE SEE PROFILE

Abdo Abdi Ayman Nashwan

Duhok Polytechnic University Duhok Polytechnic University
4 PUBLICATIONS   12 CITATIONS    1 PUBLICATION   0 CITATIONS   

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Haval Sadeeq on 15 December 2021.

The user has requested enhancement of the downloaded file.

Paper—Image Compression Using Neural Networks: A Review

Image Compression Using Neural Networks: A Review

https://doi.org/10.3991/ijoe.v17i14.26059

Haval T. Sadeeq1(*), Thamer H. Hameed2, Abdo S. Abdi1, Ayman N. Abdulfatah1

1
Duhok Polytechnic University, Dahuk, Kurdistan Region of Iraq
2
University of Duhok, Dahuk, Kurdistan Region of Iraq
[email protected]

Abstract—Computer images consist of huge data and thus require more

memory space. The compressed image requires less memory space and less trans-
mission time. Imaging and video coding technology in recent years has evolved
steadily. However, the image data growth rate is far above the compression ratio
growth, considering image and video acquisition systems’ popularization. It is
generally accepted, in particular, that further improvement in coding efficiency
within the conventional hybrid coding system is increasingly challenging. A new
and exciting image compression solution is also offered by the deep convolution
neural network (CNN), which in recent years has resumed the neural network
and achieved significant success both in artificial intelligent fields and in signal
processing. In this paper we include a systematic, detailed and current analysis of
image compression techniques based on the neural network. Images are applied
to the evolution and growth of compression methods based on neural networks.
In particular, the end-to-end frames based on neural networks are reviewed,
revealing fascinating explorations of frameworks/standards for next-generation
image coding. The most important studies are highlighted and future trends even
envisaged in relation to image coding topics using neural networks.

Keywords—image compression, neural networks, artificial neural network,

data compression

1 Introduction

Compression only minimizes the needed number of bits in order symbolize a video
file or an image without dramatically affecting (original) input quality. One of the rele-
vant compression areas is compression of images. The main aim of compression of the
image is to reduce the amount of data required to symbolize an image. Even a standard
digital image in a single band typically needs a large number of bits to display or store.
The compression of images reduces the redundancies and irrelevances in the data. In
contrast with the original image there are fewer bits needed to symbolize a compressed
image. There are fewer bits needed to symbolize a compressed image in comparison
to the original picture. The space used to store and bandwidth to convey the image
is greatly reduced. To save the transmission time and bandwidth, we can save more
images and transmit them in less time. Either image compression technique involves
three main phases Mapper, Quantizer, and Encoder [1]. In the first step, Mapper

iJOE ‒ Vol. 17, No. 14, 2021 135

Paper—Image Compression Using Neural Networks: A Review

converts the image to the frequency domain with special domain information. In order
to map many to one theory Quantizer works. More values are quantified to fewer. The
number of bits required to symbolize the image is therefore reduced. Finally, the result
is coded during the encoding process to further improve the compression. To re-create
the original image, several procedures are mapped namely; decoding, dequantization,
and inverse mapping. This is called decompression [2], [3]. Compression and decom-
pression block diagram are shown in Figure 1.

Fig. 1. Compression and decompression block diagram

In digital images, three fundamental redundancies are common:

1.1 Coding redundancy

The same number of bits is used in the normal image for more likely symbols and
for lower probability symbols. We may boost this by allocating less bits to gray levels
that seem more likely and more bits to less likely grays. In redundancy, there is no cor-
relation between pixels [4].

1.2 Psycho-visual redundancy

There is material in digital images which is not relative to ordinary visual process-
ing. Psychovisually redundancy is the name of this unimportant knowledge. The reduc-
tion of psychovisual redundancy leads to a lack of quantitative knowledge. The content
of the image cannot be detected by the human eye [5].

1.3 Interpixel redundancy

Unrecognizing and utilizing data is known as redundancy of interpixel. If a value of

its neighboring (subsequent) pixels can be predicted, the table would be redundant with
interpixel values. The resolution of the picture influences the interpixel redundancies.

136 http://www.i-joe.org
Paper—Image Compression Using Neural Networks: A Review

Redundancy between pixels of statistical dependence, particularly between adjacent

pixels [6].

2 Image compression techniques

The compression techniques of the image [7] are usually known as a Lossy and a
Lossless. Table 1 below presents different strategies for each group. The data would
be lost if there is a loss compression, as its name implies. The image restored isn’t the
same as the image of the input. Figure 2 portrays the Lossy compression block draw-
ing. Although the restored image is a complete replication of the original image with
lossless techniques. No data loss occurs; decompressed image is identical to uncom-
pressed one. Lossless methods of compression typically minimize file size for complex
images by about ten percent. Health photographs or pictures in the courts are used.
Lossless compression methods for simple images can provide significant compression.
The block diagram is seen in the Figure 3.

Table 1. Classification of compression techniques

Lossy Techniques Loseless Techniques
Transform coding Run-length encoding
Vector quantization Entropy coding
Fractal coding Preditiction coding
Block transform coding Huffman coding

Fig. 2. Lossy image compression

iJOE ‒ Vol. 17, No. 14, 2021 137

Paper—Image Compression Using Neural Networks: A Review

Fig. 3. Lossless image compression

2.1 Lossy techniques

Transform encoding. It is a sort of compression of relevant data like image or audio

signals. It translates pixel image values from spatial domain to frequency domain. More
information is given on the image with a few coefficients and most are low or zero. In
order to map the pixel value of the image to a series of coefficients, a reversible linear
transformation, such as Discrete Fourier or Discrete Cosine Transform is used. The
coefficients are quantified and coded further. As much knowledge as possible will be
gathered in a high-quality transformation in a limited number of coefficients. Then the
quantization technique is used to extract coefficients that contain the least quantity of
information. In this method, an input image of size N to N is first divided into a number
of n to n that is not overlapping, then transformed into arrays (N/n)2, each sizes array n
to n. Each block is separately transformed [8].
Vector quantization. Quantization operates on map many to one base. It maps
a value group to one. In addition, the methods of quantization are divided into two
types: scalar and vector quantization. scalar conducts a single mapping set for each
value. Vector quantization is the multidimensional creation of scalar quantization. In
both spectral and spatial domains, VQ can be implemented. The theory of data tells
us that vector quantization can achieve better compression than scalar quantization.
This technique constructs a dictionary of vectors known as vectors of code. The input
image is divided into unrelated, image-vector-marked blocks. The index of the closest
matching vector in the dictionary encodes any image vector [9].
Fractal coding. It assumes that a portion of an image is generally identical to
its neighboring parts of the input image. These sections are called “fractal codes”;
they are translated into mathematical data. The encoded image is restored by using
these fractal codes. By another way, metaheuristic algorithms such as [10] [11] [12]
represent a set of approaches used to solve hard optimization tasks with lucid resources
consumption. They are characterized by their fast convergence and reduction of

