Technium Vol. 3, Issue 7 pp.

14-26 (2021)
ISSN: 2668-778X
www.techniumscience.com

Counterfeit Currency Recognition Using Deep Learning: A

Review
Twana Nasih Ahmed 1, Sabat Salih Muhamad 2
1
Horticulture and Landscape design Department, Bakrajo Technical institute, Sulaimani Polytechnic
University [email protected]
2
Computer Science Department, Faculty of Science, Soran University, Soran,
Erbil, Kurdistan Region, Iraq [email protected]

Abstract. This review article reviews some previous work in the literature about using
deep learning techniques for recognizing fake notes from the real notes. A human-visual
system can be utilized for identifying and confirming currency-notes. Yet, our eyesight
cannot sense everything, and it is frequently hard for humans to recognize genuine currency
without using technology assistance. Deep learning methods have been revealed to be more
efficient for numerous applications. Consequently, deep learning has been used to enhance
currency recognition accuracy. The methods that are used including Classic neural
networks, CNN, and GANs for recognizing fake notes.
Keywords: Currency Recognition, Deep Learning, CNN.

1. Introduction
Today, many recognition techniques are implemented to recognize images, recognize faces,
recognize car license plates, and recognize human behaviors. Currency is the primary average for
circulation, and Various countries' currencies have different qualities. However, when the value of currency
grows, there will be an increase in counterfeit currency. Counterfeit money might damage these nations'
interests. As a result, one of the hottest subjects and a critical issue at the moment is how to use recognition
technology to the genuine of money (Zhang, 2018).
Visual examination was used in the past to identify and genuine money, particularly currency notes.
Our eyesight cannot sense everything; sometimes, it is not easy for humans to distinguish genuine currency
from auth genuine entice currency without the aid of technology. Although UV-based-recognition With the
developing sophistication of counterfeiting systems, a system is still in operation. Moreover, it's getting
more impossible to tell the difference between fake and genuine currencies (Zhang, Yan and Kankanhalli,
2019).
From day to day, the progress of automated systems and techniques for currency recognition have
improved (Jadhav, Sharma and Bhandari, 2019). Artificial intelligence (AI) has become ingrained in a
variety of fields, including civil engineering, medical, and image processing. AI is built on neural networks
used for currency recognition and for a variety of different currency related tasks, including detecting
currency note-portrait, detecting fake-notes, recognizing currency note-serial-number, and extracting and
identification of currency note features(Veeramsetty, Singal and Badal, 2020).
Deep learning techniques, such as multilayer neural-networks, have recognized to be effective in
various applications (Zhang, Yan and Kankanhalli, 2019). This review article aims to explain the impact of
deep-learning techniques in recognizing the fake currency notes from the real notes. This review presents

14
 Technium Vol. 3, Issue 7 pp.14-26 (2021)
ISSN: 2668-778X
www.techniumscience.com

some related works that support using deep learning techniques for recognizing counterfeit currency in
machine learning and image processing. This review article is structured as follows: part two, a background
on the topic is given—part three, overviewing the related literature and applied methods. The findings are
demonstrated in part four. The last part concludes this review article.

