Journal of King Saud University - Computer and Information Sciences 36 (2024) 102038

H O S T E D BY Contents lists available at ScienceDirect

Journal of King Saud University - Computer and

Information Sciences
journal homepage: www.sciencedirect.com

Full Length Article

Jordanian banknote data recognition: A CNN-based approach with

attention mechanism
Ahmad Nasayreh a, Ameera S. Jaradat a, Hasan Gharaibeh a, Waed Dawaghreh a,
Rabia Mehamad Al Mamlook b, c, Yaqeen Alqudah a, Qais Al-Na’amneh d,
Mohammad Sh. Daoud e, Hazem Migdady f, Laith Abualigah g, *
a
Department of Information Technology and Computer Sciences, Yarmouk University, Irbid 211633, Jordan
b
Department of Business Administration, Trine University, IN, USA
c
Department of Industrial Engineering, University Zawia, Tripoli, Libya
d
Department of Cyber Security and Cloud Computing, Applied Science Private University, Amman 11937, Jordan
e
College of Engineering, Al Ain University, 112612 Abu Dhabi, United Arab Emirates
f
CSMIS Department, Oman College of Management and Technology, 320 Barka, Oman
g
Hourani Center for Applied Scientific Research, Al-Ahliyya Amman University, Amman 19328, Jordan

A R T I C L E I N F O A B S T R A C T

Keywords: Identifying counterfeit banknotes is crucial in financial transactions, as the process of identification cannot be
Convolution neural networks handled by ATMs or vending machines. The recent developments in technology, particularly the smart systems
Attention mechanism that are integrated with cameras and artificial intelligence (AI) tools, allow for the distinction of currency and the
Banknotes classification
detection of counterfeit. In this study, we are suggesting an approach for identifying counterfeit Jordanian
Deep learning
banknotes and differentiating them from genuine ones. The suggested approach collaborates deep learning
through a convolutional neural network (CNN) and another attention mechanism which contributes to focusing
on features of importance while avoiding features of less importance. The proposed model has proven its ability
to recognize counterfeits with high performance and accuracy while focusing on the important features
extracted. The study made use of a data set from Kaggle that includes a collection of Jordanian banknotes in five
different denominations. Image processing techniques were employed to produce artificial images by boosting
the brightness of real ones. Eight trained models and the suggested model were compared. It demonstrated
excellence with encouraging outcomes, achieving 96% accuracy, 96.6% precision, 96.4% recall, and 94.5% f1-
score. Also, we tested our approach on two datasets Indian dataset and DS1, DS2, and DS3 datasets, we obtained
88% and 99.9% accuracy, respectively. The achievement of detecting counterfeit Jordanian banknotes is proof
that a well-established AI model contributes to dealing with security vulnerabilities in many institutions.

1. Introduction sensors are used but with their susceptibility to noise and high imple­
mentation costs their spread remains restricted and not widely available
In the age of tremendous technological development, every aspect of for the general public (Office, 2023), to overcome these limitations
our lives now depends on the use of the internet. Finance and currency computer vision and traditional Machin Learning techniques such as k-
transactions are no exception, and while we see and observe the wide­ nearest neighbor (Youn et al.,May, 2015) genetic algorithms (Cao and
spread of 40digital transactions, banknotes are still dependable and Liu, 2010) and fuzzy systems (AL-Gawda‘et al., 2016) were being used,
usable by individuals such that the automated handling of the currency although these traditional solutions tend to yield lower accuracy when
is still an issue. Due to the fact that defrauding using forged banknotes processing wrinkled or soiled banknotes and they lake for generalization
continues to rise the development of detection systems becomes pivotal. abilities (Lee et al., 2017).
Money-counting machines that have detection systems with various The capabilities of smartphones nowadays, especially their cameras,
embedded sensors such as magnetic, infrared (IR), or ultraviolet (UV) are becoming much more powerful. Considerably, modern computer

* Corresponding author.
E-mail address: [email protected] (L. Abualigah).

https://doi.org/10.1016/j.jksuci.2024.102038
Received 5 December 2023; Received in revised form 29 February 2024; Accepted 11 April 2024
Available online 13 April 2024
1319-1578/© 2024 The Authors. Published by Elsevier B.V. on behalf of King Saud University. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/by/4.0/).
A. Nasayreh et al. Journal of King Saud University - Computer and Information Sciences 36 (2024) 102038

vision methodologies, especially Convolutional Neural Networks (CNN) study, in the third section we discuss the proposed method and data
and Deep Learning (DL) are widely used to harness their efficient pattern processing, in the fourth section we demonstrate the results and anal­
extraction capabilities. CNN has proved its superior performance ysis, and in the last section we present the conclusion of our study.
compared to traditional Machine Learning techniques in image classi­
fication problems. It has been utilized in recognizing various forged 2. Related work
currencies from different countries such as the Pakistani Rupee (Ali
et al., Sep. 2019), Indian Rupee (Pham et al., Feb. 2019), Euro, and 2.1. Background
United States Dollar (Han and Kim, Aug. 2019). Moreover, proposals
from researchers have implemented the use of transfer learning for Convolutional neural Networks (CNNs) in image classification
instance the detection of Euro currency (“Evaluation of Texture De­ Convolutional Neural Networks (CNNs): revolutionized image
scriptors, 2023), others created a custom CNN architecture to be used in classification by efficiently learning hierarchical feature representations
Dollar, and Jordanian Dinar counterfeit currency detection (Park et al., from picture data. Unlike typical machine learning algorithms, these
2020) (Pham et al., 2020). Though these produced good performance networks automatically learn spatial feature hierarchies using a sys­
accuracies they lack clarity regarding which approach is more suitable tematic method that incorporates fully connected, pooling, and con­
for detecting counterfeit banknotes. volutional layers (Albawi et al., Oct. 2018). Convolutional layers use
Counterfeit banknote detection is critical for safe financial integrity convolution to extract high-level features from input pictures
and preventing fraud. This research aims to create a reliable system for (Namatēvs, 2017). Formally, the convolution technique is defined as an
identifying fake Jordanian banknotes, addressing limitations in prior equation with a kernel g(x,y) that traverses the input image to generate a
studies which lacked focus on this currency. The core challenge is feature map. This map makes it easier to identify spatial feature hier­
accurately differentiating between genuine and counterfeit Jordanian archies by emphasizing key areas of the image.
banknotes. Traditional detection methods have shortcomings, necessi­ Pooling Layers: also known as max pooling, are used after convo­
tating an improved approach leveraging modern techniques (Khan et al., lution to minimize feature map dimensionality while retaining sub­
Dec. 2020). This study proposes a novel solution employing convolu­ stantial information. Pooling techniques combine features from specific
tional neural networks (CNNs) and attention mechanisms. The key neighborhoods to minimize computational complexity and increase
contribution is tailoring these state-of-the-art deep learning techniques detection efficiency (Yu et al., 2014).
specifically for counterfeit Jordanian banknote detection using a new Fully Connected Layers: The final layer of a CNN, these layers make
dataset (Brauwers and Frasincar, Apr. 2023). This not only addresses a predictions based on the features reduced and refined by previous
gap in prior research but also provides a real-world solution for identi­ layers. This detailed feature analysis enables the network to categorize
fying fake banknotes in Jordanian financial transactions. The approach photos into distinct groups.
holds promise for greater efficiency and accuracy over traditional
methods. Attention mechanism
Therefore, by referring to previous studies, as shown in Section 2, we
found that there is no study related to the Jordanian currency in all its Neural networks with attention mechanisms can choose to focus on
denominations. We made a major contribution by using a new data set certain areas of a picture while disregarding others. This procedure re­
on Kaggle and used a method of processing images to make fake money lies on human visual attention. Adding an attention mechanism to CNNs
by increasing the brightness of the original images, so that we have Real improves the model’s ability to focus on specific visual regions that are
and fake money. The other contribution was to propose a novel more relevant to the job at hand (Vaswani et al., xxxx). This is especially
approach based on a convolutional neural network (CNN) combined useful for difficult picture classification tasks when the value of char­
with an attention mechanism, which has shown promising results in acteristics varies dramatically between photos. By dynamically
several previous studies (Xue et al., 2021); (Yao et al., Nov. 2019). Our balancing the contributions of distinct input data sections, the attention
approach was also compared with eight pre-trained learning models, mechanism increases the model’s interpretability and ability to make
where our approach was shown to outperform all other algorithms have judgments based on the most informative aspects. This technique has
promising results with a major contribution from the attention mecha­ demonstrated significant increases in model performance across a wide
nism, which has a major role in its ability to focus on important features range of disciplines, including computer vision and natural language
that increase the efficiency of the model and avoid features of little processing.
importance, which saves time and computational cost, and gives
promising results as noticed by (Alkhalili et al., 2021). This research Detection of fake and real banknotes
aims to advance counterfeit banknote detection for Jordanian currency
by developing a novel approach using state-of-the-art techniques. The One of the most difficult challenges facing the banking industry is the
goal is to design an effective system for identifying fake Jordanian detection of counterfeit banknotes, which necessitates extremely accu­
banknotes in order to enhance financial security and prevent fraud rate and dependable technologies. Traditional systems’ manual inspec­
(Alqasrawi, Nov. 2016). The proposed methodology will leverage tion and unsophisticated machine learning algorithms are insufficient to
modern technologies such as deep learning and computer vision to deal with advanced counterfeiting strategies. The usage of CNNs, which
create an optimized model tailored to the specific visual features of valid are frequently augmented with attention mechanisms, represents a sig­
and counterfeit Jordanian notes. By focusing on this under-studied area nificant step forward in this sector. These models can tell the difference
of banknote recognition research and the unique properties of Jordanian between real and counterfeit money by learning to detect the minute
bills, this work strives to make important contributions in anti- patterns and properties that distinguish authentic banknotes. This skill is
counterfeiting and fraud prevention for the nation’s currency. The critical for maintaining the financial system’s integrity and avoiding
specialized model developed through this research could serve as an financial losses (Agasti et al., Nov. 2017).
invaluable tool for law enforcement and financial institutions in iden­ Each currency’s peculiarity, such as the Jordanian dinar offers new
tifying forged banknotes and combating counterfeiting threats. Overall, traits and hurdles to recognition. It may be important to develop unique
this project aims to push boundaries in applying advanced methodolo­ models that can adapt to differences in the security features, patterns,
gies to a novel problem domain, with aspirations to meaningfully and textures of various currencies. This technique, which combines
improve security and mitigate economic risks posed by counterfeiting of attention mechanisms with CNNs, not only improves the accuracy of
the Jordanian Dinar. In this study, the structure of the paper consists of identifying counterfeit currency but also sheds insight on the traits that
the second section related work, which discusses studies related to our are most important in distinguishing between real and fake banknotes.