138 http://www.i-joe.org
Paper—Image Compression Using Neural Networks: A Review

research complexity. It is worth noting that [13] and [14] have used metaheuristics for
fractal image compression, and the performed experiments showed their effectiveness
in the resolution of such problems.
Block transform coding (BTC). In BTC, the image input is divided in blocks
where the size 8 pixels to 8 are used in every row. The coding of the block transform
benefits from the correlation of the block pixels. Each block is modified by virtue of
this. Finally, every block is individually quantified and coded [15].

2.2 Lossless techniques

Run-length encoding (RLE). Data is replaced in the RLE with a duo of (length,
value), where the number of repetitions is called ‘length’ and the repeated value is
referred to as ‘value’ This method is particularly used to compress large-scale images
because in common, gray-size images, long-term value is not common. It is intended
to broken down the gray image into the bit planes, and individual bit planes are
individually compressed. One of the variants of run time coding is an efficient run-
length coding mechanism [16].
Entropy encoding. Entropy is the lowest number of bits required to represent a
symbol in the compression frame (the total length of the code for all symbols). The
explicit features of the medium are not based. Every individual symbol in the input
is assigned a unique prefix-free code. A prefix-free codeword for the variable length
is used to replace each fixed-length input symbol. The codeword length is almost
proportional to the probability logarithm. The most widely used symbols therefore take
shorter codes [17].
Predictive coding. This technique is based primarily on the difference between the
initial and expected values. It’s often referred to as Differential Pulse Code Modulation
(DPCM). It is not necessary to decompose the input image into a group of bit planes.
Only new information is coded in each pixel after extracting the interpixel similarities.
The variations between the original and expected value of the pixel are measured as
new data. The system consists of a matching predictor in two phases: an encoder and a
decoder. The predictor will generate the likely pixel value depending on the number of
preceding inputs. The predicted value is coded to produce the next element of the data
stream with a variable code length [18].
Huffman coding. In 1952, D. Huffman created the code of Huffman. It is a code
of minimal duration. This means that the Huffman algorithm generates code as close
as possible to the lower bound, entropy. The technique results in an unprecedented
(or variable) code in which terms can vary in size [19] [20].

2.3 Assessment of compressed images

The compression ratio is the compression calculation of the image. Different factors
including mean squared error and peak noise signal are determined by the compression
efficiency.
There are also several other techniques; PSNR and MSE are mostly used because
they are simple to measure [21] [22].

iJOE ‒ Vol. 17, No. 14, 2021 139

Paper—Image Compression Using Neural Networks: A Review

Mean square error (MSE). This is calculated by the mean difference between
compressed and input (original) pixel square intensities. The MSE is indicated by the
following equation.
m −1 n −1
1
MSE =
m*n ∑ ∑ [ g ′( p, q) − g ( p, q)]
i =0 j =0
2
(1)

In these instances, the pixels of gʹ(p, q) and g(p, q) respectively are the values of the
restored images. And m & n are the spatial domain number of rows and columns [23].
Peak signal to noise ratio (PSNR). PSNR is the typical way in which fidelity
measurement is carried out [24] [25]. The term PSNR is the relation between the
highest signal value of deforming noise, which changes its representation value and the
signal value of the deforming noise. In terms of logarithmic decibel scale the PSNR is
usually represented. In decibels, PSNR is computed (dB). In general, a large difference
is said to be 0.5–1 dB. The PSNR’s mathematical representation is:

 m*n 
PSNR = 10log10  (2)
 MSE 

where MSE is Mean Square Error, and m & n are image rows & columns in spatial
form.
Compression ratio (CR). Is a further (extra) measurement metric for compression
measurement. It is the relative sum between the bits needed by the image input to
the bits required by the compressed image. The following, is the compression ratio
equation [26].
Uncompressed (original) image size
CR = (3)
Compressed image size


2.4 Image formats

Images are divided into various categories such that certain image formats support
lossless compression and lossless support. Table 2 tabulates different image formats
and supporting techniques.

Table 2. Different image formats

Lossy Compression Loseless Compression
JPG GIF
JPEG PNG
JPEG 2000 TIF
BMP
RAW

140 http://www.i-joe.org
Paper—Image Compression Using Neural Networks: A Review

JPG. JPG analyzes the image and eliminates detail that cannot be seen by a human
eye. It is used in continuous tone photographs and pictures. The data is saved in 24-bit
color. The amount of compression can be changed [27].
JPEG. Is a perfect way to store 24-bit photographic images, abbreviated as a Joint
Photographic Expert Group. It is used extensively in multimedia and online applications.
JPEG compresses the image in such a way that the detail is lost. Also for the encoding
of video is JPEG. Not suitable for diagrams and graphics [28] [29].
JPEG 2000. In JPEG the compression of the image is based on DCT coefficients,
while in JPEG 2000 wavelet method is used. It was formed by the “Joint Photographic
Experts Group committee” in the year 2000. JPEG 2000 addresses loss and lossless
strategies [30] [31].
GIF. Graphics Interchange Format is the GIF’s full format. The 8-bit colors are
allowed. It is also used for (black and white) text and images in gray scale. The biggest
downside of GIF’s is that images with over 256 colors cannot be used, but most color
pictures have over 256 colors, i.e. 24 bits per pixel. In animated pictures it is commonly
used. [32] [33].
PNG. PNG is Portable Network Graphics’ short form. This format is used for
image compression without loss. The PNG file format replaces the GIF file format as
compression is 10–30% higher than the GIF. PNG creates smaller files which allows
for more colors. PNG supports partial transparency. There are two variants of PNG-24
(224 = 16777216 supports) and PNG-8 (28 = 256 colors supported) [34] [35].
TIFF. It stands for Tagged Image File Format short form. It can be considered as
lossless format. The format is used primarily in photography and desktop publishing,
because of its extremely high quality. Saving a total of 8, 16 bits for a total 24 and 48
bits respectively per (red, green, blue). Compression of TIFF in web applications is
relatively poor and not used. They are particularly used in massive sizes in high-quality
prints [36].
BMP. The BMP is a short for the Bitmap Image file. The raster graphic data is
included with the BMP file. Showing devices do not impact this knowledge. This is
not important to see the BMP file image for a graphics adapter. It is typically used
with compression techniques without loss. It is also referred to as System Independent
Bitmap (DIB short) or a bitmap. It primarily stores digital images on the OS/2 and
Microsoft Windows operating systems in a wide variety. It can store digital images both
monochrome and 2-D color [37] [38].
From above, the key techniques for compressing images are transformed and pre-
dicted among the different coding frameworks. JPEG is the common image compres-
sion standard consisting of the fundamental modules of the transform/prediction as
shown in the Figure 4. The input image in JPEG is divided into 8_8 blocks that are not
overlapped, all of which will be translated into block DCT frequency (BDCT). A binary
stream is then compressed with DCT coefficients by quantizing and entropy coding for
each transformed block [39] [40].