2. Background
2.1 Deep Learning
“Deep Learning is a new area of Machine Learning research, which has been introduced with the
objective of moving Machine Learning closer to one of its original goals: Artificial Intelligence” (p.6).
Deep Learning explains learning numerous stages of illustration and concept that assist researchers in
analyzing data, including images, sounds, and texts (Deng and Yu, 2013).
Deep learning is commonly correlated of a neural network with several layers that could learn from
vast amounts of dataset, including a series of labeled images. Furthermore, this has mostly been applied to
the area of vision and voice (natural-language-processing). The weights of any layer are learned by
backpropagation in the deep learning process. Every layer have different effects on data analysis. Despite
its complexity, the method has been effectively extended to a wide range of classification and identification
problems (Zhang, Yan and Kankanhalli, 2019).
2.1.1 Deep Learning-Techniques:
1- Classic-Neural-Networks: Multilayer perceptron, where the neurons are associated to the continuous
layer, and it is often used to identify Fully Connected Neural Networks. There are 3 functions involved in
the presented system:
- Linear-function: Rightly-termed, this illustrates the only line that multiplies the inputs with the
continuous multiplier.
- Non Linear-function: This is categorized into three sections.:
A- “Sigmoid Curve: It is a function interpreted as an S-shaped curve with its range from 0 to 1.
B- Hyperbolic tangent (tanh) refers to the S-shaped curve having a range of -1 to 1.
C- Rectified Linear Unit (ReLU): It is a single-point function that returns 0 if the input value is less
than the set value and the linear multiple if the input value is greater.” (Vadapalli, 2020).
2- Convolutional-Neural-Networks: CNN is a more progressive version of the classic artificial neural-
network-model with a lot of potentials. It is designed to handle higher levels of complexity, as well as
pre-processing and data-compilation. It is based on the direction in which neurons in a human’s visual
cortex are arranged.
The CNN is built in four stages after the input data has been introduced in the convolutional system:
Convolution: is a procedure that generates feature maps from the input dataset and then applies a function
to these maps.
Max Pooling: It assists C-NN in detecting an image based on the assumed modification.
Flattening: In this stage, the dataset produced is then compressed for a CNN to examine.
Full Connection: This is sometimes defined as a hidden layer that collects the loss function for the
system(Veeramsetty, Singal and Badal, 2020).
3- Recurrent-Neural-Networks (RNNs): at the beginning, RNN was developed to aid in the prediction of
sequences; for instance, the Long-Short-Term-Memory (LSTM) algorithm is widely recognized for its
versatility. These networks are solely based on dataset sequences with variables input lengths (Vadapalli,
2020).
4- Generative-Adversarial-Networks: Using an efficient Generative system for generative counterfeit
examples and recognizing original data objects from the generative ones, known as (GANs)
(Goodfellow et al., 2020; Ali et al., 2019). In their research,(Ali et al., 2019) used GANs and can tell
the difference between fake and real banknotes with the highest level of precision. GAN.s are a type of
neural network which works on two elements and is quite interesting. The generative-network is a first,
and the discriminator network is the second. After applying GANs to the data-set and then using them
for classifying the genuine and counterfeit notes, promising results were obtained. Figure.1 shows the
proposed model, in which the Discriminator-Neural-Network (D) was learned using dataset from the

15
 Technium Vol. 3, Issue 7 pp.14-26 (2021)
ISSN: 2668-778X
www.techniumscience.com

training sets (original distribution) and generated-data (after perturbation i.e., added noises from the
latent space). During a training procedure of every Discriminator-network and the Generator-Network
(G), the loss functions were updated. Following training, the model could correctly categorize genuine
and counterfeit currency note with 80% accuracy. Moreover, discriminator portion of GANs served as
a classifier in image-recognition. The generator and discriminator, on the other hand, work together in
the learning process.

The basic idea behind currency note identification was that GANs highlighted two models. The
generative model-creates counterfeit banknotes, while the discriminative-model D determines whether
the dataset acknowledged from G came from the training dataset or from the generator itself. The priori
was discribed in input noise variables Pb-(b) to comprehend the generator's distribution P-g over dataset
(a), and after that characterizes a mapping to the banknotes dataset space as G-(b;θg), when G was a
discriminable function epitomized by a multilayer perceptron with parameters θg. According to GANs
networks, a second multilayer perceptron was also outlined D-(y;θd ) that outcomes only scalar.
Furthermore, banknotes that originated in the dataset, not the generator, were represented by D(a). The
discriminator D was learned on genuine banknotes to improve its chances of suitably identifying the
Dataset(a), and the example b was created using the generative-model G. This model G's loss-energy
function can be expressed mathematically as:

2.2 Single-shot-multibox-detector: is an extensive figure system in detecting items (Dai et al., 2016).
It creates bounding boxes using function maps from various layers. It would build various border
boxes based on various classifications; after that, one can find out what the item was. This system
can also be utilized in real time detecting, and it has the potential to be quicker than R_C-NN and

16
 Technium Vol. 3, Issue 7 pp.14-26 (2021)
ISSN: 2668-778X
www.techniumscience.com

Res Net. Furthermore, SSD not only makes sure that detection is quick, but also ensures detection
accuracy (Zhang, Yan and Kankanhalli, 2019).