2
A. Nasayreh et al. Journal of King Saud University - Computer and Information Sciences 36 (2024) 102038

CNNs with attention mechanisms offer the twin advantage of high per­ banknotes recognition, the methodology applied by sequential of CNN
formance and interpretability, making them a good contender for future (custom, AlexNet, VGG11, and VGG16) for the dataset which consist of
research and use in counterfeit detection. 20,800 images of real and counterfeit banknotes. The proposed model
achieved 98.4% accuracy. This pruning model evaluated reduction in
2.2. Literature review parameters by about 75% for reduction in accuracy up to 0.5%. Since the
best performance was when used VGG11 and VGG16 which achieved
In financial institutions, the detection of counterfeits has been a 99.9%. In addition, a deep learning-based method with their self-
significant concern. Over the years, researchers have developed several collected dataset by (Pham et al.,May, 2022) proposed for detecting
techniques to handle this problem. One approach is to use AI techniques counterfeit multinational banknotes using smartphone cameras. They
to analyze the features of multinational banknotes. There are many combine methods of YOLOv3 AND CNN for detection fake and genuine
studies that proposed the application of image processing used in banknotes including US dollar, Europe, Korean Yon, and Jordanian
research for fake detection. Which shows promise of improving results in dinar. Moreover technique by (Pham et al., Feb. 2018) proposed a deep
computer vision, machine learning, and deep learning for fake recog­ learning fitness classification method using the datasets of banknote
nition and detection. image of USD, INR, and KRW. This method can extract and classify
There are several approaches that can be used to successfully features by using image sensors and CNN which captured by visible-light
recognize banknotes, some of the most widely used techniques involve 1-dimintional line to classify fit and unfit banknotes. The results showed
the uses of Machine Learning algorithms, some of these techniques were that on the USD banknote 96.9%, on the INR banknote 99.6% and on the
represented by (Gopane and Kotecha, 2020) which developed a method KRW banknote 97.6%.
for detecting counterfeit Indian banknotes using the RIO cropping A Deep CNN technique proposed by (Azra, Oct. 2020) using smart­
method to separate features, the system uses SVM as a classifier which is phone cameras to capture images of the Ethiopian banknotes in a cross-
trained on several individual features extracted from genuine and dataset environment. This technique includes image acquisition, image
counterfeit banknotes. Their dataset consists of 30 images of Indian size normalization, grayscale conversion and build CNN model to clas­
banknotes. And achieves a highest accuracy of 80% when using SVM. sify and detect the Ethiopian banknotes, they get 97.6% testing accu­
Additionally, the automated method with a dataset of 290 images of racy. Additionally, authors in (Naresh Kumar et al., Dec. 2020) present a
banknotes proposed by (Wang, Sep. 2022) uses Optical Coherence To­ novel approach using Deep Convolution Neural Network for detecting
mography (OCT) and various Machine Learning algorithms (logistic the counterfeit Indian banknotes in their dataset which consist of 306
regression, K-NN, SVM, and random forest) to classify the types of images of Indian banknotes after augmentation. It involves edge
banknotes based on their internal features. SVM classifier achieved the detection, image segmentation, filtering, and pattern matching. Then
best performance with sensitivity 96.55% and specificity 98.85 %. training CNN model on real and fake banknotes and using model to
Another approach (Ashna and Momand, xxxx) proposed a machine classify new currency notes based on learned features. Which achieved
learning models (Random Forest, Naïve bayes, and PART) to detect the highest accuracy at 96.6% by using Random Forest. Moreover (Kamble
Afghan counterfeit banknotes, they are using image processing to et al., Dec. 2019) proposed a Deep CNN model build to identifies the fake
extract texture features from images in the dataset which contains 70 banknotes, this model generated and used a dataset consist of 40,000
images of Afghan banknotes. The highest accuracy achieved by Random images was preprocessed using median-blur filter. The CNN model
Forest at 99%. While (Andrushia et al., 2020) proposed an approach to consists of five convolutional layers, flat layer, and four fully connected
detect counterfeit money by extracting first-order and second-order layers which obtained testing accuracy about 85.6%, validation and
statistical features from their dataset which consist of 120 images of training accuracy were 96.55% and 98.57% respectively. An automatic
Indian banknotes. The feature vectors are given to SVM model for banknotes recognition system proposed by (Tasnim et al., Jan. 2021),
classification. This method produced a classification accuracy of 95.8%. the authors building their unique dataset which consist of 70,452 images
And another techniques by (Ayalew Tessfaw et al., 2018) designed a of Bangladeshi banknotes. They build a deep learning-based CNN model
system that used local binary pattern techniques to extract the features for automatic feature extraction and classification. And achieved 92%
from the images in the dataset which collected by camera or scanner. accuracy.
After feature extracted, they used SVM as classifier to recognize the An AI system for detection fake banknotes by combining Genetic
Ethiopian banknotes and achieved 93% accuracy. algorithm with fuzzy logic by analysis the image and extract the feature.
On the other hand of the banknote recognition, Convolutional Neural Using Fuzzy system (Fuzzy Classification, Genetic Fuzzy Classification,
Networks (CNNs) providing improved accuracy and effectiveness. As it ANFIS Classification, and Genetic ANFIS Classification.) for fraud
is in the (Awad et al., 2022) which proposed a Deep Learning-based banknotes which achieve accuracy rates of 97.64%, 98.60%, 80.83%,
system used a database with 3,961 images of the seven categories of and 97.72%, respectively, where the Genetic Fuzzy Classification had
Iraqi banknotes. Built to classify multi-class of Iraqi banknotes and to the best performance, proposed by (Dirik, 2022; Islam et al., 2021)
help blind people. The system used CNN with 19 layers for accurate present raspberry pi4 and faster R-CNN approach on their dataset which
recognition, the proposed CNN system has expanded to voice pronun­ contains of 4.507 images of Bangladeshi banknotes using to recognize
ciation based on recognized banknotes. The system achieves an accuracy Bangladeshi banknotes. They present a smart blind glass that can
of 98.6% when using Arab and English sides. Furthermore in (Azra, Oct. identify currency and send audio data to the blind man’s ears. The
2020) that used and compared the CNN algorithms (AlexNet, Google­ recognition accuracy is 97.8%.
Net, and VGG16) with transfer learning to detect counterfeits. The Most prior works utilize relatively small datasets as shown in Table 1,
researcher used a pre-trained model on a small dataset consisting of 362 ranging just in the hundreds or thousands of banknote images. Our study
banknote images which captured under visible light, they fine-tuned the aims to leverage a larger and more diverse dataset encompassing various
models for improving the accuracy. VGG16 achieved the highest accu­ international currency denominations, posing a greater challenge for
racy at 99.8%. And a study by (Mittal and Mittal, 2018) proposed an model generalization. Furthermore, many studies have focused classi­
automated method for recognizing fake banknotes by using dataset with fication efforts on only a single currency type, such as Indian rupees or
380 visible-light image that captured by smartphone camera. This model Bangladeshi takas. In contrast, our work seeks to classify multiple global
used CNN for classifying counterfeit and genuine banknotes, for currencies simultaneously, requiring the model to discern more fine-
enhanced accuracy over multiple epochs, a new SoftMax layer was grained visual differences between banknote designs.
added to the convolutional foundation of a pre-trained MobileNet. Additionally, while some studies employ only basic preprocessing
Which achieved accuracy on test set of 96.6%. Another research by like cropping, resizing or noise removal, more advanced augmentation
(Pachón et al., Dec. 2023) introduce a pruning technique for fake and normalization techniques could prove beneficial for improving