iJOE ‒ Vol. 17, No. 14, 2021 141

Paper—Image Compression Using Neural Networks: A Review

Fig. 4. JPEG compression

3 Neural networks

Also known as artificial neural networks (ANN) are one of the most common
regression and classification models for machine learning and application in computer
literature. In the sense of non-linear transformation and classification, ANN has shown
good skills. The network comprises multiple layers of basic units called neurons that
communicate through weighted connections with each other. The neurons are triggered
by means of weighted neuron connections. In order to achieve non-linearity, all inter-
mediate layers are always activated [41] [42]. In tasks like image recognition neural
networks have achieved great success; handwriting recognition and optical character
[43] [44]. In fact, the recent explosive advancement of deep learning [45] has resulted
in the neural networks that are composed of several different layers of learners. In
a task of image recognition, for example, where the machine has to classify what
types of objects are in an image one layer might learn where the lines are in an image,
while another layer could learn how these lines organize to represent different forms
(e.g., books vs. people vs. pets).
Neural networks are a type of supervised learning based on biological structures and
mechanisms of the human brain. The neural networks generate predictions in the layers
of a set of connected nodes and neurons as stated above. As the neurons of this layer
only accept variables of the data set as an input, the first layer is called the input layer.
The last layer is the output layer since the final prediction is output (s). Hidden lay-
ers are between the input and the outputs, because the output in the network is only
relevant. The neurons generate a weighted amount from their input and transform the
weighted amount by means of some kind of nonlinear function such as the logit, hyper-
bolic tangent or the linear corrected function. In the next layer of the network, the
measured value from the function is passed on the weighted sum. Knowledge flows
from the input layer through the cached layers into the output layer through the neural

142 http://www.i-joe.org
Paper—Image Compression Using Neural Networks: A Review

network in one direction. If the output layer has been reached, it is stored and translated
to predictions [46].
Neural networks may have multiple hidden layers, and the functions used inside the
neurons can differ depending on the complexity of the data for which predictions are
needed. In principle, a cached neural layer and an output layer with enough neurons
are enough to learn about any binary task, and two layers plus an output layer will
get close to each return task [47]. Tens of hidden layers can be used to support the
discovery of deep learning patterns within the data in the becoming more common
field of deep learning. The overall neural network architecture defines which neurons
in subsequent layers feed their output into other neurons. The predicted variable shape
also controls the number of neurons in the output layer. The final output layer consists
particularly of one neuron for regression tasks (continuous results) and binary clas-
sification tasks (dichotomic result). Alternatively, the final output layer consists of a
neuron per potential value for multinomial classification tasks, where there are more
than two values in the categorical outcome variable. In this case, the expected class is
the neuron with the highest value. The basic neural network architecture is illustrated
in Figure 5. ANN is consisting of one input layer, one output layer and different hidden
layers, each containing various neurons.

Fig. 5. Neural network architecture

4 Image compression using neural network

In this section we present machine learning methods for image compression, mainly
originating in the late 1980s, especially from a neural network point of view. This
section consists of the history of neural network techniques, mainly the Multilayer
Perceptron Network (MLP), the Random Neural Network, the Recurrent Networks
(RNN) and the Convolutional Neural Network (CNN). The last section will present
the recent development of the techniques for image coding with generative adversarial
networks (GAN) [48].

iJOE ‒ Vol. 17, No. 14, 2021 143

Paper—Image Compression Using Neural Networks: A Review

Neural network is a recent compression tool because data processing is in parallel

manner and thus requires less time, and it is general performance is superior to any
other technique. It is therefore important to transform a neural network image informa-
tion efficiently. The data can be transformed by techniques like PCA based on factoriz-
ing techniques developed in linear algebra [49].

4.1 Multi-layer perceptron based image coding

MLP is made of a neuronal input (nodes) layer, a variety of hidden neuron layers,
and an ultimate neuron output layer [50]. The output of any neuron I in the MLP is
referred to as,

hi = σ  ∑ x j + ci  ,
N
w (4)
 j =1 ij 

where σ is the activation function, ci denotes the linear transform biasterm and wij
indicates the weight adjustable parameter which is the layer connection. The theoret-
ical study has shown that the MLP developed with one hidden layer can bring about
arbitrary accuracy of any continuous computable function [51]. This property shows
scenarios such as reduction of sizes and compression of data. The initiative of using
MLP to establish unitary transformations for spatial data as a whole is the compression
of image. [52] proposed a complete image compression method based on nonlinear
transform coding, and a framework to optimize it end-to-end for rate–distortion perfor-
mance. The compression method offers improvements in rate–distortion performance
over JPEG and JPEG 2000 for most images and bit rates. More remarkably, although
the method was optimized using mean squared error as a distortion metric, the com-
pressed images are much more natural in appearance than those compressed with JPEG
or JPEG 2000, both of which suffer from the severe artifacts commonly seen in linear
transform coding methods. Consistent with this, perceptual quality exhibits substan-
tial improvement across all test images and bit rates. The authors believe this visual
improvement arises because the cascade of biologically-inspired nonlinear transforma-
tions in the model have been optimized to capture the features and attributes of images
that are represented in the statistics of the data.
In order to find an optimized combination of binary code, a decomposition/decision
neural network, is then utilized by authors to solve the optimization problem. Every
8_8 patch of an image was compressed with retrospect propagation by a fully linked
neural network with 16 hidden units in [53]. This technique, however, specified the
neural network parameters for the specific number of binary codes, that can hardly be
adapted in optimal form for variable compression.
[54] proposed a novel iris recognition method-based Multilayer perception neural
network (MLPNN) and particle swarm optimization algorithm (PSO) to classify iris
images. The 2D Gabor kernel algorithm was used for feature extraction. The results of
testing on the CASIA-iris V3 database and UCI machine learning repository databases
indicate that the hybrid MLPNN–PSO algorithm is an effective, appropriate, stable,
robust, and competitive recognition method for human iris recognition.