2.3 Multilayer perception: this is one of the kinds of neural-network made up of input layer and node
that work together to generate one or more hidden and outcomes layers. To calculate the output, a
nonlinear function is needed, and the back-propagation-algorithm can be utilized to train MLP's
classification and regression. Furthermore, MLPs were feedforward neural network that were made
up of numerous layer of nodes connected by a one way link and trained using back-propagation
(Zhang, Yan and Kankanhalli, 2019).
Furthermore, an input vector X of N dimensions was the basic component of MLP, and the output
vector is M-dimensions (Zanaty, 2012). Traditional approaches may be replaced by multilayer
perception. Set x is a binary vector; o to be the scalar output; and w to be the weight. The
perceptron-algorithm is then demonstrated in Equation (1):

3. Related Works
A lot of researchers have contributed to the development of currency-recognition techniques in various
ways. The previous work in currency recognition techniques is reviewed in this section and showed the
main points in Tabel.2.
(Zhang, Yan and Kankanhalli, 2019) used CNN, SSD, and MLP for currency recognition. They
chose 5-NZD, 10-NZD, and 20 -NZD as the monetary denominations for their currency-recognition
research. They started by videoing each side of these three denominations separately. They attained fifty
image samples of individual denomination, for an overall 300 image samples in the data-set, with every
particular image having a resolution of 1280x720 pixels. For deep learning, this amount of data is
insufficient. As a result, they performed datasets augmentations on the original dataset and generated novel
datasets for accomplishing well training. They had a five-step procedure for data augmentation that included
zoomings to uniform sizes, random-clip or expand, randomly rotating, zooming to uniform sizes, and
adjusting colors randomly. They compare and contrast the experimental approaches and outcomes of two
further systems. Even though the recognition time of P-CA+ BPNN using the classifier FNN was 0.4249
seconds, the accuracy was higher. After comparision of F-NN, PRFNN, C-NN, and Ada-Boost, Technique
two discovered that F-NN had the highest accuracy. Even if they only achieve 96.6 percent accuracy with
CNN, the dataset training is intense, and the detection rate is fast, mentioned in table.1.

The main goal in (Jadhav, Sharma and Bhandari, 2019) study was to use deep learning to
distinguish between fake and real currency notes. The tests were carried out in order to identify fake
currency notes from India and Saudi Arabia. The camera was used to capture the images, and then the
proposed technique was used for extracting characteristics from the captured photos. Dissimilarity and
discontinuity were used to differentiate between genuine and counterfeit currency image. Furthermore,
extracted features were used to detect counterfeit money. The comparison degrees of the two notes were
used to determine whether a note was fake or genuine. They used a MATLAB tool to automatically
recognize counterfeit and real rupees and currency from other countries based on a deep-learning process.
Their proposed method is more cost-effective and efficient than the existing one.

17
 Technium Vol. 3, Issue 7 pp.14-26 (2021)
ISSN: 2668-778X
www.techniumscience.com

(Veeramsetty, Singal and Badal, 2020), in this research, the novel-lightweight-Convolutional-

