Hindawi

Journal of Sensors
Volume 2022, Article ID 4505089, 18 pages
https://doi.org/10.1155/2022/4505089

Research Article
Ethiopian Banknote Recognition Using Convolutional Neural
Network and Its Prototype Development Using
Embedded Platform

Dereje Tekilu Aseffa , Harish Kalla , and Satyasis Mishra

Department of Electronics and Communication Engineering, Adama Science and Technology University, Adama 1888, Ethiopia

Correspondence should be addressed to Harish Kalla; harish.kalla@astu.edu.et

Received 6 September 2021; Accepted 15 December 2021; Published 12 January 2022

Academic Editor: Cong-Bin Fan

Copyright © 2022 Dereje Tekilu Aseffa et al. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.

Money transactions can be performed by automated self-service machines like ATMs for money deposits and withdrawals,
banknote counters and coin counters, automatic vending machines, and automatic smart card charging machines. There are
four important functions such as banknote recognition, counterfeit banknote detection, serial number recognition, and fitness
classification which are furnished with these devices. Therefore, we need a robust system that can recognize banknotes and
classify them into denominations that can be used in these automated machines. However, the most widely available banknote
detectors are hardware systems that use optical and magnetic sensors to detect and validate banknotes. These banknote
detectors are usually designed for specific country banknotes. Reprogramming such a system to detect banknotes is very
difficult. In addition, researchers have developed banknote recognition systems using deep learning artificial intelligence
technology like CNN and R-CNN. However, in these systems, dataset used for training is relatively small, and the accuracy of
banknote recognition is found smaller. The existing systems also do not include implementation and its development using
embedded systems. In this research work, we collected various Ethiopian currencies with different ages and conditions and
applied various optimization techniques for CNN architects to identify the fake notes. Experimental analysis has been
demonstrated with different models of CNN such as InceptionV3, MobileNetV2, XceptionNet, and ResNet50. MobileNetV2
with RMSProp optimization technique with batch size 32 is found to be a robust and reliable Ethiopian banknote detector and
achieved superior accuracy of 96.4% in comparison to other CNN models. Selected model MobileNetV2 with RMSProp
optimization has been implemented through an embedded platform by utilizing Raspberry Pi 3 B+ and other peripherals.
Further, real-time identification of fake notes in a Web-based user interface (UI) has also been proposed in the research.

1. Introduction devices prevent users from authenticating and denominating

banknotes. Feature extraction is an important method in the
In the digital world, money transactions play a vital role in process of classification of banknotes.
day-to-day life. Thousands of organizations perform digital The banknote authentication process of such machines
transactions every day. Moreover, in the digital technology consists of a series of operations. They are fake currency
world, banking operations and authentication of currencies identification, banknote denomination recognition, and
are mandatory in many applications such as smart card charg- serial number recognition. The paper currency denomina-
ing machines to pay for electricity and transport, money tion is distinguished by using the banknote images with
exchange machines, Automated Teller Machine (ATM), vend- machine learning algorithms. The result of this process is
ing machines for drinks, tollgate ticket-vending machines at used in validation and testing. Most of the countries are
highways, and parking meters at shopping malls. These using these machines adequately. These machines can be
devices are user-friendly and make jobs easier and fast. These used to Ethiopia environment by changing a few algorithms
2 Journal of Sensors

Dataset captured by
camera

Collected frames
annotation

Training dataset Validation dataset

Pre-trained model
configurations with
optimizations Training model

Model evaluation Testing dataset

Collection metrics values

and inference time

Optimally frontier
construction

Figure 1: Methodology of the proposed model.

in the software to identify and classify the Ethiopian cur- (i) In this research, classification of the Ethiopian
rency (ETB). In this research, we presented the results to banknotes using four different CNN architectures
show an effective scheme for Ethiopian banknote recogni- is employed. Those are InceptionV3, MobileNetV2,
tion. We have employed MobileNetV2 convolutional neural XceptionNet, and Resnet50
network (CNN) architecture-based banknote recognition
algorithm along with the embedded system hardware to ver- (ii) To minimize the loss function, various optimization
techniques like Adam, SGD, RMSProp, Nadam,
ify and classify the denominations of the Ethiopian bank-
Adagrad, and Adadelta are utilized to abovemen-
notes. The general objective and specific objective of the
tioned CNN architectures
research are as follows.
(iii) Applied various batch sizes 32, 64, and 128 are
1.1. General Objective. The objective of this project is to taken to understand the performance of hyperpara-
design and implement an efficient Ethiopian banknote meters of each CNN architecture with different
detector system using CNN architecture, an embedded hard- optimization techniques
ware system (Raspberry Pi B+), and a Web-based app for the
user interface. (iv) Compared the training loss, training accuracy, vali-
dation loss, validation accuracy, training time, and
1.2. Specific Objective. Particularly, this research has the fol- prediction time of all these models
lowing specific objectives: (v) Measured the performance sensitivity, specificity,
and accuracy of different classifiers
(i) To collect various Ethiopian currencies of different
ages and conditions (vi) Based upon all these results, the best-performed
model hardware system has been designed to cap-
(ii) To identify the best architecture model from exist-
ture banknote image and classify different
ing CNN architectures
denominations
(iii) Apply various optimization techniques for CNN
architects to identify the best optimization Banknotes have been effectively used in almost all the
technique money transactions like electronic currencies daily. Cur-
rency recognition and detection whether original or fake is
(iv) To develop Web-based app (software) for user an important issue in computer vision. Even though various
interface types of digital currency transaction systems exist, cash is
(v) To implement a novel architecture with the best still the favored choice in daily transactions such as vending
optimization technique using the embedded system machines, banks, shopping complexes, railway station coun-
for classifying banknotes ters, automatic teller machines, and forex. Moreover, there
may be situations where the monies drawn have fake curren-
The research contributions are as follows: cies. This issue is still unavoidable, and the banks are not
Journal of Sensors 3

Data preparation Training Testing

Bank note Denomination

dataset Selection of detection
CNN model

Separation of training, Evolution

testing and validation data
Freeze all layers in the
base model
Resize images
to 224×224
Create a model on top of the output of
layers from the base models

Data augmentation
Adding dense layers after
pre-trained model

Set optimizer with

learning rate = 0.0001

Fitting and training the CNN

model with 100 epochs

Calculation of loss function, cross

entropy category

Plotting model training and

validation result

Bank notes classification

Metric values evaluation

Figure 2: Retrain stage process of the proposed model.

able to supply the required fake currency detection system in segmenting the thin golden vertical strip which is on the cur-
all the branches [1]. rency denomination of ETB notes of 50 and 100. Test results
The four distinctive features of the banknotes such as the showed that the proposed design had an average detection rate
dominant color, the distribution of the dominant color, the of 90.42% for genuine currency, with an average prediction
hue value, and speeded-up robust features (SURF) were time of 1.68 seconds per banknote [3]. Though the researchers
extracted as the referential features of banknotes in order to have used a suiFigure method for classification as well as fraud
deduct the counterfeit currencies [2]. The above four features recognition, the techniques suffer precision issues such as clas-
in combination with local feature descriptors were considered sifying old genuine bills whose color become fading. The fad-
in a four-level classification process. During the execution, ing of color of the currencies might be due to hard-and-fast
among the four, one of the four features was extracted as a circulation of the currencies by careless peoples. The spatial
classification task. The correlation coefficient-based template information of the banknote relates to the color present in
matching was used for classification. To check the originality the note but not the color density. During the image acquisi-
of the currencies, final verification tasks were performed by tion process illumination conditions, it is expected that it
4 Journal of Sensors