144 http://www.i-joe.org
Paper—Image Compression Using Neural Networks: A Review

[55] focused on the study of dimension reduction. After using a multilayer neural
network to extract image features, the Principal Component Analysis (PCA) algorithm
is used to achieve dimension reduction. Specifically, they first leverage a multilayer
neural network to extract image features. Then they introduced and leveraged the PCA
algorithm to achieve dimension reduction. Aiming at the problem that it is difficult to
process high-dimensional sparse big data based on the PCA algorithm.

4.2 Random neural network based image coding

In 1989, a new class was developed of the random neural network. Random neural
networks operate in a different way than the MLP methods described above, where sig-
nals are optimized by the backpropagation process in the field of spaces. Random neural
network signals are transmitted as unit amplitude spikes. These neurons communicate
as a Poisson mechanism. The behavior of random neural networks have been studied
by some theoretical results [56].
[57] proposed a system that can be used to identify the image of Pekalongan batik
using Backpropagation Artificial Neural Network. Sogan’s motive is the result with the
highest level of accuracy that reaches 91.2 %% in testing the value of age = 100 and
the level of learning = 0.03. the second level of accuracy was 89.6% with age 100 and
learning level 0.02 in Jlamprang batik. The last level of 87.2% was obtained from the
Cap Combination batik and Tiga Negeri batik with the age of 100, the learning level
was 0.01 and 0.04. High and low accuracy values can be caused by structured motives
and the sharpness of batik colors will help increase the value of accuracy.
In [58] the implementation of back propagation neural network algorithm on image
compression system with good performance has been demonstrated. The back propaga-
tion neural network has been trained and tested for the analysis of different images. It
has been observed that the convergence time for the training of back propagation neural
network is very faster. Different attributes of compression such as compression ratio,
peak signal to noise ratio, bits per pixel are calculated. It has been observed that there
is significance change in compression ratio from .99 to .9556 in case of Cameraman
image. It has also been observed that there is significance improvement in peak signal
to noise ratio from 19.3181 to 20.722 in case Cameraman.

4.3 Convolutional neural network based coding

CNN’s latest success in high level computer vision tasks, including classifying
images, is outstrips conventional algorithms with a wide margin [59]. It also achieves
remarkable results, for example super-resolution and compression artifact reductions,
also in many low-level computer vision tasks. In order to characterize the link between
neighboring pixels, the CNN undertakes a convolutional operation with the cascade
convolution operations well matched with the hierarchical statistical properties. More-
over, the local receptive areas as well as the common weight of the convergence
operations often lower the trainable CNN parameters, reducing the chance of overfit-
ting significantly. Inspired by powerful CNN pictorial representation, several works
were done to explore the feasibility of compression of the CNN-based loss image. But

iJOE ‒ Vol. 17, No. 14, 2021 145

Paper—Image Compression Using Neural Networks: A Review

implementing the CNN model in the end-to-end compression is difficult straightfor-

ward. In general, CNN training relies on the back-propagation and stochastic gradient
descent algorithm, which require that the loss function is almost uniformly different
from qualified parameters, such as convergence weights and bias. Thanks to the image
compression quantization module, it generates virtually everywhere zero gradients,
avoiding CNN updating parameters. Moreover, it is difficult to optimize the classical
rate-distortion for a CNN-based compression frame. Since end-to-end CNN formation
requires an adjustable loss feature, but the rate must be estimated based on the popula-
tion distribution of whole quantifiable bins. [60] first implemented a scalar quantization
assumption, with an end-to-end, optimized CNN image compression system. By using
a pyramidal function fusion structure in an encoder and a CNN-based decoder post
processing filter.
[61] have presented a 12-layer deep convolutional neural network for compression
artifact suppression in JPEG images with hierarchical skip connections and trained
with a multi-scale loss function. The result is a new state-of-the-art ConvNet achieving
a boost of up to 1.79 dB in PSNR over ordinary JPEG and showing an improvement
of up to 0.36 dB over the best previous ConvNet result. We have shown that a network
trained for a specific quality factor is resilient to the QF used compress the input image
a single network trained for QF 60 provides a PSNR gain of more than 1.5 dB over the
wide QF range from 40 to 76. The obtained results are also qualitatively superior to
those of existing ConvNets.
[62] proposed a Deep Dual-Domain (D3) based fast restoration model to remove
artifacts of JPEG compressed images. It leverages the large learning capacity of deep
networks, as well as the problem-specific expertise that was hardly incorporated in the
past design of deep architectures. Accordingly, the authors take into consideration both
the prior knowledge of the JPEG compression scheme, and the successful practice of
the sparsity-based dual-domain approach. The successful combination of both JPEG
prior knowledge and sparse coding expertise has made D3 highly effective and efficient.

4.4 Recurrent neural network based coding

Contrary to the above-noted CNN architecture, RNN is a neural memory class to

store current comportments. Memory units in RNN, in particular, have relations with
themselves, which have previously transmitted transformed data from output. RNN
adjusts the actions of the present forward phase in order to respond to the context of the
current input by utilizing this stored information.
[63] first suggested a scaled-additive coding framework to reduce the number of
coding bits rather than approx. the estimate rate in CNN by using an RNN-based image
compression scheme.
A set of full-resolution lossy image compression methods based on neural networks
was proposed by [64]. Each of the architectures can provide variable compression rates
during deployment without requiring retraining of the network: each network need only
be trained once. All of the architectures consist of a recurrent neural network (RNN)-
based encoder and decoder, a binarizer, and a neural network for entropy coding.