Neural Network (C-NN) system for recognizing Indian currency notes was established for web and
cellphone applications efficiently. Moreover, to create the dataset, the total of 4657 images were taken. All
acceptable currency notes, such as Old-new 10-rupees note, old- new 20-rupees note, old-new 50-rupees
note, old and new 100-rupees note, new 200, 500, and 2000-rupees note, were used. Before providing the
photos to the models as inputs, they are all resized to 1024x1024 pixels. The currency note images are
augmented with data to increase the data-set size. Zoom, Rotattion-90, Rotattion-270, Tilting, Distortion,
and Fliping are the various types of augmentations used. There were 11657 images in the data-set after data
augmentation. The CNN in the following are used: one. Input-images, two. Convolution, three. Nonlinearity
operations were utilizing ReLU, four. Pooling, five. Flattening, six. Layer’s full connection. According to
findings, the proposed model outperforms six commonly recognized existing architecture in the area of
training and testing accuracies.
(Chowdhury, Jana and Parekh, 2020) proposed a system utilizing image-processing and deep-
learning methods to develop an automatic model for recognizing banknotes in India that was independent
of orienting the face sides of the banknotes. Moreover, images were gathered from scans of the originals
copies that were accessible on the internet as well as by photographing genuine banknotes. For the training
set, the system used eighty different images from eight categories, with ten sample for each catagory (this
includes both front and opposite sides). Furthermore, The eight denominational classes are – ten rupees,
twenty rupees, fifty rupees (old), fifty rupees (new), hundred rupees, two hundred rupees, five hundred
rupees, and two thousand rupees. And 34 images for test set. The samples of the notes were being switched
to their original horizontal direction after being pre-processed if they were rotated to the position. The
denominations were identified in 2 ways: first, extracting color and textures data as a feature for every
banknote and classified them by utilizing KNN. The second way was feeding pre-processed samples
banknotes in India to the C-NN; simultaneously, all the samples were twisted with a particular kernel, sub-
sampled to system feature map, after that classifying them by a dense-layer with softmax-classifier
corresponding to mentioned feature maps. The total accuracy of KNN was 91%, while C-NN has a 100%
accuracy rate.
(Ali et al., 2019) the machine_ assisted system dubbed DeepMoney was developed to recognize
counterfeit banknotes from real banknotes. State of-the-art systems of machine-learning named Generative-
Adversarial-Networks were used. Unsupervised learning was utilized by GANs to train a system that could
be used to make supervised predictions. Pakistani banknotes were applied to this technique. The total
method of the valid input was designed using image-processing and feature-recognition methods. In their
experiments, augmented examples of all image examples were utilized, demonstrating that a high accuracy
machine could be established to identify real paper-money. Their presented GANs technique for detecting
fake currency achieved an accuracy of 80%.
In (Laavanya, M. and Vijayaraghavan, 2019) The features of security-threads money note was
extracted to identify counterfeit money. The most widely used Deep-Neural Network method, transfer
training with Alex-net, was used to detect fake currency. Using augmentation, 100 images were created for
each note. Moreover, for increasing the database counts, augmentation processes such as resizing, and
rotating were used. The MATLAB 2018a were used. Transfer trained Alex networks with Adam-
optimization were used in the proposed method. The developed network was put to the test with a data-
base of 100 samples per money, 50 of which were captured images and 50 of which were augmented-
images. The image samples of Indian money 50-rupees, 200-rupees, 500-rupees, and 2000-rupees and
genuine banknotes were taken from a database. The average accuracy was equal to 81.5 percent and 75
percent for genuine and fake currencies, correspondingly. The model's overall processing time was three
seconds. Thier proposed system outperformed the VGG network by about 26% in terms of accuracy.
(Pachón, Ballesteros and Renza, 2021) introduced a comparison between custom models and
models based on transfer learning in the mission of banknote recognition and counterfeit detection. The
most suitable freezing-points in C-NN architectures (sequential, residual, and Inception) were known for
the transfer learning (TL) strategy. A unique system built on a sequential C-NN of the AlexNet type was
proposed. A Colombian banknote dataset was used to train and compare both the TL and custom models.