3
A. Nasayreh et al. Journal of King Saud University - Computer and Information Sciences 36 (2024) 102038

Table 1
Comparision with other studies closely to our work.
Author Dataset Preprocessing Method Highest Result

(Gopane and Kotecha, 30 Indian banknotes ROI Cropping LR, KNN, RF, SVM SVM: 80 %
2020) Feature Set Extraction
(Wang, Sep. 2022) 290 of YAN banknotes Image processing + LR, SVM, KNN, RF SVM
Feature extraction Sensitivity: 96.55 %
specificity: 98.85 %
(Awad et al., 2022) 3,961 of Iraqi banknotes Image Enhancement + Re-sizing
CNN 98.6 %
(Azra, Oct. 2020) 362 of of Bosnia and Herzegovina banknotes Data Augmentation AlexNet, Google Net, VGG16 VGG16: 99.8 %.

(Mittal and Mittal, 380 Indian banknotes − preprocessing and augmentation CNN 96.6 %
2018)
(Pham et al.,May, self-collected dataset of USD, Euro, won, and − preprocessing using Retinex CNN (ResNet-18, 99 %
2022) JOD banknotes. filtering. GoogleNet, and
− Banknote region detection using Inception-ResNet-v2)
YOLOv3.
− Banknote region cropping.
(Pham et al., Feb. After augmentation the samples in the dataset − Segmentation banknotes region. CNN USD banknote: 96.9
2018) was approximately 100,000 images of USD, − size normalization. %
INR, and KRW banknotes. INR banknote: 99.6
%.
KRW banknote: 97.6
%.

(Pachón et al., Dec. 20,800 images of real and counterfeit − extract features and patterns AlexNet, VGG11, VGG16 VGG11 and VGG16:
2023) banknotes. 99.9 %
(Dirik, 2022) 1,372 images of banknotes Fuzzy Classification, Genetic Fuzzy Genetic Fuzzy
− Analysis the image Classification, ANFIS Classification, and Classification: 98.60
− Extract the feature Genetic ANFIS Classification %
(Ashna and Momand, 70 images of AFN banknote − Image preprocessing RF, NB, PART RF: 99 %
xxxx) (Normalization and Resampling).
− Feature extraction.
306 images of Indian banknotes after − Image processing (smoothing Deep CNN 96.6 %
(Naresh Kumar augmentation image and removing noise). for classification.
et al., Dec. 2020) − Edge detection and
segmentation.
− feature extraction.

(Tasnim et al., Jan. 70,452 images of Bangladeshi banknotes − dataset building and CNN architecture includes automatic 92 %
2021) preprocessing (Resize and feature extraction and classification
Reshape).

(Ayalew Tessfaw 500 images of Ethiopian banknote − Image processing (image SVM 93 %
et al., 2018) smoothing and noise removal)-
feature extraction
(Islam et al., 2021) 4,507 images of Bangladeshi banknotes − ROI cropping Faster R-CNNinception V2 coco-2018 97.8 %
− normalizing model
(Andrushia et al., 12o images of Indian banknotes. − Image preprocessing. SVM as classifier 95.8 %
2020) − features extraction (color,
texture, and shape).features
vector.
(Gebremeskel et al., 6100 images of Ethiopian banknote. − image preprocessing (size a Deep CNN for classification 97.6 %
preprint, Dec. normalization and RBG to gray
2022.) conversion)
− feature extraction.
− dimensionality reduction.

classification performance. The majority of prior art also relies solely on challenges in detail. Our work will provide more thorough analyses of
standard deep CNN architectures like AlexNet or VGG rather than difficulties faced in multi-currency banknote classification, as well as
exploring potentially superior ensemble or hybrid models. There is op­ leverage more modern techniques like transformers and self-supervised
portunity for our study to examine more complex models and optimize learning that are scarcely adopted presently in this domain.
hyperparameters and cross-validation systematically to boost model
generalizability. 3. Methodology
Though some existing works report very high accuracy, this perfor­
mance is often limited to small datasets and may not translate to real- In this section, the architecture of the proposed CNN network is
world conditions with larger, more diverse banknote populations and described. The evaluation was conducted using a dataset containing
wear. Our work will aim to benchmark model performance across cur­ images of Jordanian currency banknotes, where the contribution is to
rencies and on realistic large datasets. Another limitation we intend to detect counterfeit banknotes and classify them from real banknotes. The
address is the lack of research on joint banknote verification and value main goal is to propose a CNN neural network structure and introduce
recognition, with most studies targeting only a single task. Our models attention mechanisms that in turn contribute to increasing the perfor­
will aim to concurrently classify both authenticity and denomination. mance and efficiency of the model. It was compared with six pre-trained
Finally, a few studies discuss model limitations, failure cases, and models that include VGG16, VGG, ConvNeXtBase, DenseNet121,

4
A. Nasayreh et al. Journal of King Saud University - Computer and Information Sciences 36 (2024) 102038

InceptionResNetV2, and ResNet50 as shown in Fig. 1. to machine learning models. One of the most popular methods to in­
crease the number of images and to prevent overfitting is to use the
augmentation method to increase the number of images through several
3.1. Image collection and preprocessing
types, including: (horizontal flip, vertical flip, shear range = 0.2, rota­
tion range = 45, zoom range 0.2), augmentation was done for the
3.1.1. Image acquisition
training data set, where the images were increased from 914 to 5454 to
The Jordanian paper currency denomination data set was obtained
ensure the high efficiency of the proposed mode, Fig. 3 show Sample of
from Kaggle (Jordanian Banknotes, 2023), as it contains all paper de­
augmented images.
nominations of Jordanian currencies, including five different de­
nominations (1 dinar, 5 dinars, 10 dinars, 20 dinars, and 50 dinars). The
3.2. Proposed CNN– chanel attention model (CNN-CAM)
quality of the images is considered high quality, as it was photographed
in appropriate lighting conditions, as it was photographed from different
In this paper, we investigate how Convolutional Neural Networks
angles to ensure that features were fully extracted via machine learning
(CNNs) can be used to distinguish between real and counterfeit money
and deep learning models. The number of pictures was distributed as
notes. The first phase in our method is to apply the first convolutional
follows: (135 pictures for the 1-dinar denomination, 80 for the 5-dinar
layers, which are responsible for extracting a variety of information
denomination, 80 for the 10-dinar denomination, 100 for the 20-dinar
from the note images. This first preliminary stage is required for the
denomination, and 159 for the 50-dinar denomination), based on the
subsequent integration of a channel attention mechanism. This mecha­
main contribution of this study in discovering counterfeit money from
nism critically assesses each channel’s output from the convolutional
Real money. We created counterfeit money from real money by
processes to identify which channels have the most relevant data for our
increasing the contrast of each image using the CV2 library as shown in
classification task. It expertly guides the network’s analytical focus,
Fig. 2, and we now have a group of counterfeit money with the same
giving these critical pathways processing priority. After channel atten­
number as the real one, so that it becomes new data containing real and
tion is applied, we present a spatial attention layer. The purpose of this
counterfeit money.
second stage is to examine the feature maps which have been further
refined by channel prioritization to identify image areas that are critical
3.1.2. Augmentation
for confirming the legitimacy of a message. This layer highlights key
Especially since image processing is a necessary task that has been
features like watermarks, holograms, and fine print that are crucial to
used by many studies in the field of image processing (Perez and Wang,
the identification process by creating a spatial attention map. This kind
2017), as it is necessary to process the images and prepare them before
of focused emphasis guarantees that the CNN’s subsequent phases are
entering them into the automated model. The size of the images was
focused on examining these richer, prominent areas of the picture.
standardized to (224,224) to be suitable for the various deep learning
models used. It is known that Deep learning models, especially con­
volutional neural networks (CNN), need a larger data volume compared

Fig. 1. Flow Chart of the Proposed Approach.

5
A. Nasayreh et al. Journal of King Saud University - Computer and Information Sciences 36 (2024) 102038

Fig. 2. Sample of Image created.

Fig. 3. Sample of augmented images.