146 http://www.i-joe.org
Paper—Image Compression Using Neural Networks: A Review

[65] proposed a method for lossy image compression based on recurrent, convolutional
neural networks that outperforms BPG (4:2:0), WebP, JPEG2000, and JPEG as mea-
sured by MS-SSIM. They introduced three techniques: hidden-state priming, spatially
adaptive bit rates, and perceptually-weighted training loss and showed that they boost
the performance of our baseline recurrent image compression architecture.
Recent article in [66] focused on developing deeper and more complex networks,
which significantly increased network complexity. Two effective blocks are developed:
analysis and synthesis block that employs the convolution layer and Generalized Divi-
sive Normalization (GDN) in the variable rate encoder and decoder side. The proposed
network utilizes a pixel RNN approach for quantization. Furthermore, to improve the
whole network, they encode a residual image using LSTM cells to reduce unneces-
sary information. Experimental results demonstrated that the proposed variable-rate
framework with novel blocks outperforms existing methods and standard image codecs.

4.5 Generative adversarial network based coding

GANs for short, GANs can be defined as the unsupervised learning task in the field of
machine learning that automatically requires the discovery and learning of regularities
and models in the input data, enabling the creation or production of new examples that
may have been extracted from the original data set. GAN’s are an intelligent way to
construct a generative model by posing the issue with two substructures as a supervised
learning problem: the model generator used to produce more features, and the model
that tries to classify features as true or false examples (generated).
[67] are proposing one of interesting works and highly optimized GAN-based image
compression, with a remarkable improvement of the compression ratio and the use of
large parallel computing cores from GPU in real time, networks compress the input
image into a very compact function space and the generative network is used to recon-
stituting the decoded image from the features. The clearest distinction between GAN
and RNN or CNN-based image compression is the inclusion of an opposite loss that
dramatically improves the subjective quality of the reconstructed image. The genera-
tive network and opposing network are jointly trained to increase the efficiency of the
generative model significantly.
The GAN approach in [68] greatly enhances compression, for example producing
standardized images for all compressed file quality levels, 2.5 times smaller than JPEG
and JPEG 2000, 2x smaller than WebP and 1.7x smaller than BPG. In this sense, quality
can be calculated using MS-SSIM, while PSNR metrics still don’t work. The compres-
sion of the light field image (LF) could achieve substantial coding gains based on the
progress made in GAN-based view synthesis by generating the missing views with
sampled background views in LF. GAN creates content, in fact, more compatible than
basic textures with the semantics of the original content. In particular, we can see the
differences in contents in specific textures while expanding reconstructed images.
[69] proposed a unified binary generative adversarial network (BGAN+) to simul-
taneously convert images to binary codes for both image compression and retrieval in
a multi-task fashion and an unsupervised way. By restricting the input noise variable
of generative adversarial networks (GAN) to be binary and conditioned on the features

iJOE ‒ Vol. 17, No. 14, 2021 147

Paper—Image Compression Using Neural Networks: A Review

of each input image, BGAN+ can simultaneously learn two binary representations per
image: one for image retrieval and one for image compression. To equip the binary
representation with the ability of accurate image retrieval and compression, we design
a novel loss function. The results show that the proposed method outperforms the exist-
ing retrieval methods with significant margins and the multi-task strategy is beneficial
for both tasks.
[70] tried to compact conceptuality by producing the most possible image semantic
zed information. A GAN-based image compression system, targeting bitrates below 0.1
bpp, is studied in depth, allowing various content generation degrees. Currently, GAN
compression is effective in narrow-domain images such as faces and still needs more
research on natural-image modeling in general [71] [72] [73].

5 Discussion and recommendations

On the basis of the analysis, we believe that the neural network advantages are three-
fold in picture and video compression. Firstly, the outstanding quality data adaptability
of a neural network exceeds the signal processing model because the network parame-
ters are based on a great deal of realistic details, while the models are designed by hand
using previous imaging and visual expertise in the state-of-the-art coding standards.
Second, the wider region of reception is commonly used in neural network models,
which not only utilizes the information given in the vicinity, but also improves coding
efficiency by utilizing samples from afar, but only use the conventional coding methods
for neighboring and distant samples are difficult to use. Third, both the structure and
the function that enables the joint compression optimization both for human view and
machine vision analysis can be described by the neural network. However, only high
compression efficiency for human viewing is achieved by the new coding standards.
We intend to further research the deep learning image and video compression in the
representation and distribution of image and videos with better quality and less bit rates
and the memory and computational efficiency in the realistic photo. Computing and
memory burdens are the greatest obstacle to the deployment of a deep learning picture.
For higher performance, larger neural networks are typically taken into consideration
with more layers and nodes but different network parameters’ efficiency is not well
explored.

6 Conclusion

Image compression is intended to pursue more efficient visual signals representa-

tion during retaining high-quality and growing significance in large-scale visual data.
The neural network compression techniques, in particular the recent deep learning
and video compression techniques, were discussed in this paper. The results are now
available. The survey previously presented here shows that the new end-to-end image
compression based on neural network is still in its infancy and is just overperforming
the JPEG2000 and battling against HEVC.

148 http://www.i-joe.org
Paper—Image Compression Using Neural Networks: A Review

On the basis of the discussion in this article, the neural network showed encour-
aging results for different compression tasks in the images and videos. Even if it is
machine sophistication and memories usage are still troublesome, the neural network
has achieved significant coding gains in addition to the cutting-edge video encoding
frameworks due to its high efficiency of prediction or compact presentation of image
and video signals.