0 0 0 0 0
25 25 25 25 25
50 50 50 50 50
75 75 75 75 75
100 100 100 100 100
125 125 125 125 125
150 150 150 150 150
175 175 175 175 175

0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150
0 0 0 0 0
25 25 25 25 25
50 50 50 50 50
75 75 75 75 75
100 100 100 100 100
125 125 125 125 125
150 150 150 150 150
175 175 175 175 175

0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150
0 0 0 0 0
25 25 25 25 25
50 50 50 50 50
75 75 75 75 75
100 100 100 100 100
125 125 125 125 125
150 150 150 150 150
175 175 175 175 175

0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150
0 0 0 0 0
25 25 25 25 25
50 50 50 50 50
75 75 75 75 75
100 100 100 100 100
125 125 125 125 125
150 150 150 150 150
175 175 175 175 175

0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150
0 0 0 0 0
25 25 25 25 25
50 50 50 50 50
75 75 75 75 75
100 100 100 100 100
125 125 125 125 125
150 150 150 150 150
175 175 175 175 175

0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150

Figure 3: Sample images of Ethiopian currency collected from the bank.

Table 1: Splitting the banknote datasets. The Ethiopian currency may have similar color features;
because of this, it is not able to classify accurately the denom-
Type of currency Training Validation Testing ination based on color features. In addition, the template
Fifty 1430 231 427 matching classifiers are variant to intensity value change.
Five 1534 238 449 Some researchers proposed the following methods to enhance
Hundred 1393 219 419 the feature extraction of the images [3]. SIFT, GLCM, color
Ten 1345 221 427 momentum, CNN, and a combination of SIFT, GLCM, and
color momentum techniques. Another researcher proposed
Total 5702 909 1722
and improved the feature extractions using feed-forward arti-
Total images: 8333 ficial neural network as a classifier for the design of the Ethio-
pian paper currency recognition system. Convolutional neural
network (CNN) model utilizes a total of 2400 banknotes
would not vary, but the notes may lose their intensities images which were collected through the scanner for training
because of the worn-out or they may have dirt. The banknote and testing [3]. A CNN-based Ethiopian currency recognition
recognizing process is sufficient in using RGB space density. model for vision-impaired people has been developed in
Journal of Sensors 5

Table 2: Image quantity after data augmentation generation.

train_data_generator = ImageDataGenerator(
zca_epsilon = 1e − 06, Type of currency Training Validation Testing
rotation_range = 45, Fifty 4079 320 3227
width_shift_range = 0.2,
Five 4290 334 3355
height_shift_range = 0.2,
brightness_range = none, Hundred 4039 322 3176
fill_mode = “nearest”, Ten 4004 304 3130
horizontal_flip = true, Total 16,412 1280 12,888
vertical_flip = false, Total images: 30,580
rescale = none,
)

trum method is employed for the classification of Korean

Pseudocode 1: Data generation.
won notes. The wavelet domain features of Korean won bill
images were extracted using the Sobel filter. Additionally, to
reduce the features of image space coordinates employed the
which they used pretrained models such as TensorFlow object canonical analysis. Finally, it is required to send the output
detection API and the pretrained Faster R-CNN (region-based image to the K-nearest neighbor method to recognize the
CNN), Inception, and SSD (single-shot multibox detection) Korean won currency notes [11].
MobileNetV2 models with CNN architecture and tested them Modern countries have been investigating to develop
in real-time scenarios. They predicted the outcomes from the innovative technologies for authenticating their banknotes.
system by applying various input methods such as partially A Canadian dollar scanner has been developed with prere-
damaged banknotes, fully damaged banknotes, and folded corded voice output. The user can keep the banknote under
currencies. This model showed an average accuracy value of the scanner a prerecorded voice output informing the
91.8% which is considered as very less when compared to denomination value. This scanner works only for Canadian
our proposed method. In this approach, for every 100 transac- dollars [12]. A Sri Lankan (LKR) rupee classification system
tions, approximately 8 transactions are found to be abnormal. has been designed by utilizing a unique linear transforma-
Hence, this approach has not been considered for commercial tion function, edge detection, and backpropagation neural
usage. Moreover, one of the drawbacks of their approach is networks [13]. The banknote recognition was conducted
that they have not come up with the procedures to calculate using AdaBoost and SURF-based methods [14–16]. Further-
the total processing time for the prediction of banknotes [4]. more, an image processing-based Bangladesh bill recogni-
The traditional approach of paper money recognition is tion technique has been designed in which the researcher
the identification of visible features found on a paper cur- extracted the features from Latent Image, watermark, and
rency like the historical people images, like father of the microprinting of the bill using the Histogram of Gradient
nation and freedom fighters, and landmark buildings, like (HOG) method. To classify the Bangladesh bills, they uti-
parliament buildings, numbers, color, texture, signature, lized the SVM (support vector machine) algorithm. This
and size of banknote. These textures are extracted as features design was found to provide 100% of classification accuracy.
using texture-modeling techniques. The disadvantage of this However, the designed system tested only on the 1000 and
approach is that it is difficult to recognize the old age of 500 Bangladesh banknote denominations [1]. In the same
banknotes torn due to their regular usage [5]. The researcher view, the portable system for blind people for Euro banknote
predicted the denomination classification based on geomet- detection and recognition is presented in [17]. Finally, a
ric features of banknotes [6]. This approach is applicable researcher presented a currency recognition system for six
only for countries that have banknote denominations of var- different types of US dollars in which the banknote images
ious sizes like China and Italy. It was unsuitable for US dol- are processed by utilizing the line sensor. To extract the
lars, which had a similar size for various denominations. banknote image features, the Principal Component Analysis
Moreover, most of the country’s banknotes’ denomination (PCA) algorithm was adopted and Linear Vector Quantiza-
colors are different [7]. The color of Indian paper currencies tion (LVQ) network was utilized for classification [14]. Sev-
is different for all denominations. Using these various colors, eral researchers have recommended server-client model-
the author developed an Indian banknote denomination based online banknote recognition systems for Slovak EURO
classification [8]. Similarly, another researcher also investi- currency. This model employed a SIFT detector. However,
gated the currency recognition system using color conver- this model had a problem with a long recognition time
sion techniques. The researcher extracted the color features [18]. A currency recognition system for Ethiopian banknotes
of Indian currencies by using HSV (Hue, Saturation, and using a support vector machine has been discussed in [19],
Value) color space [9]. Furthermore, a researcher urged on which recognizes the front part of the currency well.
extracting the Euro currency note features using unique Another researcher proposed an Indian banknote detection
characteristics of image parts. Similarly, contributed method using side invariant technique. In this technique,
research on feature extraction techniques using coordinate the system used a template matching method. It compared
data size possesses the same characteristics of image color the edges of the input test images and reference images from
[10]. Likewise, a new approach called the visible light spec- the database. Finally, the system classified the paper
6 Journal of Sensors