3.2.1. Architecture of CNN model entered with the shape [224, 224, 3]. The image is inserted and fed into
The design of our neural network plays an important role in the first convolutional layer. The filters are applied using 16 filters, as
achieving promising performance in classification. In this section, we these filters are concerned with extracting important features from the
provide a detailed description of the structure of the convolutional image, and therefore the outputs of this layer contain maps of the pa­
neural network (CNN). In the structure of the CNN, images are initially rameters, where each map corresponds to the feature extracted from the

6
A. Nasayreh et al. Journal of King Saud University - Computer and Information Sciences 36 (2024) 102038

original image. The size of these features is 3x3, and the best was also
chosen. activation function is randomly selected between ReLU (Recti­
fied Linear Unit) and LeakyReLU (Leaky Rectified Linear Unit), and it is
known according to the study (Eckle and Schmidt-Hieber, Feb. 2019)
that it is responsible for reducing the vanishing gradients in the network
during training and converting any negative value to zero. As for Lea­
kyReLU, it is Responsible for reducing dead neurons by multiplying
negative values by alpha, where alpha is a small positive constant,
regarding max pooling, which contributes to reducing the spatial di­
mensions of the input feature maps while preserving important features
with stride (2,2), as it moves by two from the right. And two down for
each step.
After that, it is passed to the flat layer through a dense layer to
perform a linear transformation of the flat inputs. This transformation
contributes to a deeper understanding of the complex patterns within
the data, where each weight is multiplied by each of the corresponding
inputs, and then the result is summed, and space is saved by introducing
bias. Equation (1) shows how to calculate the output vector Y by
multiplying the input vector x with the weight matrix W, and the bias
vector b.
y = wx + b (1)

Nonlinearity is introduced to the output layer and any real number is

converted to a value between 0 and 1 via the sigmoid activation function
commonly used for binary classification, as shown in the equation.
1
ν(x) = x
(2)
1 + e−

3.2.2. The convolutional Block attention Module (CBAM)

The Convolutional Block Attention Module (CBAM) was used within
the CNN structure, where, through two important dimensions, channel Fig. 5. Channel Attention Module.
and spatial attention maps are extracted and then combined via element-
wise multiplication with the input feature map (Zhang et al., 2022). ([ ])
Mc (g) = β(k3×3 gsmean′ gsmax ) (4)
Fig. 4 shows the CBAM structure, which consists of the channel attention
unit responsible for measuring the results of the convolutional layer,
β represent to the sigmoid activation function , andk3×3 represent to the
after which the final result is obtained through processing via the spatial
kernel size of the convolution.
attention unit (Ganguly et al., preprint, Aug. 2023.).
Fig. 5 shows the mechanism by which the channel attention unit
3.3. Transfer learning approaches
works to output a one-dimensional vector by condensing the feature
map, which is done by compressing the input feature maps using Max
In this study, we employed the concept of transfer learning a well-
Pooling and Mean Pooling. Then the compressed feature maps are
known realm in Deep Learning. It has such a great reputation and
collected and transferred to the fully connected layer and subjected to
attention in research since it has the ability to transfer already-gained
the sigmoid activation function, which results in the normalized channel
knowledge, allowing researchers not to worry about the need for a
attention weights denoted by Mc (g) which are mathematically repre­
massive dataset and expensive learning time (Alzubaidi, Jan. 2020). In
sented by the equation (3):
( ( )) ( ( )) the context of our work, with the intention of potentially enhancing the
Mc (g) = β(w1 w0 gcmean + w1 w0 gcmax ) (3) performance of our problem, we investigated the use of transfer learning
by using six different pre-trained models. Those models are explored in
β represent the sigmoid activation function, g represents the input more detail in this section.
feature map w1 represents the weights of the fully connected layer, and
w0 represents the hidden layer. 3.3.1. VGG16 VGG19 models
Fig. 6 shows the spatial attention module. It also contains Mean VGG16 was first proposed in 2014 by the Visual Geometry Group
Pooling and Max Pooling through which it is done, which contribute to (VGG) from the University of Oxford, hence the name VGG16 (Simonyan
merging the previously extracted feature maps that contain one channel and Zisserman, 2015) their pre-trained convolutional neural network
to then become a feature map with two channels. The weight of the consists of 16 layers trained on the ImageNet dataset (Deng et al., 2009)
spatial attention Ms (g) was obtained through the equation (4): which has 14 million training images. Hence, it’s considered relatively

Fig. 4. CBAM structure.

7
A. Nasayreh et al. Journal of King Saud University - Computer and Information Sciences 36 (2024) 102038

larger receptive fields (like 5x5 or 7x7) but with fewer parameters and
less computational complexity (A Transfer Learning Strip Steel Surface,
xxxx).

3.3.2. InceptionResNetV2
InceptionResNet adopts the Network in Network (NIN) (Lin et al.,
2014) architecture this design principle aims to improve the power of
each individual conv layer by applying more complex operations
without compromising the overall reduction of model complexity, like
1x1 convolutions, within each layer. Hence, it helped the feature
learning and representation capabilities of InceptionResNet (Liu, et al.,
2021). The Residual Block design helps transmit the features between
different layers both forward and backward, therefore speeding up the
parameter optimization and network training. In this paper, we used
two-layer learning units incorporating the residual network. To reduce
the dimensions, we first applied a 1x1 convolution, afterward a 3x3
convolution operation was applied.

3.3.3. DenseNet
This network was proposed embracing the fact that Convolutional
Networks with shorter connections between layers close to the input and
those close to the output can be significantly deeper, more efficient to
train, and more accurate. The Dense Conv Net is a feed-forward network
that connects each layer to every other layer. The feature maps learned
by each layer are fed as input to the next layer. DenseNets offer various
benefits: they solve the Vanishing gradient problem, improve feature
Fig. 6. Spatial Attention Module. propagation, and mainly reduce the number of hyperparameters (Lin
et al., 2014). A Densely Connected Network consists of three main
extensive with a total number of parameters of 138 million, thus it components which are the Dense Block, a transition layer, and the
gained its competitive performance capabilities when it is used on bottleneck connection. Fig. 9 shows the dense block components, the
different image classification problems (Zhang et al., 2022). VGG16 number of layers this block has depends on the vision of DenseNet you
outperforms AlexNet (Alom et al., xxxx) by replacing big filters with a are using in our work we have used DenseNet201 and DenseNet121 both
sequence of smaller 3 X 3 filters, such that our data was reshaped into having the same composite function which includes Batch Normaliza­
224 X 224 images and fed to Vgg16 and Vgg19 as well, a sigmoid tion (BN), Rectified Linear Unit Function (ReLU), and Convolution op­
function was used on the last dense layer as the activation function to be erations. Each layer in this network has “k” inputs, implying that the
compatible with our binary classification problem as we want to detect “kth” layer receives outputs from all preceding layers resulting in a
whether the input currency image is Real or Fake Fig. 7 illustrates the highly interconnected structure (He et al., 2016).
workflow of VGG16 implementation. Another version of VGG16 that DenseNet201 and DenseNet121 have the same architecture and
was developed with 19 weighted layers (Gharaibeh, Sep. 2023), there­ workflow, the only differentiating factor is the number of layers each
fore it was called VGG19. The main idea of both these networks is the has. DenseNet201 has a total number of 201 layers while DenseNet121
same: using many layers with small filters can emulate the effect of has only 121 layers.

Fig. 7. VGG16 Implementation & Structure Design.