18
 Technium Vol. 3, Issue 7 pp.14-26 (2021)
ISSN: 2668-778X
www.techniumscience.com

7280 images, including rotations, partial views of the banknotes, and non-aligned banknotes, were captured
and stored (560 images per class). Natural light may enter the system due to environmental conditions. Data
augmentation by varying lighting conditions was used to solve this problem. 60% of the data-sets were used
for learning, 20% for validating, and 20% for testing. After comparison between the system and the four
TL_based models (AlexNet, SqueezeNet, ResNet18, InseptionV3), ResNet18 had the highest accuracy,
with a score of 100 percent. The proposed custom network had the shortest inference times, with
performance up to 6.48 times quicker in C-P-U and 16.29 times quicker in G-P-U than the inference time
with transfer learning systems.
In (Pham et al., 2020), this research aimed to recognize fake banknotes that made by re-production
with general purpose imaging equipment for human with vision disabilities. They suggested a method that
used CNN for detecting fake banknotes using banknote samples taken by cellphone cameras in noticeable
light. Experiments on a self collected data-set of USD, Euro€, Korean-won₩, and Jordanian-dinarJOD
banknotes revealed that their system outperforms state-of-the-art methods in the use of fake recognition.
In (Navya Krishna et al., 2019) Automatic Fake Currency Recognition System (AFCRS) was
created with CNN and the VGG 16 architecture to detect counterfeit paper currency and determine if it was
genuine. They used Keras deep learning library in python to made their CNN. After fitted model, the
screenshots were captured in their experimented method, AFCRS. The difference between a fake and an
original currency note was clearly noticeable. As a result, it was advisable to use it as a smartphone
application. application that was merely beneficial to society.
(Hoang and Vo, 2019) presented a method for recognition of paper currency based fundamental
image processing using deep-learning for feature-extraction and recognition. Furthermore, image set such
as 10000, 20000, 50000, 100000, 200000, 500000 Vietnamese dong and some other national currencies
like dollars and euro were used and set of images taken from a counterfeit banknote that was made with a
different material than the original banknote's paper. There was a total of 9,736 real banknote samples and
1,083 fake banknote samples, all of which were standardized to a similar resolution of 200x450 pixel. They
used deep CNN for both feature extraction and recognition and SVM for classification. In place of face
value recognition and counterfeit detection, the CNN model was used to extract features. Using the SVM
technique, the color sample was divided into classes based on face value and national currency. This method
was discovered to be useful in real-life applications, with a 99.97% accuracy rate.
(Sakazawa et al., 2019) proposed applying a watermark to the learned D-NN model to secure their
copyrights. In their paper presented a cumulative and visual watermark decoder and demonstrated that this
performs in all cases. In the case of the MNIST water-mark input image samples, twenty image samples
were chosen from a total of 60,000 to decode the watermark. At the twentieth sample in the term of MNIST,
the decoded patterns were sufficiently obvious. Lum-[0][1], Lum-[2][0], Lum-[2][2] of#20 have identical
pixels value in Cifar+WM, but if a meaningful pattern was embedded, it could be visually identifiable.
Higher contrast was reached with lower accuracy. It means that embedding a strong watermark devalues
the host DNN model's original mission. After comparing MNIST performed better (greater accuracy at the
same contrast) than the other tests. It was because MNIST image samples for embedding had totally
diverse characteristics from Cifar-10 dataset for initial learning. Cifar+WM images, on the other side, share
a lot of similar features with Cifar10 data. As a result, watermark embedding input images could be of a
different kind than the host-model training dataset.
In (Tajane et al., 2018) paper, they presented the DeepLearning method to recognize and detect
Indian-coins. Their proposed system was divided into two phases, the first of which was the software
application of the work using MATLAB, then the second of which was the use of a Raspberry-Pi as a
controlling board to control the system's operation. The project's hardware implementation began with
capturing a photo of the implanted coin, which was after that processed by using the Raspberry Pi controller.
The system was learned on over 1600 image samples and could differentiate between four types of items,
including 1, 2, 5, and 10 rupee-coins. Thier trained system was put to the test on a variety of normal and
own documented data-sets that include rotational, interpreted, and shifted image samples. They trained
AlexNet to be able to recognize different coins in the first phase, and then they placed the 1-rupee coin in
the face of the camera, and the learned network correctly recognized the coin as 1 rupee and 2 rupees.

19
 Technium Vol. 3, Issue 7 pp.14-26 (2021)
ISSN: 2668-778X
www.techniumscience.com

AlexNet was trained and tested on a dataset of five- and ten-rupee coins. Recognition accuracy and response
time were used to calibrate the output system. The obtained results presented method exceed traditional
systems.
(Kamble et al., 2019) created a Deep CNN model that identifies counterfeit notes without manually
extracting image features. The model learns from the created dataset and can detect a counterfeit note by
training it on it. They used 40,000 images as a dataset, 32,000 image samples for training, and 8,000 image
samples for testing. Moreover, the median_blur filter was used to pre-process the images for training. The
model's architecture consisted of five convolution-layers, a flatten-layer, and four fully connected-layers.
Furthermore, training dataset is subdivided once more into 30 000 image samples for training and 2 000
image samples for cross validation. The system was learned, and the training and cross-validations accuracy
was determined based on the built architecture. After that, of image in the testing data set was preprocessed
(using the median-blur filter). The accuracy of the testing was around 85.6 percent. Thus, the training and
validation accuracy was 98.57 percent and 96.55 percent.

Table.2. Summarizing the main points of the literature review.

Authors Aim Method Data-set Result Limitatio Yea
n r
(Zhang, They gave used 5 NZD, 10 NZD, They compare and Their 201
Yan and a thorough Convoluti
and 20 NZD. contrast the system 9
Kankanhall overview onal Total of 300 experimental methods just work
i, 2019) of Neural images. and outcomes of two for NZD
currency Networks other methods. denominat
recognitio (CNN), Although the ion
n. Single recognition time of currency
shot PCA+ BPNN using recognitio
multibox the classifier F-NN n.
detector was 0.4249 seconds,
(SSD), the accuracy was
and higher. After the
Multilayer comparisons of FNN,
perception PRFNN, C-NN, and
(MLP) for AdaBoost, Technique
currency II discovered that
recognitio FNN had the highest
n accuracy. Even if
they only achieve
96.6 percent accuracy
with CNN, The
dataset is well-
trained, the detecting
speed was fast.
(Jadhav, used deep used bank to identify fake The proposed system Extracting 201
Sharma and learning to notes from currency notes was more profitable security 9
Bhandari, distinguis different from India and and effective in features of
2019) h between countries Saudi Arabia. comparison to the currency
fake and by current system. is not
real extracting considered
currency and .
notes. analyzing
their
features in