0 0 0 0 0
25 25 25 25 25
50 50 50 50 50
75 75 75 75 75
100 100 100 100 100
125 125 125 125 125
150 150 150 150 150
175 175 175 175 175

0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150
0 0 0 0 0
25 25 25 25 25
50 50 50 50 50
75 75 75 75 75
100 100 100 100 100
125 125 125 125 125
150 150 150 150 150
175 175 175 175 175

0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150
0 0 0 0 0
25 25 25 25 25
50 50 50 50 50
75 75 75 75 75
100 100 100 100 100
125 125 125 125 125
150 150 150 150 150
175 175 175 175 175

0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150
0 0 0 0 0
25 25 25 25 25
50 50 50 50 50
75 75 75 75 75
100 100 100 100 100
125 125 125 125 125
150 150 150 150 150
175 175 175 175 175

0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150
0 0 0 0 0
25 25 25 25 25
50 50 50 50 50
75 75 75 75 75
100 100 100 100 100
125 125 125 125 125
150 150 150 150 150
175 175 175 175 175

0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150

Figure 4: Output images of image data augmentation techniques rotation, shifts, and flips.

currency depending on the generated match score. Due to in the input signal which can lead to an unpredictable
the limitations of the proposed system, it showed a low accu- change in the transform coefficient. The drawback of this
racy rate because of the positioning of the input currency method is that it generates a spatial-domain feature in geo-
image, input image brightness, and image quality system metrical positions [23]. Similarly, a new approach, grid fea-
[20]. Furthermore, currency classification by employing a tures of images, was invented to classify the banknote
robust method called block local-binary pattern was investi- images. In this approach, the banknote image feature vector
gated by researchers [21]. Similarly, a multicurrency detec- is calculated by using the average gray value of partitioned
tion system for Chinese Yuan and Euro paper currencies m × n divisions of the input image. The system also had
has been designed and tested where they followed the gener- other flaws such as determining the best m × n partition size
alized Gaussian density (GGD), quaternion wavelet trans- and number of features for efficient classification [24]. The
form (QWT), and a backpropagation neural network deep convolutional neural networks are implemented and
(BPNN) to detect the multicurrency notes [22]. Addition- trained on the DenseNet-121 architecture [25]. Finally, the
ally, Persian banknote detection is projected by using Haar SVM learning algorithm-based Nigerian currency classifica-
wavelet transform (HWT) and multilayer perceptron tion system was proposed and used a genetic algorithm for
(MLP) neural network. The benefit of HWT is as a change parameter optimization. A predefined decision function was
Journal of Sensors 7

3 channel input Filters ⁎ 5 Maps ⁎ 5 Filters ⁎ 5 Maps ⁎ 5 Filters ⁎ 4 Maps ⁎ 4

Pointwise Depthwise Pointwise

Figure 5: MobileNetV2 model architecture.

correction algorithms. These banknote images are trained

using an efficient AdaBoost algorithm. This approach works
on various mobile operating systems like Windows and
Symbian. However, though the mobile phone-based system
1 × 1 “Expansion” layer
used efficient algorithms, they took more processing time
which needed to be improved for real-time scenarios [28].
Batch normalization Similarly, a technique proposed for determining the US dol-
lar employed an image identification approach called eigen-
ReLU6 faces based on the Principal Component Analysis (PCA).
The system analyzed the banknote images on two different
Nokia smartphones. It is found that the system had a preci-
3 × 3 Depthwise convolution sion rate of 99.8%. The system had limitations such as image
Batch normalization
brightness and segmenting the background of banknote
images [29]. Correspondingly, a bionic eyeglass was imple-
ReLU6 mented to recognize the banknotes. Bionic eyeglass was
interfaced with a Samsung Galaxy S mobile phone, and this
experiment was tested with five visually challenged persons.
1 × 1 “Projection” layer This system utilized K-means and CNN architecture for the
classification of banknote images. Bionic eyeglass has an
Batch normalization integrated micro camera to capture the banknote image.
After that, image is transmitted to the computer system by
using server/client architecture for recognition. Here, Sam-
+ sung mobile acts as an interface device just to send the
images to the computer system through the Bionic eyeglass
with no processing capability [30]. Finally, a Jordan paper
currency recognition system for both colored and grayscale
Bottleneck residual block images with a dataset size of 400 has been developed. This
system used to test on both grayscale and color images.
Figure 6: MobileNetV2 convolutional layer models [35].
The SIFT features were extracted from both colored and
grayscale images and compared with the classification rate.
applied in the SVM learning algorithm. However, an efficient However, the system faced the same issues with the illumi-
set of features was not calculated for Nigerian banknote classi- nation difficulty and the system was comparatively slow
fication. Further, the system was not evaluated using other when interfaced on a cell phone [31]. Optimization is one
global banknotes [26]. of the core components of machine learning, “convolution
Globally, various countries’ banknote recognition sys- neural network (CNN) with RMSProp” (Root Mean Square
tems have been designed to help visually impaired persons. Propagation) on a standard “UCI machine learning reposi-
Few of these systems are LookTel Money Scanner and tory” dataset [32]. A technique the PSOSGD proposed by
IDEAL Currency Identifier to predict any denomination of Albeahdili is superior to the traditional method in training
US dollars. Apart from this, there are dedicated mobile [33]. To boost the classification performance, four different
readers, namely, Money Talker, Note TEller 2, and KNFB optimizers are employed with the TF-CNN, Adagrad, Prox-
designed for Australian and US dollar currency recognition imal Adagrad, Adam, and RMSProp [34].
[27]. A mobile phone-based US banknote detection system Banknote recognition and classification systems play a
was employed with background subtraction and perspective vital role in automatic money transaction machines. Most
8 Journal of Sensors

Monitor

Camera

Raspberry pi module
Wireless mouse
Power adaptor

er
hold
u r rency Wireless keyboard
C

Figure 7: Experimental setup.

Table 3: Training accuracy of batch sizes 32, 64, and 128.

Model
Optimization Batch size 32 Batch size 64 Batch size 128
IV3 MN XN RN IV3 MN XN RN IV3 MN XN RN
Adam 96.38 96.96 94.80 94.78 96.30 96.96 95.44 94.85 96.46 96.96 94.61 94.83
SGD 95.75 94.56 91.59 96.56 96.30 92.65 94.12 96.42 96.46 92.93 92.28 96.47
RMSProp 96.59 96.98 92.28 94.65 96.58 96.98 92.20 94.57 96.57 96.98 96.20 94.56
Nadam 96.45 96.96 96.29 94.93 96.35 96.96 96.24 96.00 96.37 96.96 96.25 96.02
Adadelta 59.34 87.31 89.61 94.75 51.15 86.09 91.35 94.99 46.46 87.10 94.30 94.89
Adagrad 91.48 94.82 94.47 96.88 85.34 94.89 94.44 96.87 91.94 96.00 94.61 96.88

Table 4: Test accuracy of batch sizes 32, 64, and 128.