7 References
[1] J. Nowaková, M. Prílepok, and V. Snášel, “Medical Image Retrieval Using Vector Quanti-
zation and Fuzzy S-tree,” J. Med. Syst., vol. 41, no. 2, 2017, doi: https://doi.org/10.1007/
s10916-016-0659-2
[2] R. K. Netalkar, H. Barman, R. Subba, K. V. Preetam, and U. S. N. Raju, “Distributed Com-
pression and Decompression for Big Image Data: LZW and Huffman Coding,” J. Electron.
Imaging, vol. 30, no. 5, 2021, doi: https://doi.org/10.1117/1.JEI.30.5.053015
[3] M. Xu, C. Li, S. Zhang, and P. Le Callet, “State-of-the-Art in 360 Video/Image Processing:
Perception, Assessment and Compression,” IEEE J. Sel. Top. Signal Process., vol. 14, no. 1,
pp. 5–26, 2020, doi: https://doi.org/10.1109/JSTSP.2020.2966864
[4] K. Ding, K. Ma, S. Wang, and E. P. Simoncelli, “Comparison of Full-Reference Image Qual-
ity Models for Optimization of Image Processing Systems,” Int. J. Comput. Vis., vol. 129,
no. 4, pp. 1258–1281, 2021, doi: https://doi.org/10.1007/s11263-020-01419-7
[5] M. Krishna, G. Srinivas, and P. V. G. D. Prasad Reddy, “Image Smoothening and Morpho-
logical Operators Based JPEG Compression,” J. Theor. Appl. Inf. Technol., vol. 85, no. 3,
pp. 252–259, 2016.
[6] Z. Chen, X. Hou, X. Qian, and C. Gong, “Efficient and Robust Image Coding and Trans-
mission Based on Scrambled Block Compressive Sensing,” IEEE Trans. Multimed., vol. 20,
no. 7, pp. 1610–1621, 2018, doi: https://doi.org/10.1109/TMM.2017.2774004
[7] A. J. Hussain, A. Al-Fayadh, and N. Radi, Image compression techniques: A survey in loss-
less and lossy algorithms, vol. 300. 2018. https://doi.org/10.1016/j.neucom.2018.02.094
[8] H. Zhang and L. Hu, “A Data Hiding Scheme Based on Multidirectional Line Encoding
and Integer Wavelet Transform,” Signal Process. Image Commun., vol. 78, no. January,
pp. 331–344, 2019, doi: https://doi.org/10.1016/j.image.2019.07.019
[9] P. P. Chavan, B. S. Rani, M. Murugan, and P. Chavan, “A Novel Image Compression Model
by Adaptive Vector Quantization: Modified Rider Optimization Algorithm,” Sadhana –
Acad. Proc. Eng. Sci., vol. 45, no. 1, 2020, doi: https://doi.org/10.1007/s12046-020-01436-9
[10] H. Sadeeq, A. Abdulazeez, N. Kako, and A. Abrahim, “A Novel Hybrid Bird Mating
Optimizer With Differential Evolution for Engineering Design Optimization Problems,”
Lect. Notes Data Eng. Commun. Technol., vol. 5, pp. 522–534, 2018, doi: https://doi.
org/10.1007/978-3-319-59427-9_55
[11] H. T. Sadeeq, A. M. Abdulazeez, N. A. Kako, D. A. Zebari, and D. Q. Zeebaree, “A New
Hybrid Method for Global Optimization Based on the Bird Mating Optimizer and the Differ-
ential Evolution,” Proc. 7th Int. Eng. Conf. “Research Innov. Amid Glob. Pandemic”, IEC
2021, no. August, pp. 54–60, 2021, doi: https://doi.org/10.1109/IEC52205.2021.9476147
[12] H. Sadeeq and A. M. Abdulazeez, “Hardware Implementation of Firefly Optimization Algo-
rithm Using FPGAS,” ICOASE 2018 – Int. Conf. Adv. Sci. Eng., pp. 30–35, 2018, doi:
https://doi.org/10.1109/ICOASE.2018.8548822

iJOE ‒ Vol. 17, No. 14, 2021 149

Paper—Image Compression Using Neural Networks: A Review

[13] R. Menassel, B. Nini, and T. Mekhaznia, “An Improved Fractal Image Compression Using
Wolf Pack Algorithm,” J. Exp. Theor. Artif. Intell., vol. 30, no. 3, pp. 429–439, 2018, doi:
https://doi.org/10.1080/0952813X.2017.1409281
[14] R. Menassel, I. Gaba, and K. Titi, “Introducing BAT Inspired Algorithm to Improve Fractal
Image Compression,” Int. J. Comput. Appl., vol. 42, no. 7, pp. 697–704, 2020, doi: https://
doi.org/10.1080/1206212X.2019.1638631
[15] S. Limuti, E. Polo, and S. Milani, “A Transform Coding Strategy for Voxelized Dynamic
Point Clouds,” Proc. – Int. Conf. Image Process. ICIP, pp. 2954–2958, 2018, doi: https://
doi.org/10.1109/ICIP.2018.8451254
[16] H. H. Nuha, “Lossless Text Image Compression using Two Dimensional Run Length Encod-
ing,” J. Online Inform., vol. 4, no. 2, p. 75, 2020, doi: https://doi.org/10.15575/join.v4i2.330
[17] U. Sharma, M. Sood, E. Puthooran, and Y. Kumar, “A Block-based Arithmetic Entropy
Encoding Scheme for Medical Images,” Int. J. Healthc. Inf. Syst. Informatics, vol. 15, no. 3,
pp. 65–81, 2020, doi: https://doi.org/10.4018/IJHISI.2020070104
[18] Z. Chen, S. Member, J. Xu, C. Lin, and W. Zhou, “Stereoscopic Omnidirectional Image
Quality Assessment Based on Predictive Coding Theory,” pp. 1–13.
[19] J. V. C. I. R, S. Yuan, and J. Hu, “Research on Image Compression Technology Based on
Huffman Coding Q,” J. Vis. Commun. Image Represent., vol. 59, pp. 33–38, 2019, doi:
https://doi.org/10.1016/j.jvcir.2018.12.043
[20] M. Testolina, T. Ebrahimi, S. Diego, and U. States, “Evaluation,” no. August, 2021, doi:
https://doi.org/10.1117/12.2597813
[21] X. Min et al., “Unified Blind Quality Assessment of Compressed Natural, Graphic, and
Screen Content Images,” IEEE Trans. Image Process., vol. 26, no. 11, pp. 5462–5474, 2017.
https://doi.org/10.1109/TIP.2017.2735192
[22] X. Zhang, W. Lin, and S. Wang, “Fine-Grained Quality Assessment for Compressed

Images,” IEEE Trans. Image Process., vol. 28, no. 3, pp. 1163–1175, 2019, doi: https://doi.
org/10.1109/TIP.2018.2874283
[23] D. Minnen, J. Ballé, and G. Toderici, “Joint Autoregressive and Hierarchical Priors for
Learned Image Compression,” no. Nips, 2018.
[24] C. Dong, Y. Deng, C. C. Loy, and X. Tang, “Compression Artifacts Reduction by a Deep
Convolutional Network,” pp. 576–584.
[25] K. Gregor and F. Besse, “Towards Conceptual Compression,” no. Nips, 2016.
[26] F. Jiang, W. Tao, S. Liu, J. Ren, X. Guo, and D. Zhao, “An End-to-End Compression Frame-
work Based on Convolutional Neural Networks,” pp. 1–13.
[27] G. K. Dziugaite and D. M. Roy, “A Study of the Effect of JPG Compression on Adversarial
Images arXiv : 1608.00853v1 [cs.CV] 2 Aug 2016,” no. Isba, 2016.
[28] N. Das et al., “Shield: Fast, Practical Defense and Vaccination for Deep Learning using
JPEG Compression,” pp. 196–204, 2018. https://doi.org/10.1145/3219819.3219910
[29] X. Fu, X. Wang, A. Liu, J. Han, and Z. Zha, “Learning Dual Priors for JPEG Compression
Artifacts Removal,” pp. 4086–4095, 2019.
[30] L. Theis and W. Shi, “Lossy Image Compression with Compressive Autoencoders,”