20
 Technium Vol. 3, Issue 7 pp.14-26 (2021)
ISSN: 2668-778X
www.techniumscience.com

detail
using deep
learning
(Veeramset a novel A Old-new 10- Their developed The 202
ty, Singal lightweigh lightweigh rupees note, old- model, which was drawbacks 0
and Badal, t t three new 20-rupees trained with an image of their
2020) Convoluti layers note, old-new 50- size of 1024x1024, a research is
onal CNN rupees note, old batch-size of 16, the their
Neural and new 100- learning degree system
Network rupees note, new 0.0001, had a 100% cannot
(CNN) 200, 500, and accuracy; also the identify
model for 2000-rupees note. testing-accuracy is counterfeit
recognizin There were 11657 87.5 percent. currency
g Indian photos in all. The notes.
currency model was
notes was trained with 9326
developed images and
for validated with
efficient 2331 images.
web and
mobile
applicatio
ns.
(Chowdhur Utilizing CNN, The eight The total accuracy of Each of 202
y, Jana and methods KNN denominational KNN was 91%, while the rotated 0
Parekh, including employed classes are – ten C-NN had a 100% photos
2020) image- on the rupees, 20 rupees, accuracy rate. were shot
processing extracted fifty rupees (old), in the
and deep- techniques fifity rupees equal
learning, . (new), hundred backdrop,
build an Classifier. rupees, two and the
integrated G-LCM is hundred rupees, rotating at
method utilized five hundred 180 was
for for color, rupees, and two ignored
recognizin feature thousand rupees.
g Indian extraction.
banknotes.
(Ali et al., They used Generativ Photos of rupees. GANs framework for Their 201
2019) a e 50 rupees. 500, counterfeit money system 9
machine- Adversari and rupees. 1000 detection achieved an just work
assisted al were used in the accuracy of 80%. with
device Networks dataset. Pakistani
called (GANs) banknotes.
DeepMon
ey to solve
the
problems
of
classifying
currency
notes as

21
 Technium Vol. 3, Issue 7 pp.14-26 (2021)
ISSN: 2668-778X
www.techniumscience.com

counterfeit
or real.
(Laavanya, The Transferin
100 images per For genuine currency Extracting 201
M. and characteris g learned
currency, 50 of and faking, the security 9
Vijayaragh tics of Alex- which were average accuracy features of
avan, 2019) security- networkscaptured images were 81.5 percent , 75 currency
threads with and 50 of which percent is not
currency adam- were augmented considered
note was optimizati
images. The .
extracted on images of Indian
to identify money 50-rupees,
fake 200-rupees, 500-
currency. rupees, and 2000-
rupees genuine
banknotes were
used.
(Pachón, compariso custom 7280 images of a The proposed custom The 202
Ballesteros n between model Colombian network had the biggest 1
and Renza, custom based on a banknotes. shortest inference drawback
2021) models sequential times, with of a
and CNN of performance more custom
models the than 6.48 times model is
based on AlexNet quicker in C-P-U, the length
transfer type 16.29 times quicker in of time it
learning in GPU than the takes to
the inference time with train it, as
mission of transmission learning well as the
banknote systems. need for a
recognitio sufficientl
n and y diverse
counterfeit dataset to
detection. allow for
generaliza
tion
during
training.
(Pham et recognize CNN self-collected Their system Their 202
al., 2020) fake data-set of USD, overachieve state of system for 0
banknotes Euro€, Korean- the art techniques in detecting
that made won₩, and the course of fake and
by Jordanian- detection. localizing
reproducin dinarJOD banknotes
g with banknotes. using
general photograp
purpose hs taken
imaging by mobile
devices cameras
for human and
with combining
visual false
detection