Model
Optimization Batch size 32 Batch size 64 Batch size 128
IV3 MN XN RN IV3 MN XN RN IV3 MN XN RN
Adam 96.26 96.94 90.94 79.06 94.30 95.95 96.32 77.26 96.29 96.86 96.34 79.47
SGD 94.21 93.20 92.18 96.87 94.30 91.34 92.70 96.85 96.29 93.46 92.76 96.72
RMSProp 96.45 96.96 92.76 76.06 96.42 96.93 90.33 77.16 96.52 96.90 92.96 78.57
Nadam 92.30 96.92 91.12 75.63 94.01 96.90 94.59 80.48 94.04 96.98 94.25 80.62
Adadelta 61.61 88.15 90.69 96.00 50.28 89.05 93.52 96.33 46.95 87.91 94.21 96.24
Adagrad 90.19 96.21 96.38 96.96 90.19 96.22 96.32 96.97 90.53 96.29 96.34 96.97

of the reviewed research mentioned played a substantial role are concentrated on different neural network architectures
in the problem area of banknote recognition and classifica- and do not consider the effect of optimization techniques.
tion. Even if they have their limitations/gap such as small Most of the researchers concentrated on various techniques
datasets, they require special knowledge on the usage of sys- and algorithms, but they did not mention the type and spec-
tems, and moreover, they are found to be expensive. Some of ifications of the computer system to run these algorithms
the research needs more processing time to see the results efficiently. Though most of the researchers do analysis on
because of more computational algorithms and internal high-speed computers, there are limitations like the cost
architecture layers. In addition, other researchers are and complexity of the system to use and operate. Hence,
expected to come up with issues like folding, lighting condi- apart from the software programs, algorithms, and architec-
tions, more processing time, paper currency quality, and tures, it is also necessary to implement it on sophisticated
worn currency notes. Moreover, all the above research works embedded systems with lightweight architectures in order
Journal of Sensors 9

Table 5: Train loss of batch sizes 32, 64, and 128.

Model
Optimization Batch size 32 Batch size 64 Batch size 128
IV3 MN XN RN IV3 MN XN RN IV3 MN XN RN
Adam 0.02 0.00 0.07 0.05 0.02 0.00 0.07 0.05 0.02 0.00 0.07 0.05
SGD 12.75 18.01 21.80 2.66 0.02 0.17 0.18 0.03 0.02 0.17 0.18 0.03
RMSProp 0.01 0.00 0.18 0.10 0.02 0.00 0.09 0.10 0.02 0.00 0.06 0.09
Nadam 0.36 0.00 0.03 0.04 0.35 0.00 0.03 0.04 0.24 0.00 0.03 0.04
Adadelta 1.15 0.49 0.34 0.06 1.24 0.48 0.32 0.05 1.26 0.50 0.31 0.05
Adagrad 0.33 0.07 0.07 0.01 0.33 0.07 0.07 0.01 0.32 0.07 0.07 0.01

Table 6: Test loss of batch sizes 32, 64, and 128.

Model
Optimization Batch size 32 Batch size 64 Batch size 128
IV3 MN XN RN IV3 MN XN RN IV3 MN XN RN
Adam 3.28 0.34 104.48 183.74 27.65 0.21 3.99 182.01 2.44 0.50 3.79 250.94
SGD 7.92 15.81 14.42 1.06 37.65 15.30 13.57 1.30 2.44 14.83 13.88 1.44
RMSProp 1.83 0.11 13.88 291.95 2.30 0.18 110.69 373.63 2.72 0.24 160.91 455.04
Nadam 1.38 0.24 88.62 484.78 2.94 0.44 87.80 189.05 5.56 0.08 48.57 416.08
Adadelta 112.05 46.75 28.59 4.19 124.38 45.63 26.05 3.92 126.22 47.32 25.34 3.93
Adagrad 25.90 5.83 3.92 0.29 25.86 5.96 3.99 0.34 25.38 5.90 3.79 0.34

Table 7: Validation time of batch sizes 32, 64, and 128 for classifiers and optimizers.

Model
Optimization Batch size 32 Batch size 64 Batch size 128
IV3 MN XN RN IV3 MN XN RN IV3 MN XN RN
Adam 34.15 36.23 128.91 36.82 35.15 35.09 310.34 59.58 43.33 52.57 132.35 36.23
SGD 33.92 35.12 155.89 35.44 38.92 35.09 124.84 51.14 43.33 50.91 304.66 35.6
RMSProp 35.87 50.91 135.54 37.35 37.87 36.22 132.87 53.21 34.66 53.05 129.42 36.4
Nadam 37.6 55.36 187.16 40.85 39.6 36.58 124.09 56.84 37.49 55.93 125.39 40.26
Adadelta 36.65 51.23 127.73 69.61 42.65 45.51 126.27 51.77 237.57 51.44 664.41 71.11
Adagrad 37.79 52.42 128.13 89.92 37.79 36.84 123.32 59.58 36.7 52.57 123.41 66.07

to provide high processing speed with less validation time. 2.1. Flowchart of the Proposed Methodology.
Hence, there is a need to come up with a development of
the best accuracy currency recognition models using CNN 2.2. Preparing Dataset. The Ethiopian banknote images were
classifier with a good optimization technique on sophisti- captured using a Logitech C920 digital camera with a resolu-
cated embedded systems. Therefore, in this research project, tion of 1080 pixels. All banknotes were collected from the
we designed and implemented an efficient Ethiopian bank- Cooperative Bank of Oromia, Deka Adi Branch, Adama,
note detector system using CNN architecture, an embedded Ethiopia. This collected dataset contains four denomina-
hardware system (Raspberry Pi B+), and a Web-based app tions/classes such as five, ten, fifty, and hundred with dem-
for the user interface. onstrated sample images in Figure 3. Images of banknotes
are in RGB format and are stored in Joint Photographic
Experts Group (JPEG) format. They possess some challenges
2. Materials and Methods such as blurry images, partially occluded signs, low resolu-
tion, poor illumination, and furthermore different dimen-
This part is to propose the methodology to determine the sions. These images are filtered and renewed in size into
best CNN model for the real-time Ethiopian banknote clas- 500 × 500 pixels.
sification by iteratively varying the pretrained model config-
urations. Figure 1 demonstrates the block diagram of the 2.3. Splitting Data. When the samples in the dataset increase
proposed model procedure. Figure 2 shows retrain stage enormously and are diverse, splitting data with ratio of train
process of the proposed model, which is described below. dataset 70%, validation dataset 10%, and test dataset 20% is a
10 Journal of Sensors

The average training accuracy of batch sizes 32, 64 and 128

98 120
97
96 100
95 80
94
93 60
92
91 40
90 20
89
88 0
IV3 MN XN RN IV3 MN XN RN IV3 MN XN RN
Batch size 32 Batch size 64 Batch size 128
Adam Nadam
SGD Adadelta
RMSProp Adagrad

Figure 8: Graph of the average training accuracy for each dataset with each classifier and optimizer. IV3: InceptionV3; MN: MobileNetV2;
XN: XceptionNet; RN: ResNet50.