8
A. Nasayreh et al. Journal of King Saud University - Computer and Information Sciences 36 (2024) 102038

3.3.4. ResNet50 Unit Dense with sigmoid activation function for determining whether an
ResNet50 stands for Residual Network, it’s a specific type of con­ image is Real or Fake.
volutional neural network, it was first introduced in 2015 (He et al.,
2016), as its name indicates it consists of 50 layers (48 convolutional
3.4. Performance evaluation
layers, one Max Pool layer, and 1 Average Pool), ResNet50 uses a
bottleneck design for the building block as shown in Fig. 8, this design
Among the most famous measurements in the field of machine
helped to reduce the number of parameters and matrix multiplications,
learning: accuracy, precision, recall, and f1-score, where, according to
enabling a much faster training in each layer.
the equation (5), accuracy expresses the ratio of correct predictions
A renowned version of ResNet50 called ResNet50V2 is designed to
relative to all predictions. In our case, refer to the proportion of images
enhance performance and reduce training time. It consisted of archi­
that are correctly classified as fake or real money. In the equation (6),
tectural improvements, incorporates bottleneck blocks, making the
precision refers to the ratio of the prediction of positive cases to the ratio
overall design computationally effective. In the field of convolutional
of all positive cases expectation is positive, and according to the current
neural networks ResNetV2 proven its significant enhancement contrib­
study, it indicates to the proportion of images expected to be fake money
uting as a top tier performance in many image classifications related
that are fake.The equation (7) refers to recall, which expresses the
tasks (Rasheed, Dec. 2020).
percentage of actual cases of fake money that were predicted correctly.
The f1-score in the equation (8) indicates the average of precision and
3.3.5. ConvNext
recall.
ConvNeXt integrates the strengths of transformer architectures into
its network structure and parameter selection. Through modification of Tp + TN
Accuracy = (5)
the training parameters, optimizers, and convolution kernel sizes, it has T p + T N + Fp + FN
even surpassed the well-known and highly rated transformer model
known as Swin-Transformer (The Impact of Fatigue, 2023) on the Precision =
Tp
(6)
ImageNet-1K dataset. ConvNeXt has a modified structure compared to T p + Fp
the original ResNet which shows its superiority in the overall perfor­
mance. The similarity in both ConvNext and ResNet50 structure-wise Tp
Recall = (7)
becomes clearly observable in the head’s feature extraction layer, a T P + FN
middle layer with separate stacking of bottleneck structures of four
precision × Recall
different dimensions, and the final high-dimensional feature classifica­ F1 − Score = 2 × (8)
Precision + Rrecall
tion layer. on the other hand, the differences in the inner working
mechanisms of each layer and the strategy of stacking are shown in
4. Results and discussion
Fig. 9.
In our work and to adopt the transfer learning task we employed all
To evaluate the proposed method for detecting and classifying
these pre-trained models. The previously mentioned models with their
counterfeit Jordanian currency notes, we implemented the CNN-CAM
layers, including Conv layers were loaded with the ImageNet pre-trained
approach and evaluated it on the collected data set, and it was
weights. With the aim of preserving their original hyperparameters, we
compared with eight pre-trained deep learning models that include
froze the pre-trained layers in each model, Table 2 demonstrates those
VGG16, VGG19, ResNet50V2, DenseNet121, DenseNet201, and Convex-
hyperparameters in each of the implemented pre-trained models.
Tiny, Inceptionresnetv2, and CNN.We implemented the proposed
Moreover, a custom classification header was added so that each model
approach on the environment in which the evaluation was made for the
architecture properly fit our classification problem, this header con­
proposed approach is explained, where all experiments for this study
sisted of a GlobalAveragePooling2D layer for feature aggregation, pur­
were conducted on a laptop using a Python tool on a computer with an
sued by a 128 units’ Dense layer with ReLU activation, and lastly, a 1
i5-10H CPU, NVIDIA GeForce GTX 1660Ti 4 GB, 8 GB RAM, 250 GB SSD

Fig. 8. DenseNet Architecture Design.

9
A. Nasayreh et al. Journal of King Saud University - Computer and Information Sciences 36 (2024) 102038

Fig. 9. (a) ConvNeXT block Design and (b) ResNet Block Design.

Table 2
Pre-Trained Models Hyperparameters.
VGG16 ResNet50 DenseNet- DenseNet- Inception- ConvNeX- VGG19
121 201 ResNetV2 Tiny

Input 256x256x3 256x256x3 224x224x3 224x224x3 299x299x3 224x224x3 224x224x3

layer
Conv 13 4 blocks 121 dense blocks, 201 dense blocks, The model consists of 4 blocks 4 blocks with several 16
Layers with several conv transition layers, and a transition layers, and a with multiple combinations of conv layers organized
layers organized final classification layer final classification layer conv layers with varying as follows (3, 3, 9, 3)
as follows filters
(3,3,3,3)

Dense 3 1 1 1 1 3
Layers
Dropout 0.3 0.3 0.3 0.3 0.3 0.3 0.3
optimizer Adam Adam Adam Adam Adam Adam Adam
Epochs 30 30 30 30 30 30 30
Learning 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001
Rate

and 1 TB HDD, where the data set was divided into 60% for training,
20% for verification and 20% for testing. Table 3
The Performance of Models.

4.1. Result analysis Model Accuracy Precision Recall F1-Score

VGG16 90 93.7 87.2 90.3

After obtaining poor results from some pre-trained deep learning VGG19 81.2 85.7 76.7 80.9
models in the problem of detecting and classifying counterfeit money, ResNet50V2 78.7 78.6 81.3 80
DenseNet121 90.3 90.6 90.6 90.6
we decided to propose an approach based on CNN with attention DenseNet121 88.4 83 97.6 89.8
mechanism, as it turned out that the proposed approach demonstrated Convex-Tiny 78 77.7 81 79.5
promising results compared to other models. Table 3 shows the perfor­ Inceptionresnetv2 81 77.7 89.5 83
mance of the nine models on the Jordanian currency data set, as it turns CNN 95 95.4 95.2 95.3
CNN-CAM 96 96.6 96.4 96.5
out that the CNN-CAM model outperformed the other models with a
performance of 96% for accuracy, precision, recall, and f1-score. This
result is considered the highest among other models, as the CNN model
is also considered to have achieved results close to the proposed model
with 95% accuracy.

10
A. Nasayreh et al. Journal of King Saud University - Computer and Information Sciences 36 (2024) 102038

The attention mechanism is considered an essential factor in per­

formance, as can be seen in Table 3, where the improvement was
observed on the CNN by increasing the accuracy after adding the
attention mechanism from 95% to 96%. Some pre-trained models ach­
ieved good results, such as the two models, DenseNet121 and the VGG16
model, with an accuracy of 90.3% and 90%, respectively. The two worst
models were Convex-Tiny and ResNet50V2, with an accuracy of 78%
and 78.8%, respectively.
The chart in Fig. 10, shows the performance of all the models used in
our study, as it shows the strength of our CNN-CAM model with its su­
periority over the rest of the other models with an accuracy of 96%,
precision 95.6%, recall 95.5% and F1-Score 95.5. There is no doubt that
the clear superiority of the model The proposal is due to the modification
of the CNN by integrating it with the CAM, which focuses on the
important features and achieves high results, in addition to also
choosing the most optimal hyperparameters through experience. The
CNN model comes with an accuracy of 95%, which is considered a
promising result compared to other models, and there is no doubt about
the VGG16 and DenseNet121 model as well, achieving results of 90%
and 90.3%, respectively. Due to the difficulty of our images and the
complex features, the other models did not achieve satisfactory results. Fig. 11. confusion Matrix of CNN-CAM.
As shown in the Fig. 10, the results for the pre-trained models are
considered unsatisfactory. as shown in Fig. 13, where the performance was very good, as training
The errors predicted by the proposed CNN-CAM model is presented. and validation accuracy increases between epochs. In contrast, the loss
Fig. 11 shows the confusion matrix, which provides an overview of all values decrease, which indicates that the model learns well by classi­
correct and incorrect predictions, where the vertical axis represents the fying real and counterfeit money, as training stopped at epoch 25 based
true contacts and the horizontal axis represents the expected labels, on stop learning, due to the apparent instability in the period from 20 to
where, as shown in Fig. 11, there are six errors. Incorrect prediction 25, so it was stopped at the best performance, which is approximately
between real and fake money is considered a very small percentage, and 96%.
this shows the ability of our model to make the correct prediction be­ To support the superiority of the proposed model, the model was
tween fake and real money. The integration between CNN and the validated on five samples from outside the data that were not included in
attention mechanism is considered a fundamental contribution to the training or testing process. These samples consist of three real im­
obtaining promising results. ages and three fake images, and the model succeeded in classifying these
This Fig. 12 indicates the ROC curve. We can see that the perfor­ samples correctly. Hence, we see the success of the proposed model in
mance is distinguished by the high AUC on the ROC curve, where high real scenarios. Fig. 14 shows samples that have been verified as
accuracy expresses low prediction error rates in the classification be­ counterfeit.
tween real and fake money.
It can be better understood if the proposed model fits the data. This is
done by drawing learning curves that include training and validation
loss, training, and validation accuracy over the 30 epochs in the data set

Fig. 10. Chart for performance Deep Learning Models.

11
A. Nasayreh et al. Journal of King Saud University - Computer and Information Sciences 36 (2024) 102038

3. Increasing Detection Accuracy: By improving the precision and

dependability of counterfeit currency detection, the proposed
method decreases the potential of financial losses caused by false
positives.
4. The technology significantly reduces the time and resources required
for human checks by automating the detection process and enabling
real-time analysis, hence increasing overall efficiency.
5. Improved detection capabilities aid regulatory and law enforcement
efforts to reduce the availability of counterfeit money, strengthening
financial and legal institutions in the process.
6. Advancing Technological Innovation: By demonstrating the effective
application of complicated models in detecting counterfeit currency,
this work not only addresses a critical financial issue but also ad­
vances the disciplines of artificial intelligence and machine learning.