pp. 1–19, 2017.
[31] F. Yang, L. Herranz, Y. Cheng, and M. G. Mozerov, “Slimmable Compressive Autoencoders
for Practical Neural Image Compression,” pp. 4998–5007.
[32] B. O. Ayinde, “A Fast and Efficient Near-Lossless Image Compression using Zipper Trans-
formation arXiv : 1710.02907v2 [eess.IV] 10 Oct 2017,” pp. 1–13.
[33] C. Oswald, E. Haritha, A. A. Raja, and B. Sivaselvan, “An Efficient and Novel Data Cluster-
ing and Run Length Encoding Approach to Image Compression,” no. June 2020, pp. 1–16,
2021, doi: https://doi.org/10.1002/cpe.6185

150 http://www.i-joe.org
Paper—Image Compression Using Neural Networks: A Review

[34] K. Nakanishi, S. Maeda, T. Miyato, and D. Okanohara, “Neural Multi-scale Image Com-
pression,” 2018.
[35] M. Reconstruction, C. Framework, B. On, and P. N. G. Image, “Multiple Reconstruction
Compression Framework,” doi: https://doi.org/10.5121/ijcsity.2019.7401
[36] A. J. Qasim, R. Din, F. Qasim, and A. Alyousuf, “Review on Techniques and File Formats of
Image Compression,” Bulletin of Electrical Engineering and Informatics, vol. 9, no. 2, pp.
602–610, 2020, doi: https://doi.org/10.11591/eei.v9i2.2085
[37] C. Series, “Comparing Freeman Chain Code 4 Adjacency Algorithm and LZMA Algo-
rithm in Binary Image Compression Comparing Freeman Chain Code 4 Adjacency
Algorithm and LZMA Algorithm in Binary Image Compression,” 2021, doi: https://doi.
org/10.1088/1742-6596/1783/1/012045
[38] S. Ezilarasi, “Enhanced AMBTC Based Adaptive Image Compression Technique,” vol. 7,
no. 8, pp. 5690–5702, 2020.
[39] N. A. Kako, H. T. Sadeeq, and A. R. Abrahim, “New Symmetric Key Cipher Capable of
Digraph to Single Letter Conversion Utilizing Binary System,” no. May, pp. 1028–1034,
2020, doi: https://doi.org/10.11591/ijeecs.v18.i2.pp1028-1034
[40] N. A. Kako, A. M. Abdulazeez, and H. T. Sadeeq, “Effect of Colored Noise on Neuron Mem-
brane Size Using Stochastic Hodgkin-Huxley Equations,” 2021. https://doi.org/10.1109/
IEC52205.2021.9476110
[41] D. Q. Zeebaree, H. Haron, and A. M. Abdulazeez, “Gene Selection and Classification of
Microarray Data Using Convolutional Neural Network,” 2018 Int. Conf. Adv. Sci. Eng.,
no. November, pp. 145–150, 2018. https://doi.org/10.1109/ICOASE.2018.8548836
[42] J. N. Saeed and A. M. Abdulazeez, “Facial Beauty Prediction and Analysis Based on Deep
Convolutional Neural Network: A Review,” vol. 1, pp. 1–12, 2021. https://doi.org/10.30880/
jscdm.2021.02.01.001
[43] M. R. Mahmood, “A New Hand Gesture Recognition System Using Artificial Neural,”
no. April 2017, 2019.
[44] S. I. Saleem, A. M. Abdulazeez, and Z. Orman, “A New Segmentation Framework for Arabic
Handwritten Text Using Machine Learning Techniques A New Segmentation Framework for
Arabic Handwritten Text,” no. April, 2021, doi: https://doi.org/10.32604/cmc.2021.016447
[45] X. Zhang, S. Wang, Y. Zhang, and W. Lin, “High-Efficiency Image Coding via Near-Optimal
Filtering,” vol. 24, no. 9, pp. 1403–1407, 2017. https://doi.org/10.1109/LSP.2017.2732680
[46] K. Andersson, M. Zhou, and G. Van Der Auwera, “HEVC Deblocking Filter,” vol. 22,
no. 12, pp. 1746–1754, 2012, doi: https://doi.org/10.1109/TCSVT.2012.2223053
[47] C. Fu et al., “Sample Adaptive Offset in the HEVC Standard,” vol. 22, no. 12, pp. 1755–
1764, 2012. https://doi.org/10.1109/TCSVT.2012.2221529
[48] J. Lin, Y. Chen, Y. Huang, S. Lei, and T. He, “Motion Vector Coding in the HEVC Standard,”
no. c, 2013. https://doi.org/10.1109/JSTSP.2013.2271975
[49] X. Zhang, R. Xiong, and W. Lin, “Low-Rank based Nonlocal Adaptive Loop Filter for
High Efficiency Video Compression,” vol. 8215, no. c, pp. 1–12, 2016, doi: https://doi.
org/10.1109/TCSVT.2016.2581618
[50] S. Ma, X. Zhang, J. Zhang, C. Jia, S. Wang, and W. Gao, “Nonlocal In-Loop Filter :
The Future Way Towards Next-Generation Video Coding?” pp. 1–9, 2016. https://doi.
org/10.1109/MMUL.2016.16
[51] C. Tsai et al., “Adaptive Loop Filtering for Video Coding,” vol. 7, no. 6, pp. 934–945, 2013.
https://doi.org/10.1109/JSTSP.2013.2271974
[52] J. Ball and E. P. Simoncelli, “E—o i c,” 2017.
[53] M. W. Gardner and S. R. Dorling, “Artificial Neural Networks (The Multilayer
Perceptron)—A Review Of Applications In The Atmospheric Sciences,” vol. 32, no. 14,
pp. 2627–2636, 1998. https://doi.org/10.1016/S1352-2310(97)00447-0