The average test accuracy of batch sizes 32, 64 and 128

120 120

100 100

80 80

60 60

40 40

20 20

0 0
IV3 MN XN RN IV3 MN XN RN IV3 MN XN RN
Batch size 32 Batch size 64 Batch size 128
Adam Nadam
SGD Adadelta
RMSProp Adagrad

Figure 9: Graph of the average test accuracy for each dataset with each classifier and optimizer.

good choice. Table 1 represents the splitting ratio of Ethio-

pian banknotes.
The average train loss of batch sizes 32, 64 and 128
25 1.4 2.4. Augmenting Dataset. In short, data augmentation is a
1.2
preprocessing technique to help us to generate more image
20 variations to both avoid overfitting and increase the accuracy
1 of the model. Here is the index of techniques we used in the
15 0.8 research.
Image rotation helps our model to become more robust
10 0.6
to the changes in the orientation of banknotes. The informa-
0.4 tion of the image remains the same, for example, a ten-birr
5 note is a ten-birr note even if we see it from a different angle.
0.2
When the image is rotated, some pixels will move outside
0 0
the image and leave an empty area that needs to be filled
IV3 MN XN RN IV3 MN XN RN IV3 MN XN RN
in. To fill this empty area, use the nearest pixel values. Here,
Batch size 32 Batch size 64 Batch size 128
we rotated images into 45°. All the bank currency notes may
Adam Nadam
not always be in the center of the image. To overcome this
SGD Adadelta
RMSProp Adagrad
problem, we shifted the pixels of the image either horizon-
tally or vertically.
Figure 10: Graph of the average train loss for each dataset with Random flip technique should be according to the object
classifier and optimizer. in the image. The pretrained network will be trained on
Journal of Sensors 11

The average test loss of batch sizes 32, 64 and 128

500 600
450
400 500
350 400
300
250 300
200
150 200
100 100
50
0 0
IV3 MN XN RN IV3 MN XN RN IV3 MN XN RN
Batch size 32 Batch size 64 Batch size 128
Adam Nadam
SGD Adadelta
RMSProp Adagrad

Figure 11: Graph of the average test loss for each dataset with each classifier and optimizer.

The average validation time

350 700
300 600
250 500
200 400
150 300
100 200
50 100
0 0
IV3 MN XN RN IV3 MN XN RN IV3 MN XN RN
Batch size 32 Batch size 64 Batch size 128
Adam Nadam
SGD Adadelta
RMSProp Adagrad

Figure 12: Graph of the average validation time for each dataset with each classifier and optimizer.

patches of images which are randomly flipped from the orig- feature extraction of data into training and validation sets.
inal dataset. The various techniques of data augmentation After evaluation of the model, save the weights into .h5py
output images are shown in Figure 4. So here we flipped format. These saved files are loaded and tested using Rasp-
the banknotes horizontally. After the data augmentation, berry Pi 4 B+ model processor. The hardware setup is shown
we obtained 30,580 samples as shown in Table 2. in Figure 7.
2.5. MobileNetV2. MobileNetV2 is a CNN architecture model
for image classification and mobile vision. MobileNetV2 has 4. Results and Discussions
two new global hyperparameters (width multiplier and resolu-
tion multiplier) that allow model developers to tradeoff latency For performance evaluation, the designed framework has
or accuracy for speed and low size depending on their require- been trained on four (4) types of convolutional neural net-
ments. The convolutional layer and model architectures are work architectures including InceptionV3, MobileNetV2,
shown in Figures 5 and 6. XceptionNet, and ResNet50, individually. Six different opti-
mization techniques, namely, Adam, SGD, RMSProp,
3. Experimental Setup Nadam, Adadelta, and Adagrad, have been utilized in the
CNN models. All these models and optimization techniques
3.1. Training and Testing Infrastructure. The hardware uti- are conducted simultaneously for 100 epochs. The experi-
lized comprised a Core i7, 2.4 GHz processor with 16 GB ments are performed on a high-performance computer
RAM, and NVIDIA RTX2060 GPU. The operating system using NVIDIA RTX2060 GPU. When training a CNN
used was Windows 10, 64-bit. Image processing techniques model, one of the important aspects is to avoid overfitting.
and other related Python, Keras, and Transfer learning tool- This is occurring when the model fits the training data well,
boxes were utilized in the development of the system. The but it cannot generalize and make accurate predictions for
research utilized cross-validation for data analysis. It is a sta- data that has not been seen before. To avoid overfitting, a
tistical technique used in evaluating predictive models by cross-validation technique has been employed, where we
12 Journal of Sensors

Table 8: Sensitivity of batch sizes 32, 64, and 128 for classifiers and optimizers.

Model
Optimization Batch size 32 Batch size 64 Batch size 128
IV3 MN XN RN IV3 MN XN RN IV3 MN XN RN
Adam 95.98 93.87 93.52 89.96 96.1 96.26 92.44 90.83 95.9 95.84 94.96 89.38
SGD 94.56 95.56 91.59 93.56 95.3 96.10 94.42 93.6 94.46 95.7 91.92 93.5
RMSProp 95.95 96.53 89.81 96.02 95.85 96.45 95.32 96.08 95.75 96.49 95.12 95.82
Nadam 94.54 94.35 95.29 89.26 96.01 95.95 96.04 90.36 95.73 95.02 95.25 93.36
Adadelta 69.34 91.57 91.17 79.31 69.41 96.10 92.66 82.73 66.57 94.99 93.43 84.53
Adagrad 87.84 96.01 95.47 85.82 89.34 96.15 89.67 86.62 87.65 96.08 94.18 84.82

Sensitivity of batch sizes 32, 64 and 128 for classifiers Figure 8, the highest accuracy is obtained in MobileNetV2
and optimizers using the RMSProp optimization technique. Apart from this,
98 120 using the combinations of MobileNetV2 with Adam and
MobileNetV2 Nadam optimizations, it is found that accu-
96 100 racy of 96.96% is achieved which is the second-best accuracy
94 value among other optimization methods. However, the
80
92 Resnet50 model also shows better accuracy value of 96.88%
60 using Adagrad optimization with batch sizes 32 and 128.
90
40 But it is also almost nearer to the value of the first model
88 which is the MobileNetV2 model. Except for the model
86 20 InceptionV3 with Adadelta optimization technique, all other
84 0
optimized combinations produced more than 90% accuracy.
IV3 MN XN RN IV3 MN XN RN IV3 MN XN RN The Adadelta optimization technique shows low perfor-
Batch size 32 Batch size 64 Batch size 128 mance with an accuracy of 46.46% for the InceptionV3
Adam Nadam
model with a batch size of 128. Hence, this optimization
SGD Adadelta technique is unfit to predict all the Ethiopian currencies.
RMSProp Adagrad Hence, MobileNetV2 is more accurate than InceptionV3,
XceptionNet, and ResNet50 while being smaller and more
Figure 13: Sensitivity of the proposed architecture models. than 2.5 times less computation. The graph in Figure 6
shows the persistence of training accuracy for all the pro-
split the data into three parts—the training dataset, the test posed models and optimization techniques. It is found that
dataset, and the validation dataset. The splitting data are when compared to MobileNetV2, Resnet50 is showing
performed in three different ways with the ratio of 70% for almost constant values for all optimization techniques and
training, 20% for testing, and 10% for validation. The loss batch sizes.
and accuracy are measures of loss and accuracy on the train-
ing set, while test loss and test accuracy are measures of loss 4.2. Test Accuracy. The test accuracy, we must measure per-
and accuracy on the test data set. Tables 3–7 present the formance on unseen data, which are not used for training
average of train accuracy, test accuracy, train loss, test loss, data. To measure the test accuracy, use the test dataset.
and validation time of banknote recognition using Incep- The banknote classification test accuracy results are pre-
tionV3, MobileNetV2, XceptionNet, and ResNet50 architec- sented in Table 4. It is understood that the highest training
tures with various batch sizes of 32, 64, and 128, respectively. accuracy of 96.97% is achieved by ResNet50 using Adagrad
Figures 8–11 represent the persistence graph of train accu- optimization in batch sizes 64 and 128. The results are also
racy, test accuracy, train loss, test loss, and validation time presented graphically and are shown in Figure 9. As shown
of the abovementioned models and optimizations. in Figure 9, the highest accuracy is obtained in ResNet50
using the Adagrad optimization technique. However, the
4.1. Train Accuracy. A train accuracy metric is used to mea- MobileNetV2 model also shows better accuracy value of
sure the algorithm’s performance in an interpretable way. It 96.96% using RMSProp optimization with batch size 32.
is the measure of how accurate the proposed model’s predic- But it is also almost nearer to the value of the first model
tion is compared to the true data. The banknote classifica- which is the ResNet50 model. The test accuracy values are
tion of training accuracy results is presented in Table 4. It very low for Adadelta optimization for all proposed models.
is understood that the highest training accuracy of 96.98% The Adadelta optimization technique shows low perfor-
is achieved by MobileNetV2 using RMSProp optimization mance with all proposed models. Hence, this optimization
in all the batch sizes. The second hyperparameter resolution technique is unfit to predict all the Ethiopian currencies.
multiplier of MobileNetV2 is reducing the computational The graph in Figure 9 shows the persistence of test accuracy
cost of a neural network. The results are also presented for all the proposed models and optimization techniques. It
graphically and are shown in Figure 8. As shown in is found that when compared to MobileNetV2, XceptionNet
Journal of Sensors 13