4.3. Comparison and discussion

We have made a comparison between our work and other proposed

Fig. 12. ROC Curve of CNN-CAM. models as shown in Table 4. Our proposed model may not have the
highest accuracy among other studies, but it is important to emphasize
4.2. Significant of this study that the accuracy rate we achieved is largely accurate and reliable in
classifying our specific data. The approach is considered novel and can
1. Dealing with the Challenges of Counterfeit Currency: Counterfeit be tested on a wider set of data if it is obtained. The superiority of our
currency undermines financial stability and encourages criminal approach over previous studies is that we are the first study conducted
conduct, which has a significant impact on the global economy. The on Jordanian money with a comprehensive qualitative comparison with
growing sophistication of counterfeiting schemes frequently renders pre-trained deep learning models, as our proposed model is considered
existing detection methods ineffective, emphasizing the need for the first model that works on classifying money in all countries, as the
more intelligent, precise, and timely detection systems. addition of attention mechanism is considered a fundamental addition
2. Using Attention Mechanisms and CNN: A novel technique for picture to the CNN model because of its It is important to focus on the most
recognition for currency identification combines Convolutional important features and avoid weak features, which saves time and
Neural Networks (CNNs) with an attention mechanism. This com­ computational cost during training.
bination enables the exact identification of the most revealing in­ In the branch of banknote classification research, numerous models
dicators of authenticity or counterfeiting, with a focus on critical have been developed using diverse datasets, methodologies, and pre­
aspects to improve detection accuracy and dependability.’ processing techniques as shown in Table 4. These models aim to accu­
rately classify different world currencies, with achieved accuracy rates

Fig. 13. Training and Validation (Accuracy and Loss).

Fig. 14. validating of our model on unseen data.

12
A. Nasayreh et al. Journal of King Saud University - Computer and Information Sciences 36 (2024) 102038

Table 4
The Comparision with Relate Work.
Author Dataset Preprocessing Method Highest Result

(Gopane and 30 Indian banknotes ROI Cropping LR, KNN, RF, SVM
Kotecha, 2020) Feature Set Extraction SVM: 80 %
(Wang, Sep. 290 of YAN banknotes Image processing + LR, SVM, KNN, RF SVM
2022) Feature extraction Sensitivity:
96.55 %
specificity:
98.85 %
(Awad et al., 3,961 of Iraqi banknotes Image Enhancement + Re-sizing
2022) CNN 98.6 %
(Azra, Oct. 2020) 362 of of Bosnia and Herzegovina banknotes Data Augmentation AlexNet, Google Net, VGG16
VGG16: 99.8 %.

(Mittal and 380 Indian banknotes − preprocessing and augmentation CNN 96.6 %
Mittal, 2018)
(Pham et al., self-collected dataset of USD, Euro, won, and JOD − preprocessing using Retinex filtering. CNN (ResNet-18, 99 %
May, 2022) banknotes. − Banknote region detection using GoogleNet, and
YOLOv3. Inception-ResNet-v2)
− Banknote region cropping.
(Pham et al., Feb. After augmentation the samples in the dataset was − Segmentation banknotes region. CNN USD banknote:
2018) approximately 100,000 images of USD, INR, and KRW − size normalization. 96.9 %
banknotes. INR banknote:
99.6 %.
KRW banknote:
97.6 %.

(Pachón et al., 20,800 images of real and counterfeit banknotes. − extract features and patterns AlexNet, VGG11, VGG16 VGG11, VGG16:
Dec. 2023) 99.9 %

(Tasnim et al., 70,452 images of Bangladeshi banknotes − dataset building and preprocessing CNN architecture includes 92 %
Jan. 2021) (Resize and Reshape). automatic feature extraction and
classification
(Ayalew Tessfaw 500 images of Ethiopian banknote − Image processing (image smoothing SVM 93 %
et al., 2018) and noise removal)- feature extraction
Jordanian Dinar Banknotes Data Augmentation
Our Model CNN-CAM 96 %

ranging from 80% to as high as 99.9%. Our research focuses specifically 10 0, 20 0, 50 0, and 20 0 0 rupees. The dataset of banknote images was
on developing a model to classify Jordanian Dinar banknotes. Like many gathered in a range of lighting conditions, as well as busy and gloomy
other banknote classification models, our approach utilizes Convolu­ backdrops. Images of partially folded or concealed banknotes are also
tional Neural Networks (CNNs), a powerful deep learning technique taken. We are the first research to conduct an experiment on this data
well-suited for image classification tasks. Additionally, we employ data set, and we achieved an accuracy of 88%, which is considered good
augmentation techniques to expand our dataset of Jordanian Dinar accuracy on a selection data set.
images. This involves artificially manipulating the images through ro­ Our research’s specialized approach focusing solely on Jordanian
tations, shifts, and other transformations to generate more training data, Dinar banknotes sets it apart. Combined with strong technical imple­
enhancing the model’s ability to generalize. mentations like CNN architectures and tailored data augmentation, our
However, our research is distinguished by its specialized focus on model delivers state-of-the-art accuracy in classifying this specific cur­
Jordanian dinar banknotes. By focusing specifically on this currency, we rency. The research excels in developing highly customized, currency-
can perform more detailed analysis and tuning to identify the complex specific representations needed for precise Jordanian Dinar banknote
visual features unique to the Jordanian Dinar banknote. This allows the classification.
model to learn the highly specialized representations needed for accu­ This research has demonstrated promising classification accuracy on
rate classification. Our model achieves an accuracy rate of up to 96% in Jordanian Dinar banknotes using a specialized deep learning model.
classifying Jordanian dinar banknotes into their respective de­ While these results are encouraging, there remain opportunities to build
nominations. This places our model among the best performing models on this foundation through several avenues of future work. One valuable
in JD classification research, demonstrating a high level of efficiency. extension would be growing the scope of the model to encompass
Furthermore, the data augmentation techniques we use are custom and multiple global currencies. Retraining on a diverse dataset of interna­
specific to JOD images. This contributes to improving the training data, tional banknotes could exhibit the model’s capabilities beyond Jorda­
which enables our model to effectively recognize the subtle differences nian Dinar classification. Furthermore, expanding the dataset size and
between the different denominations of the Jordanian dinar. Our model variability by collecting more banknote images under differing condi­
was also tested on two other data sets, the first of which contains 7280 tions would enable the model to become more robust. Another area to
images (Pachón et al., Dec. 2023). The dataset is organized into three explore is implementing more advanced convolutional neural network
volumes: 60% for training, 20% for validation, and 20% for testing. To architectures. Experimenting with cutting-edge deep learning tech­
use triple cross-validation, three datasets are used in three different niques could potentially push classification accuracy even higher.
ways: DS1, DS2, and DS3. An accuracy of 99.9% was achieved, which Additional worthwhile directions include integrating serial number
demonstrated the strength of our proposed approach. An India bank­ recognition for authentication, evaluating performance on damaged
notes dataset was also used in our dataset (Meshram et al., Apr. 2022). banknotes to showcase real-world viability, and developing practical
Ten distinct varieties of Indian banknotes are included in it: 10, 20, 50, counting and validation systems deployable on portable devices. Finally,