iJOE ‒ Vol. 17, No. 14, 2021 151

Paper—Image Compression Using Neural Networks: A Review

[54] N. Ahmadi and G. Akbarizadeh, “A Hybrid Robust Iris Recognition Approach Using Iris
Image Preprocessing, 2D Hybrid Robust Iris Recognition Approach Using Iris Image
Pre-processing, Two-dimensional Gabor Features and Multi-layer Perceptron Neural Net-
work/PSO,” no. May 2018, 2017, doi: https://doi.org/10.1049/iet-bmt.2017.0041
[55] J. V. C. I. R, J. Ma, and Y. Yuan, “Dimension Reduction of Image Deep Feature Using PCA
Q,” vol. 63, 2019, doi: https://doi.org/10.1016/j.jvcir.2019.102578
[56]
K. Dimililer, “Backpropagation Neural Network Implementation for Medical
Image Compression,” Journal of Applied Mathematics, vol. 2013, 2013. https://doi.
org/10.1155/2013/453098
[57] R. A. Surya, A. Fadlil, and A. Yudhana, “Identification of Pekalongan Batik Images Using
Backpropagation Method Identification of Pekalongan Batik Images Using Backpropaga-
tion Method,” 2019, doi: https://doi.org/10.1088/1742-6596/1373/1/012049
[58] S. S. Panda, M. S. R. S. Prasad, M. N. M. Prasad, and C. S. Naidu, “Image Compression
Using Back Propagation Neural Network,” pp. 74–78, 2012.
[59] S. A. Dianat, N. M. Nasrabadi, S. Venkataraman, and R. Inst, “A . m .,” vol. 2, pp. 2793–2796,
1991.
[60] V. Quantization, O. Images, B. U. A. Neural-network, and C. Algorithm, “Vector Quan-
tization Of Images Clustering Algorithm,” no. October 1988, 2021, doi: https://doi.
org/10.1117/12.968954
[61] L. Cavigelli, P. Hager, and L. Benini, “CAS-CNN: A Deep Convolutional Neural Network
for Image Compression Artifact Suppression.”
[62] Z. Wang, D. Liu, S. Chang, Q. Ling, Y. Yang, and T. S. Huang, “D3: Deep Dual-Domain
Based Fast Restoration of JPEG-Compressed Images.”
[63] E. Gelenbe, “Random Neural Networks with Negative and Positive Signals and Prod-
uct Form Solution,” Neural Comput., vol. 1, no. 4, pp. 502–510, 1989, doi: https://doi.
org/10.1162/neco.1989.1.4.502
[64] G. Toderici et al., “Full Resolution Image Compression With Recurrent Neural Net-
works,” Proc. – 30th IEEE Conf. Comput. Vis. Pattern Recognition, CVPR 2017, vol. 2017,
pp. 5435–5443, 2017, doi: https://doi.org/10.1109/CVPR.2017.577
[65] N. Johnston et al., “Improved Lossy Image Compression with Priming and Spatially Adap-
tive Bit Rates for Recurrent Networks,” Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pat-
tern Recognit., pp. 4385–4393, 2018, doi: https://doi.org/10.1109/CVPR.2018.00461
[66] K. Islam, L. M. Dang, S. Lee, and H. Moon, “Image Compression with Recurrent Neural
Network and Generalized Divisive Normalization,” pp. 1875–1879, 2021, doi: https://doi.
org/10.1109/CVPRW53098.2021.00209
[67] F. Hai, K. F. Hussain, E. Gelenbe, and R. K. Guha, “Video Compression With Wavelets and
Random Neural Network Approximations,” Appl. Artif. Neural Networks Image Process.
VI, vol. 4305, no. April, pp. 57–64, 2001, doi: https://doi.org/10.1117/12.420926
[68] Y. Lecun, Y. Bengio, and G. Hinton, “Deep Learning,” Nature, vol. 521, no. 7553, pp. 436–
444, 2015, doi: https://doi.org/10.1038/nature14539
[69] J. Song, T. He, L. Gao, X. Xu, A. Hanjalic, and H. T. Shen, Unified Binary Generative
Adversarial Network for Image Retrieval and Compression, vol. 128, no. 8–9. 2020. https://
doi.org/10.1007/s11263-020-01305-2
[70] J. Ballé, D. Minnen, S. Singh, S. J. Hwang, and N. Johnston, “Variational Image Compres-
sion With a Scale Hyperprior,” 6th Int. Conf. Learn. Represent. ICLR 2018 – Conf. Track
Proc., 2018.
[71] Y. He, “Application of Panoramic Image Technology in Distance Teaching System,” Int. J.
Emerg. Technol. Learn., vol. 10, no. 6, pp. 27–31, 2015, doi: https://doi.org/10.3991/ijet.
v10i6.4848

152 http://www.i-joe.org
 Paper—Image Compression Using Neural Networks: A Review

[72] L. Favario and E. Masla, “A New Architecture for Cross-Repository Creation and Sharing of
Educational Resources,” Int. J. Emerg. Technol. Learn., vol. 12, no. 2, pp. 185–209, 2017,
doi: https://doi.org/10.3991/ijet.v12i02.6058
[73] S. Zhu, “An Online Interaction Mode for International Trade Practice Course Under Net-
work Environment,” Int. J. Emerg. Technol. Learn., vol. 12, no. 7, pp. 32–43, 2017, doi:
https://doi.org/10.3991/ijet.v12i07.7219

8 Authors

Haval Tariq Sadeeq is a lecturer in the Information Technology Department at

Duhok Polytechnic University in Duhok, Kurdistan Region of Iraq. Currently, he is a
PhD student at the Technical College of Informatics/DPU. His interests are Soft Com-
puting, Swarm Intelligence, Image Processing, Data Cryptography and Compression.
E-mail: [email protected]
Thamer Hassan Hameed is a lecturer at the University of Duhok, Duhok, in the
Kurdistan Region of Iraq. His interests are Swarm Intelligence, Machine Learning and
Image Processing. E-mail: [email protected]
Abdo Sulaiman Abdi is a lecturer in the Information Technology Department at
Duhok Polytechnic University, Duhok, Kurdistan Region of Iraq. His interests are com-
puter hardware, networking, and communications. E-mail: [email protected]
Ayman Nashwan Abdulfatah is a lecturer in the Information Technology Depart-
ment at Duhok Polytechnic University in Duhok, Kurdistan Region of Iraq. his interests
are DSP, Web programming, Computer Network. E-mail: [email protected].
krd

Article submitted 2021-08-06. Resubmitted 2021-10-12. Final acceptance 2021-10-12. Final version
published as submitted by the authors.

iJOE ‒ Vol. 17, No. 14, 2021 153

View publication stats