Table 9: Specificity of batch sizes 32, 64, and 128 for classifiers and optimizers.

Model
Optimization Batch size 32 Batch size 64 Batch size 128
IV3 MN XN RN IV3 MN XN RN IV3 MN XN RN
Adam 92.34 90.87 91.43 86.23 95.65 92.36 88.54 92.38 95.32 90.74 89.45 89.29
SGD 89.64 90.34 94.23 92.37 94.47 94.95 90.67 92.93 93.67 90.34 86.67 91.1
RMSProp 93.82 95.81 89.67 92.93 95.84 95.98 92.23 90.28 94.97 96.28 88.24 94.25
Nadam 90.12 94.45 87.34 91.1 95.43 93.93 93.46 89.53 94.34 89.86 89.75 92.36
Adadelta 52.32 89.32 83.95 89.02 65.27 89.72 89.87 93.17 56.93 85.95 86.83 93.75
Adagrad 77.98 94.87 91.37 93.47 87.94 81.23 89.12 93.83 86.26 86.36 90.46 92.64

Specificity of batch sizes 32, 64 and 128 for classifiers and optimizers
98 120
96
100
94
92 80
90
60
88
86 40
84
20
82
80 0
IV3 MN XN RN IV3 MN XN RN IV3 MN XN RN
Batch size 32 Batch size 64 Batch size 128
Adam Nadam
SGD Adadelta
RMSProp Adagrad

Figure 14: Graph of the average specificity.

(a) Raspberry Pi captured and tested the 10-birr note front side

(b) Raspberry Pi captured and tested the 50-birr note front side

Figure 15: Embedded-based hardware system for capturing and testing the sample banknote images.

is showing almost constant values for all optimization tech- 4.3. Train Loss. In this research, we have employed the loss
niques and batch sizes. But XceptionNet is showing less test function as categorical cross-entropy to classify multiple
accuracy compared to MobileNetV2 and ResNet50. classes.
14 Journal of Sensors

Table 10: Banknote image classification test time using embedded-based hardware system.

Model
Optimization Batch size 32 Batch size 64 Batch size 128
MN RN MN RN MN RN
RMSProp 87 s 113 s 105 s 118 s 130 s 145 s

Banknote image classification test time using embedded based

hardware system
140 160

120 140
120
100
100
80
80
60
60
40
40
20 20
0 0
Batch size 32 Batch size 64 Batch size 128
MN
RN

Figure 16: Banknote image classification test time using embedded-based hardware system.

Figure 17: Placement of the banknote on Web-based system to scan.

The MobileNetV2 architecture showing training loss is The MobileNetV2 architecture contains a smaller num-
zero for Adam, RMSProp, and Nadam optimization tech- ber of layers compared to other network models. Hence, it
niques in batch sizes 32, 64, and 128. Compared to other is showing that training loss becomes zero. The Mobile-
models with various optimization techniques and batch NetV2 and ResNet50 showed almost equal training and test
sizes, MobileNetV2 shows low loss values. The ResNet50 accuracy, but the test loss is very high for the ResNet50
showed good training and test accuracy, but it has not because its gradient is too large, so this model is overfitting
achieved zero loss for any optimization techniques shown the data.
in Figure 10.
4.5. Validation Time. Table 7 presents the average process-
4.4. Test Loss. When we increased the number of layers, the ing (testing) time for the Ethiopian bank currency 5-birr,
training and test error rate also increases. From Figure 11, 10-birr, 50-birr, and 100-birr, employed using various
we observed that a 50-layer ResNet, 48-layer InceptionV3, CNN models. To improve the recognition time, each system
and 36-layer XceptionNet CNN give more error rate on both was pretrained before testing on a GPU-based laptop. The
training and testing datasets than MobileNetV2 28-layer time for each set was computed with different classifiers as
CNN architecture. each technique varied because of the architecture’s
Journal of Sensors 15

Figure 18: Capture the snapshot of banknote.

Figure 19: Save the snapshot of banknote images.

Figure 20: Output window of Web-based banknote recognition system.

16 Journal of Sensors

DEFINE FUNCTION predict(filename):

#Step 1
SET my_image TO plt.imread(os.path.join(‘uploads’, filename))
#Step 2
SET my_image_read TO resize(my_image (224,224,1))
#Step 3
CALL #F with graph.as_default():
CALL #F set_session(sess)
SET #Fmodel TO tf.keras.models.load_model(“model_name.h5”)
CALL #F model.compile(optimizer = tf.keras.optimizers.Adam(learning_rate =1e − 05),#‘adam’,
loss = ‘categorical_cross-entropy’,
metrics = [‘accuracy’])
SET #F probabilities TO model.predict(np.array([my_image_read])) [0,:]
CALL #F OUTPUT(probabilities)
#Step 4
SET number_to_class TO [‘fifty’, ‘five’, ‘hundred’, ‘ten’, ‘twohundred’]
SET #F index TO np.argsort(probabilities)
SET predictions TO {“class1”: number_to_class[index [3]],
“class2”:number_to_class[index [2]],
“class3”:number_to_class[index [1]],
“prob1”:probabilities[index [3]],
“prob2”:probabilities[index [2]],
“prob3”:probabilities[index [1]]}
#Step 5
CALL #F RETURN render template(‘predict.html’, predictions = predictions)

Pseudocode 2: Web development.