13
A. Nasayreh et al. Journal of King Saud University - Computer and Information Sciences 36 (2024) 102038

publicly releasing the banknote dataset and model details would pro­ merging it with the CAM, which selects the most ideal hyperparameters
mote further innovation. Allowing the research community full access through experience and concentrates on the critical components to
can lead to new ideas for enhancing currency classification and recog­ produce excellent outcomes. In future work, we intend to refine our
nition systems. By broadening the model’s applicability, generaliz­ methodology and apply it to a dataset of polymer banknotes, with the
ability, and accessibility, future work can build upon these promising goal of addressing the unique issues of counterfeit identification in these
beginnings. This provides an exciting roadmap for evolving the materials. We will investigate the application of advanced machine
specialized capabilities of the model to make an impact in real-world learning algorithms to improve accuracy and adaptability. Partnerships
applications. with financial institutions will also be sought to guarantee that our
strategy is applicable and meets real-world criteria.
4.4. Limitations Compliance with ethical standards
Ethical approval: This article does not contain any studies with
To have a complete understanding of the study’s breadth and human participants or animals performed by any of the authors.
applicability, the limitations of “Jordanian Banknote Data Recognition: Informed consent: Informed consent was obtained from all indi­
A CNN-Based Approach with Attention Mechanism” must be addressed. vidual participants included in the study.
One notable limitation is the specificity and diversity of the data used; Funding: Not Applicable.
our dataset, which primarily consists of banknotes from specific series or
eras, may not accurately capture the range of conditions encountered in
real-world settings, such as varying degrees of wear and tear or different Declaration of competing interest
lighting and angle conditions. This feature may limit the model’s resil­
ience and generalizability. The authors declare that they have no known competing financial
Furthermore, the computational complexity of modern CNN archi­ interests or personal relationships that could have appeared to influence
tectures, when combined with attention mechanisms, may limit imple­ the work reported in this paper.
mentation in real-time applications or on devices with low processing
resources. While channel and spatial attention methods refine the References
model’s emphasis on key features, they may accidentally bias the model
A Transfer Learning Strip Steel Surface Defect Recognition Network Based on VGG19 |
toward specific banknote qualities, thereby omitting subtle but impor­
SpringerLink.” Accessed: Oct. 20, 2023. [Online]. Available: https://link.springer.
tant traits. This bias may impair the model’s performance in edge com/chapter/10.1007/978-981-15-2341-0_41.
scenarios. Agasti, T., Burand, G., Wade, P., Chitra, P., Nov. 2017. Fake currency detection using
image processing. IOP Conf. Ser. Mater. Sci. Eng. 263, 052047 https://doi.org/
Additionally, because different banknotes have different security
10.1088/1757-899X/263/5/052047.
characteristics, our methodology—which was developed specifically for Albawi, S., Bayat, O., Al-Azawi, S., Ucan, O.N., Oct. 2018. Social Touch Gesture
Jordanian banknotes—might not be immediately applicable to other Recognition Using Convolutional Neural Network. Comput. Intell. Neurosci. 2018,
currencies without requiring major modification. Another problem is 1–10. https://doi.org/10.1155/2018/6973103.
M. AL-Gawda, Z. Beiji, and N. Mohammed, “Yemeni Mobile Counterfeit Detection
the way counterfeit technology is developing; as counterfeiters grow System Using Support Vector Machines, Fuzzy Logic and Image Processing
more skilled, our model might need to be updated frequently to stay Techniques,” J. Comput. Theor. Nanosci., vol. 13, no. 5, pp. 2965–2977, May 2016,
relevant. doi: 10.1166/jctn.2016.4945.
Ali, T., Jan, S., Alkhodre, A., Nauman, M., Amin, M., Siddiqui, M.S., Sep. 2019.
Finally, the range of data collecting and the application regions of the DeepMoney: counterfeit money detection using generative adversarial networks.
model may be constrained by ethical and privacy considerations around PeerJ Comput. Sci. 5, e216.
the gathering and use of banknote data, particularly when addressing Alkhalili, M., Qutqut, M.H., Almasalha, F., 2021. Investigation of Applying Machine
Learning for Watch-List Filtering in Anti-Money Laundering. IEEE Access 9,
counterfeits. By being aware of these limitations, we not only identify 18481–18496. https://doi.org/10.1109/ACCESS.2021.3052313.
the limitations of the work that we have done so far, but we also point Z. Alom, T. M. Taha, C. Yakopcic, S. Westberg, P. Sidike, and M. S. Nasrin, “The History
out directions that future research should take and develop. This will Began from AlexNet: A Comprehensive Survey on Deep Learning Approaches”.
Alqasrawi, Y., Nov. 2016. Natural Scene Image Annotation Using Local Semantic
increase the validity and applicability of our findings in the larger
Concepts and Spatial Bag of Visual Words. Int. J. Sens. Wirel. Commun. Control 6
context of document verification and counterfeit detection. (3), 153–173. https://doi.org/10.2174/2210327906666160606152043.
L. Alzubaidi et al., “Towards a Better Understanding of Transfer Learning for Medical
Imaging: A Case Study,” Appl. Sci., vol. 10, no. 13, Art. no. 13, Jan. 2020, doi:
5. Conclusion and future work
10.3390/app10134523.
D. Andrushia, M. Neebha, and T. Bella Mary, “An Intelligent Method for Indian
The rise in banknote fraud has made the development of detection Counterfeit Paper Currency Detection,” J. Inf. Technol. Manag., vol. 12, no. Special
technologies necessary. Although there are classic solutions being used, Issue: The Importance of Human Computer Interaction: Challenges, Methods and
Applications., Dec. 2020, doi: 10.22059/jitm.2020.79189.
such as computer vision and traditional machine learning techniques, H. Ashna and Z. Momand, “Applications of Machine Learning in Detecting Afghan Fake
their accuracy and generalization capacities are poor. This study pro­ Banknotes”.
posed a Jordanian banknote fraud detection model that integrates deep S. R. Awad, B. T. Sharef, A. M. Salih, and F. L. Malallah, “Deep learning-based iraqi
banknotes classification system for blind people,” East.-Eur. J. Enterp. Technol., vol. 1,
learning, using CNN and an attention mechanism. Utilizing a data set no. 2(115), pp. 31–38, Feb. 2022, doi: 10.15587/1729-4061.2022.248642.
from Kaggle Realizing that no research has been done on Jordanian E. Ayalew Tessfaw, B. Ramani, and T. Kebede Bahiru, “Ethiopian Banknote Recognition
currency in any of its denominations, we were able to make a significant and Fake Detection Using Support Vector Machine,” in 2018 Second International
Conference on Inventive Communication and Computational Technologies (ICICCT),
contribution by enhancing the richness of the original data set. This was Coimbatore: IEEE, Apr. 2018, pp. 1354–1359. doi: 10.1109/ICICCT.2018.8473013.
done by applying a technique for processing images to create fictional Azra, Y., Oct. 2020. Banknotes Counterfeit Detection Using Deep Transfer Learning
currency by boosting the brightness of the source images. Within the Approach. Int. J. Adv. Trends Comput. Sci. Eng. 9 (5), 8115–8122. https://doi.org/
10.30534/ijatcse/2020/172952020.
CNN structure, the Convolutional Block Attention Module (CBAM) was Brauwers, G., Frasincar, F., Apr. 2023. A General Survey on Attention Mechanisms in
employed. Here, two crucial dimensions are utilized to extract the Deep Learning. IEEE Trans. Knowl. Data Eng. 35 (4), 3279–3298. https://doi.org/
channel and spatial attention maps, which are subsequently integrated 10.1109/TKDE.2021.3126456.
B.-Q. Cao and J.-X. Liu, “Currency Recognition Modeling Research Based on BP Neural
by element-wise multiplication with the input feature map. To assess the
Network Improved by Gene Algorithm,” in 2010 Second International Conference on
suggested methodology, the CNN-CAM technique was employed and Computer Modeling and Simulation, Jan. 2010, pp. 246–250. doi: 10.1109/
evaluated on the acquired dataset. It was then paralleled with eight pre- ICCMS.2010.270.
trained deep learning models, namely VGG16, VGG19, ResNet50V2, J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “ImageNet: A large-scale
hierarchical image database,” in 2009 IEEE Conference on Computer Vision and Pattern
DenseNet121, DenseNet201, Convex-Tiny, Inceptionresnetv2, and CNN. Recognition, Jun. 2009, pp. 248–255. doi: 10.1109/CVPR.2009.5206848.
The superiority of the proposal is a result of the CNN being modified by M. Dirik, “Detection of Counterfeit Banknotes Using Genetic Fuzzy System,” 2022.