22
 Technium Vol. 3, Issue 7 pp.14-26 (2021)
ISSN: 2668-778X
www.techniumscience.com

impairmen and
t. denominat
ion
labeling
does not
perform
well.
(Navya to detect CNN and The images were The method of The data 201
Krishna et counterfeit the VGG gathered from a detection a fake note collection 9
al., 2019) paper 16 variety of Google was fast and simple was
currency architectur sources as well as under the trained limited
and e the Churan model. and did
determine labeled Children's not
if it was Bank of India. represent
genuine. Initial notes and the real-
false notes are world
divided into two situation
categories in this of
data collection. counterfeit
currencies
.
(Hoang and recognitio deep CNN 10k, 20k, 50k, This method was High 201
Vo, 2019) n of paper for both 100k, 200k, 500k discovered to be computati 8
currency feature VND and some useful in real-life onal cost
extraction other national applications, with a and More
and currencies like 99.97% accuracy rate. national
recognitio dollars and euro. currencies
n and There was a total and face
SVM for of 9,736 real values
classificati banknote samples were not
on. and 1,083 fake taken into
banknote samples considerat
ion when
applying
for the
real-time
banknote
inspection
scheme.

(Sakazawa To secure 1:integrati In M-N-I-S-T The initial task's Their 201

et al., 2019) the ng watermark input precision was 0.752 method is 9
copyright, cumulativ data, the with the MINST case unable to
watermark e- and watermark was and 0.725 with Cifar- handle
ing is visual decoded using 20 WM, respectively. broader
applied to watermark images out of watermark
trained decoding 60,000. patterns,
DNN from D- such as
models. NN company
systems, logos and
T-Ms, due

23
 Technium Vol. 3, Issue 7 pp.14-26 (2021)
ISSN: 2668-778X
www.techniumscience.com

2: to a
Enabling limitation
third party of class
identificati size.
on by
including
a subclass
of the
massive
decoding
dataset

(Tajane et recognizin Deep 1600 images The obtained results In real- 201
al., 2018) g and learning could confirmed that the time 8
detecting approache differentiate presented method embedded
Indian s. between 4 item exceed traditional systems,
coins. types, including systems. no
1, 2, 5, and 10 ₹ algorithms
coins. were used.

(Kamble et counterfeit Deep 500 rupees real The accuracy of the Since their 201
al., 2019) notes CNN and fake, 2000 testing was around data set 9
recognitio rupees real and 85.6 percent. The was small,
n. fake. 40,000 training outcome was the model
images as a data- 98.57 percent, and the was
set, 32,000 validation accuracy unable to
image-samples was 96.55 percent. learn and
samples for improve.
learning and
8,000 image
samples for
testing, and 2000
for
crossvalidation.

4. Discussion
After studying all the reviewed articles, we came to understand the following points. First of all, most
of the articles chose to use CNN to recognize currency notes after comparing with other previous works
because it can get more accuracy. However, some of the articles used GANs for the same reason.
The use of CNN has several advantages, including, such as, CNN is well-known for its architecture,
and the best part is that no feature extraction is needed. The main advantage of C-NN over its predecessor
is that it can identify crucial feature without requiring people interaction.
Some drawbacks of CNN are, Images in different positions are classified; because of operations like max
pool, a Convolutional neural network is substantially slower, If the CNN has many layers, the training phase
can take a long time, if the machine does not have a powerful GPU and a CNN needs a big Data-set to
process and train the neural network.

24
 Technium Vol. 3, Issue 7 pp.14-26 (2021)
ISSN: 2668-778X
www.techniumscience.com

5. Conclusion
After reviewing the above articles, we understood that deep learning techniques are crucial to currency
note recognition. Based on deep learning literature review, most of the researchers using Convolutional
Neural Networks for recognizing fake notes from the real notes. This technique can get more accuracy after
comparing this technique with other previous works.
Future researchers should build a system that can be used for currency recognition for all countries
around the world because some of the models identify few different currencies. On the other hand, some of
the current models cannot identify fake money. Furthermore, the researchers should create an application
model for cellphones and web-application to recognize fake and real money, especially for people with
visual disabilities. Finally, the researchers should crucially work on extracting security thread features.