approach. It shows the validation time for the dataset with 4.7. Hardware Implementation. As we observed the results
each method; the InceptionV3 using Adam optimization from sensitivity and specificity metric values, the Mobile-
technique in batch size 32 had the shortest validation time NetV2 architecture showed the best values. The second-
34.15 seconds. But Inceptionv3 showed less training accu- best performance was showed by ResNet50. So, we designed
racy and test accuracy. The second fastest validation time the embedded-based hardware and tested the MobileNetV2
classifier is the MobileNetV2 model. It is showing 35.09 s. and ResNet50 architecture models in batch sizes 32, 64,
Figure 12 shows the graph of average validation time for and 128. The embedded-based hardware system, i.e., Rasp-
each dataset with each classifier. The InceptionV3 and berry Pi captured the banknote images of 10-birr and 100-
MobileNetV2 had the shortest processing time when optimi- birr using a Web camera. Using Python code called the
zation technique using the Adam and SGD. The Xception- saved architecture model to test the input image, Raspberry
Net had the highest validation time 664.41 seconds on the Pi successfully classified the banknote images and displayed
proposed classification methods. the output on the screen shown in Figure 15. Table 10 shows
the Raspberry Pi time consumption for processing and vali-
4.6. Sensitivity and Specificity. dating the banknotes. From the table, it is clear to under-
stand that the MobileNetV2 model with RMSProp
True Positive optimization technology in batch size 32 shows less time
Sensitivity = : ð1Þ
True Positive + False Negative 87 seconds due to its lightweight architecture. Compared
to all the proposed architecture, MobileNetV2 is having a
The MobileNetV2 model with RMSProp in batch size 32 few parameters which are 4.2 M and a small size of 16 MB.
showed high sensitivity 96.53% compared to remaining The Resnet50 consumes more time for processing and vali-
models shown in Table 8. Figure 13 shows the graph of the dation compared to MobileNetV2 as shown in Figure 16.
proposed model’s sensitivity.
4.8. Web Development. Researchers involved in banknote
recognition and classification are required to possess deep
True Negative
Specificity = : ð2Þ knowledge in fake note detection, banknote classification
True Negative + False Pasitive technologies, and hi-tech production of banknote products.
In such cases, use of a special authenticity verification
The MobileNetV2 model with RMSProp in batch size method is extremely helpful. This Web development-based
128 showed high specificity 96.28% compared to remaining currency recognition system is helpful to users and bank
models shown in Table 9. Figure 14 shows the graph of the employees to verify the Ethiopian currencies (ETB) globally.
proposed model’s specificity. The procedure to recognize the denomination is very simple.
Journal of Sensors 17

If a user or bank employee wants to identify the authentica- Data Availability

tion of currency, they have to keep the banknote under the
camera with good brightness as shown in Figure 17. Then, In this research, we collected the raw images of Ethiopian
click on the “Take Snapshot” button to capture the banknote banknotes from the Cooperative Bank of Oromia, Deka
image as mentioned in Figure 18. Next, save the captured Adi Branch, Adama, Ethiopia.
image from the camera using the “Save Snapshot” button
as shown in Figure 19. After saving the captured image, Conflicts of Interest
the system can authenticate the Ethiopian currency and clas-
sify the denomination. The output window of the banknote The authors of the paper declare that they have no conflict of
image recognition appears as in Figure 20 to show the out- interest or financial conflicts.
come of the Ethiopian banknote recognition and denomina-
tion to be checked [36].
This research article is the part of the preprint published Acknowledgments
in the Research Gate [36]. We must acknowledge the academic resource that we
received from Adama Science and Technology University,
Adama, Ethiopia, giving us a comfortable and active envi-
5. Conclusion ronment for pursuing our research work. We would like to
express special thanks to Mr. Tesfaye Balcha, Branch Man-
In this research, a novel method for distinguishing and clas- ager, and Mrs. Tigist, Technical Assistant, Cooperative Bank
sifying original banknotes from counterfeits (another of Oromia, Deka Adi Branch, Adama, Ethiopia, for all their
denomination) using deep learning technology has been help in the collection of research data.
developed. A prototype for an Ethiopian banknote recogni-
tion system that could be able to classify the various denom-
inations of Ethiopian banknotes has been implemented. References
Nearly about 8500 old Ethiopian paper currency images
have been collected and placed in order to train, test, and [1] M. S. Uddin, P. P. Das, and M. S. A. Roney, “Image-based
validate according to their denominations. For the perfor- approach for the detection of counterfeit banknote of Bangla-
desh,” in 2016 5th International Conference on Informatics,
mance evaluation, the designed framework of the proposed
Electronics and Vision (ICIEV), pp. 1067–1072, Dhaka, Ban-
techniques is trained on four (4) types of convolutional neu- gladesh, 2016.
ral network (CNN) architectures including InceptionV3,
[2] Information Network Security Agency, Addis Ababa, Ethiopia,
MobileNetV2, XceptionNet, and ResNet50 sequentially. J. Fentahun Zeggeye, and Y. Assabie, “Automatic recognition
The proposed architectures fused six different optimization and counterfeit detection of Ethiopian paper currency,” Inter-
techniques, namely, Adam, SGD, RMSProp, Nadam, Ada- national Journal of Image, Graphics and Signal Processing,
delta, and Adagrad, used along all the architectures for the vol. 8, no. 2, pp. 28–36, 2016.
training purposes. For all these models and optimization [3] A. Sheferaw and M. Meshesha, “Ethiopian paper currency rec-
techniques, we have conducted iterations simultaneously ognition system: an optimal feature extraction,” IEEE-SEM,
for 100 epochs. CNN architectures with optimization tech- vol. 7, no. 8, pp. 2320–9151, 2019.
niques programmed using Python language have been [4] M. L. Meharu, Real-time Ethiopian currency recognition for
implemented on Raspberry Pi computer along with an inte- visually disabled peoples using convolutional neural network,
grated Logitech camera. When compared to all the proposed Research Square, 2020.
CNN architectures, the MobileNetV2 architecture with [5] H. Hassanpour, A. Yaseri, and G. Ardeshiri, “Feature extrac-
RMSProp optimization technique in a batch size of 64 tion for paper currency recognition,” in 2007 9th International
showed the best accuracy of about 96.80%. Furthermore, Symposium on Signal Processing and Its Applications, pp. 1–4,
we designed the Web-based program to evaluate the same Sharjah, United Arab Emirates, 2007.
identification of various denominations of Ethiopian curren- [6] N. Kato, M. Suzuki, S. Omachi, H. Aso, and Y. Nemoto, “A
cies (ETB) globally. Using a Web-based design, we have cap- handwritten character recognition system using directional
tured various Ethiopian banknotes and classified the element feature and asymmetric Mahalanobis distance,” IEEE
Transactions on Pattern Analysis and Machine Intelligence,
denominations and identifications successfully. This robust
vol. 21, no. 3, pp. 258–262, 1999.
system is found to be helpful for ATMs for cash deposits
[7] D. G. Pérez and E. B. Corrochano, “Recognition system for
and withdrawals, as well as for financial transactions, bank-
Euro and Mexican banknotes based on deep learning with real
note counters, smart transportation systems, and automatic scene images,” Computación y Sistemas, vol. 22, no. 4,
vending machines, into which money is inserted to purchase pp. 1065–1076, 2018.
goods and automatic smart card charging machines. The [8] H. Aggarwal and P. Kumar, “Indian currency note denomina-
findings reported from this research work are expected to tion recognition in color images,” International Journal on
bring benefits to the development of a fast and efficient Advanced Computer Engineering and Communication Tech-
banknote recognition system using banknote images and nology, vol. 1, no. 1, pp. 12–18, 2014.
contribute to further leading into the development of more [9] P. D. Pawar and S. B. Kale, “Recognition of Indian currency
accurate point-of-care detection tools for Ethiopian bank- note based on HSV parameters,” International Journal of Sci-
note recognition and classification. ence and Research, vol. 3, no. 6, pp. 132–137, 2014.
18 Journal of Sensors