14
A. Nasayreh et al. Journal of King Saud University - Computer and Information Sciences 36 (2024) 102038

Eckle, K., Schmidt-Hieber, J., Feb. 2019. A comparison of deep networks with ReLU E. P. Office, “European publication server.” Accessed: Oct. 20, 2023. [Online]. Available:
activation function and linear spline-type methods. Neural Netw. 110, 232–242. https://data.epo.org/publication-server/document?iDocId=4897799&iFormat=0.
https://doi.org/10.1016/j.neunet.2018.11.005. Pachón, C.G., Ballesteros, D.M., Renza, D., Dec. 2023. An efficient deep learning model
“Evaluation of Texture Descriptors for Validation of Counterfeit Documents | IEEE using network pruning for fake banknote recognition. Expert Syst. Appl. 233,
Conference Publication | IEEE Xplore.” Accessed: Oct. 20, 2023. [Online]. Available: 120961 https://doi.org/10.1016/j.eswa.2023.120961.
https://ieeexplore.ieee.org/document/8270135. Park, C., Cho, S.W., Baek, N.R., Choi, J., Park, K.R., 2020. Deep Feature-Based Three-
A. Ganguly, A. U. Ruby, and G. C. C. J, “Evaluating CNN Architectures Using Attention Stage Detection of Banknotes and Coins for Assisting Visually Impaired People. IEEE
Mechanisms: Convolutional Block Attention Module, Squeeze, and Excitation for Access 8, 184598–184613. https://doi.org/10.1109/ACCESS.2020.3029526.
Image Classification on CIFAR10 Dataset,” In Review, preprint, Aug. 2023. doi: L. Perez and J. Wang, “The Effectiveness of Data Augmentation in Image Classification
10.21203/rs.3.rs-3245213/v1. using Deep Learning.” arXiv, Dec. 13, 2017. Accessed: Oct. 20, 2023. [Online].
G. Gebremeskel, T. A. Tadele, D. W. Girmaw, and A. O. Salau, “Developing a Model for Available: http://arxiv.org/abs/1712.04621.
Detection of Ethiopian Fake Banknote Using Deep Learning,” In Review, preprint, Pham, T.D., Park, C., Nguyen, D.T., Batchuluun, G., Park, K.R., 2020. Deep Learning-
Dec. 2022. doi: 10.21203/rs.3.rs-2282764/v1. Based Fake-Banknote Detection for the Visually Impaired People Using Visible-Light
Gharaibeh, M., et al., Sep. 2023. Optimal Integration of Machine Learning for Distinct Images Captured by Smartphone Cameras. IEEE Access 8, 63144–63161. https://doi.
Classification and Activity State Determination in Multiple Sclerosis and org/10.1109/ACCESS.2020.2984019.
Neuromyelitis Optica. Technologies 11 (5), 131. https://doi.org/10.3390/ Pham, T.D., Lee, Y.W., Park, C., Park, K.R., May 2022. Deep Learning-Based Detection of
technologies11050131. Fake Multinational Banknotes in a Cross-Dataset Environment Utilizing Smartphone
Gopane, S., Kotecha, R., 2020. Indian Counterfeit Banknote Detection Using Support Cameras for Assisting Visually Impaired Individuals. Mathematics 10 (9), 1616.
Vector Machine. SSRN Electron. J. https://doi.org/10.2139/ssrn.3568724. https://doi.org/10.3390/math10091616.
Han, M., Kim, J., Aug. 2019. Joint Banknote Recognition and Counterfeit Detection Pham, T., Nguyen, D., Kim, W., Park, S., Park, K., Feb. 2018. Deep Learning-Based
Using Explainable Artificial Intelligence. Sensors 19 (16), 3607. https://doi.org/ Banknote Fitness Classification Using the Reflection Images by a Visible-Light One-
10.3390/s19163607. Dimensional Line Image Sensor. Sensors 18 (2), 472. https://doi.org/10.3390/
K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” in s18020472.
2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, Pham, T., Nguyen, D., Park, C., Park, K., Feb. 2019. Deep Learning-Based Multinational
NV, USA: IEEE, Jun. 2016, pp. 770–778. doi: 10.1109/CVPR.2016.90. Banknote Type and Fitness Classification with the Combined Images by Visible-Light
Md. T. Islam, M. Ahmad, and A. S. Bappy, “Real-Time Bangladeshi Currency Recognition Reflection and Infrared-Light Transmission Image Sensors. Sensors 19 (4), 792.
Using Faster R-CNN Approach for Visually Impaired People,” in Communication and https://doi.org/10.3390/s19040792.
Intelligent Systems, vol. 204, H. Sharma, M. K. Gupta, G. S. Tomar, and W. Lipo, Eds., Rasheed, J., et al., Dec. 2020. A survey on artificial intelligence approaches in supporting
in Lecture Notes in Networks and Systems, vol. 204. , Singapore: Springer Singapore, frontline workers and decision makers for the COVID-19 pandemic. Chaos Solitons
2021, pp. 147–156. doi: 10.1007/978-981-16-1089-9_13. Fractals 141, 110337. https://doi.org/10.1016/j.chaos.2020.110337.
Jordanian Banknotes(old and new edition).” Accessed: Oct. 20, 2023. [Online]. Simonyan, K., Zisserman, A., 2015. Very Deep Convolutional Networks for Large-Scale
Available: https://www.kaggle.com/datasets/halabilbesei/jordanian-banknotesold- Image Recognition. arXiv, Apr. 10 https://doi.org/10.48550/arXiv.1409.1556.
and-new-edition. Tasnim, R., Pritha, S.T., Das, A., Dey, A., Jan. 2021. Bangladeshi Banknote Recognition
K. Kamble, A. Bhansali, P. Satalgaonkar, and S. Alagundgi, “Counterfeit Currency in Real-time using Convolutional Neural Network for Visually Impaired People. In:
Detection using Deep Convolutional Neural Network,” in 2019 IEEE Pune Section In 2021 2nd International Conference on Robotics, Electrical and Signal Processing
International Conference (PuneCon), Pune, India: IEEE, Dec. 2019, pp. 1–4. doi: Techniques (ICREST). IEEE, DHAKA, Bangladesh, pp. 388–393. https://doi.org/
10.1109/PuneCon46936.2019.9105683. 10.1109/ICREST51555.2021.9331182.
Khan, A., Sohail, A., Zahoora, U., Qureshi, A.S., Dec. 2020. A survey of the recent The Impact of Fatigue on Complex CT Case Interpretation by Radiology Residents -
architectures of deep convolutional neural networks. Artif. Intell. Rev. 53 (8), ScienceDirect.” Accessed: Oct. 20, 2023. [Online]. Available: https://www.
5455–5516. https://doi.org/10.1007/s10462-020-09825-6. sciencedirect.com/science/article/abs/pii/S107663322030355X.
J. W. Lee, H. G. Hong, K. W. Kim, and K. R. Park, “A Survey on Banknote Recognition A. Vaswani et al., “Attention is All you Need”.
Methods by Various Sensors,” Sensors, vol. 17, no. 2, Art. no. 2, Feb. 2017, doi: Wang, L., et al., Sep. 2022. Automated detection and classification of counterfeit
10.3390/s17020313. banknotes using quantitative features captured by spectral-domain optical coherence
M. Lin, Q. Chen, and S. Yan, “Network In Network.” arXiv, Mar. 04, 2014. doi: 10.48550/ tomography. Sci. Justice 62 (5), 624–631. https://doi.org/10.1016/j.
arXiv.1312.4400. scijus.2022.09.004.
J. Liu et al., “Inception Convolution With Efficient Dilation Search,” presented at the Xue, Z., Yu, X., Liu, B., Tan, X., Wei, X., 2021. HResNetAM: Hierarchical Residual
Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Network With Attention Mechanism for Hyperspectral Image Classification. IEEE J.
Recognition, 2021, pp. 11486–11495. Accessed: Oct. 20, 2023. [Online]. Available: Sel. Top. Appl. Earth Obs. Remote Sens. 14, 3566–3580. https://doi.org/10.1109/
https://openaccess.thecvf.com/content/CVPR2021/html/Liu_Inception_ JSTARS.2021.3065987.
Convolution_With_Efficient_Dilation_Search_CVPR_2021_paper.html. Yao, H., Zhang, X., Zhou, X., Liu, S., Nov. 2019. Parallel Structure Deep Neural Network
Meshram, V., Patil, K., Chumchu, P., Apr. 2022. Dataset of Indian and Thai banknotes Using CNN and RNN with an Attention Mechanism for Breast Cancer Histology
with annotations. Data Brief 41, 108007. https://doi.org/10.1016/j. Image Classification. Cancers 11 (12), 1901. https://doi.org/10.3390/
dib.2022.108007. cancers11121901.
S. Mittal and S. Mittal, “Indian Banknote Recognition using Convolutional Neural Youn, S., Choi, E., Baek, Y., Lee, C., May 2015. Efficient multi-currency classification of
Network,” in 2018 3rd International Conference On Internet of Things: Smart Innovation CIS banknotes. Neurocomputing 156, 22–32. https://doi.org/10.1016/j.
and Usages (IoT-SIU), Bhimtal: IEEE, Feb. 2018, pp. 1–6. doi: 10.1109/IoT- neucom.2015.01.014.
SIU.2018.8519888. D. Yu, H. Wang, P. Chen, and Z. Wei, “Mixed Pooling for Convolutional Neural
I. Namatēvs, “Deep Convolutional Neural Networks: Structure, Feature Extraction and Networks,” in Rough Sets and Knowledge Technology, vol. 8818, D. Miao, W. Pedrycz,
Training,” Inf. Technol. Manag. Sci., vol. 20, no. 1, Jan. 2017, doi: 10.1515/itms- D. Ślȩzak, G. Peters, Q. Hu, and R. Wang, Eds., in Lecture Notes in Computer Science,
2017-0007. vol. 8818. , Cham: Springer International Publishing, 2014, pp. 364–375. doi:
Naresh Kumar, S., Singal, G., Sirikonda, S., Nethravathi, R., Dec. 2020. “A Novel 10.1007/978-3-319-11740-9_34.
Approach for Detection of Counterfeit Indian Currency Notes Using Deep Zhang, J., Wang, Y., Wang, H., Wu, J., Li, Y., 2022. CNN Cloud Detection Algorithm
Convolutional Neural Network”, IOP Conf. Ser. Mater. Sci. Eng. 981 (2), 022018 Based on Channel and Spatial Attention and Probabilistic Upsampling for Remote
https://doi.org/10.1088/1757-899X/981/2/022018. Sensing Image. IEEE Trans. Geosci. Remote Sens. 60, 1–13. https://doi.org/
10.1109/TGRS.2021.3105424.

15