References

[1]. Ali, T. et al. (2019) ‘DeepMoney: Counterfeit money detection using generative adversarial
networks’, PeerJ Computer Science, 2019(9). doi: 10.7717/peerj-cs.216.
[2]. Chowdhury, U. R., Jana, S. and Parekh, R. (2020) ‘Automated System for Indian Banknote
Recognition using Image Processing and Deep Learning’, in 2020 International Conference on
Computer Science, Engineering and Applications, ICCSEA 2020, pp. 1–5. doi:
10.1109/ICCSEA49143.2020.9132850.
[3]. Dai, J. et al. (2016) ‘R-FCN: Object detection via region-based fully convolutional networks’, in
Advances in Neural Information Processing Systems.
[4]. Deng, L. and Yu, D. (2013) ‘Deep learning: Methods and applications’, Foundations and Trends
in Signal Processing, pp. 197–387. doi: 10.1561/2000000039.
[5]. Goodfellow, I. et al. (2020) ‘Generative adversarial networks’, Communications of the ACM,
63(11). doi: 10.1145/3422622.
[6]. Hoang, V. D. and Vo, H. T. (2019) ‘Hybrid discriminative models for banknote recognition and
anti-counterfeit’, in NICS 2018 - Proceedings of 2018 5th NAFOSTED Conference on Information
and Computer Science, pp. 394–399. doi: 10.1109/NICS.2018.8606900.
[7]. Jadhav, M., Sharma, Y. K. and Bhandari, G. M. (2019) ‘Currency Identification and Forged
Banknote Detection using Deep Learning’, in Proceeding - 1st International Conference on
Innovative Trends and Advances in Engineering and Technology, ICITAET 2019, pp. 178–183.
doi: 10.1109/ICITAET47105.2019.9170225.
[8]. Kamble, K. et al. (2019) ‘Counterfeit Currency Detection using Deep Convolutional Neural
Network’, in 2019 IEEE Pune Section International Conference, PuneCon 2019, pp. 1–4. doi:
10.1109/PuneCon46936.2019.9105683.
[9]. Laavanya, M., &amp and Vijayaraghavan, V. (2019) ‘Real Time Fake Currency Note Detection
using Deep Learning’, International Journal of Engineering and Advanced Technology, 9(1S5),
pp. 95–98. doi: 10.35940/ijeat.a1007.1291s52019.
[10]. Navya Krishna, G. et al. (2019) ‘Recognition of fake currency note using convolutional neural
networks’, International Journal of Innovative Technology and Exploring Engineering, 8(5).
[11]. Pachón, C. G., Ballesteros, D. M. and Renza, D. (2021) ‘Fake banknote recognition using deep
learning’, Applied Sciences (Switzerland), 11(3), pp. 12–81. doi: 10.3390/app11031281.
[12]. Pham, T. D. et al. (2020) ‘Deep Learning-Based Fake-Banknote Detection for the Visually
Impaired People Using Visible-Light Images Captured by Smartphone Cameras’, IEEE Access, 8,
pp. 63144–63161. doi: 10.1109/ACCESS.2020.2984019.
[13]. Sakazawa, S. et al. (2019) ‘Visual Decoding of Hidden Watermark in Trained Deep Neural
Network’, in Proceedings - 2nd International Conference on Multimedia Information Processing
and Retrieval, MIPR 2019. doi: 10.1109/MIPR.2019.00073.
[14]. Tajane, A. U. et al. (2018) ‘Deep Learning Based Indian Currency Coin Recognition’, in 2018
International Conference On Advances in Communication and Computing Technology, ICACCT

25
 Technium Vol. 3, Issue 7 pp.14-26 (2021)
ISSN: 2668-778X
www.techniumscience.com

2018, pp. 130–134. doi: 10.1109/ICACCT.2018.8529467.

[15]. Vadapalli, P. (2020) Top 10 Deep Learning Techniques You Should Know About, upGrad blog.
[16]. Veeramsetty, V., Singal, G. and Badal, T. (2020) ‘Coinnet: platform independent application to
recognize Indian currency notes using deep learning techniques’, Multimedia Tools and
Applications, 79(31–32), pp. 22569–22594. doi: 10.1007/s11042-020-09031-0.
[17]. Zanaty, E. A. (2012) ‘Support Vector Machines (SVMs) versus Multilayer Perception (MLP) in
data classification’, Egyptian Informatics Journal, 13(3), pp. 177–183. doi:
10.1016/j.eij.2012.08.002.
[18]. Zhang, Q. (2018) ‘Currency recognition using deep learning’, (Doctoral dissertation, Auckland
University of Technology).
[19]. Zhang, Q., Yan, W. Q. and Kankanhalli, M. (2019) ‘Overview of currency recognition using deep
learning’, Journal of Banking and Financial Technology, 3(1), pp. 59–69. doi: 10.1007/s42786-
018-00007-1.

26