[10] Y. Liang, L. Liu, Y. Xu, Y. Xiang, and B. Zou, “Multi-task gloh Journal of Engineering Research and Technology, vol. 1, no. 3,
feature selection for human age estimation,” in 2011 18th IEEE pp. 1–7, 2012.
International Conference on Image Processing, pp. 565–568, [25] J. Guo, Y. Zhao, and A. Cai, “A reliable method for paper cur-
Brussels, Belgium, 2011. rency recognition based on LBP,” in 2010 2nd IEEE Internatio-
[11] E. Choi, J. Lee, and J. Yoon, “Feature extraction for banknote nalConference on Network Infrastructure and Digital Content,
classification using wavelet transform,” in 18th International pp. 359–363, Beijing, China, 2010.
Conference on Pattern Recognition, pp. 934–937, Hong Kong, [26] S. Gai, G. Yang, and M. Wan, “Employing quaternion wavelet
China, 2006. transform for banknote classification,” Neurocomputing,
[12] CNIB, “Canadian bank note reader,” 1996, http://www.cnib vol. 118, pp. 171–178, 2013.
.ca/en/services/products/bank-note-reader. [27] F. Ahangaryan, T. Mohammadpour, and A. Kianisarkaleh,
[13] D. A. K. S. Gunaratna, N. D. Kodikara, and H. L. Premaratne, “Persian banknote recognition using wavelet and neural net-
“ANN based currency recognition system using compressed work,” in 2012 International Conference on Computer Science
grayscale and application for Sri Lankan currency note- and Electronics Engineering, vol. 3, pp. 679–684, Hangzhou,
SLCRec,” International Journal of Computer, Electrical, Auto- China, 2012.
mation, Control and Information Engineering, vol. 2, no. 9, [28] G. Baykal, U. Demir, I. Shyti, and G. Ünal, “Turkish lira bank-
pp. 235–240, 2008. notes classification using deep convolutional neural networks,”
[14] L. D. Dunai, M. C. Pérez, G. P. Fajarnés, and I. L. Lengua, in 2018 26th Signal Processing and Communications Applica-
“Euro banknote recognition system for blind people,” Sensors, tions Conference (SIU), pp. 1–4, Izmir, Turkey, 2018.
vol. 17, no. 12, pp. 184–197, 2017. [29] J. F. Liu, S. B. Liu, and X. L. Tang, “An algorithm of real-time
[15] K. Verma, B. K. Singh, and A. Agarwal, “Indian currency rec- paper currency recognition,” Journal of Computer Research
ognition based on texture analysis,” in 2011 Nirma University and Development, vol. 40, no. 7, pp. 1057–1061, 2003.
International Conference on Engineering, pp. 1–5, Ahmeda- [30] M. Hasanuzzaman, X. Yang, and Y. Tian, “Robust and effec-
bad, India, 2011. tive component-based banknote recognition for the blind,”
[16] K. N. Oo and A. K. Gopalkrishnan, “Zernike moment based IEEE Transactions on Systems, Man, and Cybernetics, Part C
feature extraction for classification of Myanmar paper curren- (Applications and Reviews), vol. 42, no. 6, pp. 1021–1030, 2012.
cies,” in 2018 18th International Symposium on Communica- [31] J. He, H. Zhang, O. Jin, Z. Hu, and J. Zhang, “A novel recogni-
tions and Information Technologies (ISCIT), pp. 1–6, tion method for Nigeria paper currency based on support vec-
Bangkok, Thailand, 2018. tor machine & genetic algorithms,” International Symposium
[17] O. K. Oyedotun and A. Khashman, “Banknote recognition: on Signal Processing, Biomedical Engineering and Informatics,
investigating processing and cognition framework using com- vol. 6, pp. 1087–1094, 2013.
petitive neural network,” Cognitive NeuroDynamics, vol. 11, [32] R. V. K. Reddy, B. S. Rao, and K. P. Raju, “Handwritten Hindi
no. 1, pp. 67–79, 2017. digits recognition using convolutional neural network with
[18] “Orbit research, iBill, retrieved from http,” 2013, http://www RMSprop optimization,” in 2018 Second International Confer-
.orbitresearch.com/ibill_details.php. ence on Intelligent Computing and Control Systems (ICICCS),
[19] F. Takeda, L. Sakoobunthu, and H. Satou, “Thai banknote pp. 45–51, Madurai, India, 2018.
recognition using neural network and continues learning by [33] C. L. Huang, Y. C. Shih, and C. M. Lai, “Optimization of a con-
DSP unit,” in International Conference on Knowledge-Based volutional neural network using a hybrid algorithm,” Interna-
and Intelligent Information and Engineering Systems: tional Joint Conference on Neural Networks, vol. 1, pp. 14–19,
Knowledge-Based Intelligent Information and Engineering Sys- 2019.
tems, Lecture Notes in Computer Science, pp. 1169–1177, [34] A. M. Taqi, F. Al-Azzo, and M. Milanova, “The impact of
Springer, 2003. multi-optimizers and data augmentation on TensorFlow con-
[20] A. Ahmadi, S. Omatu, T. Fujinaka, and T. Kosaka, “Improve- volutional neural network performance,” in 2018 IEEE Confer-
ment of reliability in banknote classification using reject ence on Multimedia Information Processing and Retrieval
option and local PCA,” Information Sciences, vol. 168, no. 1- (MIPR), pp. 140–145, Miami, FL, USA, 2018.
4, pp. 277–293, 2004. [35] M. R. Rani, M. Z. C. Mustafar, N. H. F. Ismail, M. S. F. Mansor,
[21] T. Reiff and P. Sincak, “Multi-agent sophisticated system for and Z. Zainuddin, “Road peculiarities detection using deep
intelligent technologies,” in 2008 IEEE International Confer- learning for vehicle vision system,” IOP Conference Series:
ence on Computational Cybernetics, pp. 37–40, Stara Lesna, Materials Science and Engineering, vol. 1068, pp. 1–10, 2021.
Slovakia, 2008. [36] H. Kalla and D. T. Aseffa, Ethiopian Banknotes Recognition
[22] M. E. Ayalew Tessfaw, M. B. Ramani, and M. T. Kebede Using Convolutional Neural Network, ResearchGate GmbH,
Bahiru, “Ethiopian banknote recognition and fake detection 2021.
using support vector machine,” in 2018 Second International
Conference on Inventive Communication and Computational
Technologies (ICICCT), pp. 1354–1359, Coimbatore, India,
2018.
[23] A. Rashid, A. Prati, and R. Cucchiara, “On the design of
embedded solutions to banknote recognition,” Optical Engi-
neering, vol. 52, no. 9, p. 093106, 2013.
[24] B. V. Chetan and P. A. Vijaya, “A robust side invariant tech-
nique of Indian paper currency recognition,” International