International Journal of

Intelligent Systems and Applications in Engineering

ISSN:2147-67992147-6799 [Link] Original Research Paper

Banknote Classification Using Artificial Neural Network Approach

Esra Kaya *1, Ali Yasar 2, Ismail Saritas 1

Accepted 10th August 2015 DOI: 10.18201/ijisae.5525010.1039/b000000x

Abstract: In this study, clustering process has been performed using artificial neural network (ANN) approach on the pictures belonging
to our dataset to determine if the banknotes are genuine or counterfeit. Four input parameters, one hidden layer with 10 neurons and one
output has been used for the ANN. All of these parameters were real-valued continuous. Data were extracted from images that were taken
from genuine and forged banknote-like specimens. For digitization, an industrial camera usually used for print inspection was used. The
final images have 400x 400 pixels. Due to the object lens and distance to the investigated object gray-scale pictures with a resolution of
about 660 dpi were gained. Wavelet Transform tool were used to extract features from images. Four input parameters are processed in the
hidden layer with 10 neurons and the output realizes the clustering process. The classification process of 1372 unit data by using ANN
approach is sure to be a success as much as the actual data set. The regression results of the clustering process is considerably well. It is
determined that the training regression is 0,99914, testing regression is 0,99786 and the validation regression is 0,9953, respectively. Based
on the results obtained, it is seen that classification process using ANN is capable of achieving outstanding success.

Keywords: ANN, Banknote, Classification, Machine Learning Database

study.
1. Introduction
Image processing is changing the image and the features of a 2.1. Artificial Neural Network:
photo, which is converted form real life image to a digital image, ANN is a system which is modelled as an inspiration of biological
and creating a new photo. Due to the conversion of the photo from neural network but with a simpler structure. The main feature of
analog media to digital media, the photo contains noise. Image these systems is that they have fully parallel, adaptive, learning and
processing can be used to remove the noise from the photo. In this parallel distributed memories[3][4].
study clustering process has been performed using ANN[1]. It is ANN is composed of processor components connected with
possible to detect the counterfeit money using banknote neurons. Generally, it consists of three layers, i.e. an input layer,
classification. For banknote classification process (BC), one or more hidden layers and an output layer. Each layer has a
classification is realized according to banknote authentication data certain number of components attached to one another called
set which can be found in UCI Machine Learning Repository neurons or nodes. Each neuron in the input layer is connected to
(center for Machine Learning and Intelligent Systems)[2]. In this each neuron in the intermediate(hidden) layer. Also, each neuron
study, our purpose is to perform the classification process in the hidden layer is connected to each neuron in the output layer.
according to the information that belongs to our data set using a Each of the neurons is connected to the other with weights and
data mining method, ANN. accompanying communication networks. ANN is trained with
sample data to learn the relationship between inputs and outputs.
2. Proposed Methodology The process of learning and training is performed by varying
weight values of the connections between neurons[5][6]. Signals
In this study, the data from banknote authentication data set is used move through neurons over weights. Each neuron receives
which can be found in UCI Machine Learning Repository (center multiple inputs from other neurons depending on their weights and
for Machine Learning and Intelligent Systems). With the given generates an output signal that may also be generated by other
data, the variance of Wavelet Transformed image (V) {-7,04 neurons[7][8][9]. In our system, randomly chosen 206 data has
6,82}, skewness of Wavelet Transformed image (S){-13,77 been used as testing data, randomly chosen 206 data from the
12,95}, curtosis of Wavelet Transformed image (C) {-5,28 17,92} remaining data has been used as validation data and the remaining
, entropy of image (E) {-8,54 2,44}, the banknote classification 960 data has been used as training data.
has been realized using ANN. The ANN model which forms our system is shown below, in Fig.
Neural Network Toolbox of Matlab R2013b has been used for this 1 and as it is shown in the figure, our neural network consists of
_______________________________________________________________________________________________________________________________________________________________
total 5 inputs as of 4 + 1 inputs, one hidden layer of 10 neurons and
1
Selcuk University,Konya – 42075, Turkey
2
Selcuk University,Konya – 42075, Turkey one output.
3
Gneysinir Vocational School of Higher Education, Konya – 42190,
Turkey
* Corresponding Author:Email: esrakaya@[Link]
#
This paper has been presented at the International Conference on
Advanced Technology&Sciences (ICAT'15) held in Antalya (Turkey),
August 4-7, 2015.
International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2016, 4(1), 16-19 |16
 Table 1. The parameters and properties used in ANN
Parameters Properties
Number of neurons in the input
layer 4
Number of the hidden layers 1
Number of neurons in the hidden
layer 10
Number of neurons in the output
layer 1
Learning rate (α) 0,4
Coefficient of momentum (β) 0,3
Learning algorithm Gradient descent (traingd)
Logarithmic sigmoid
Transfer function (logsig)
The regression graph of the training data set is shown in Figure.2.

Figure 1. The structure of ANN

B. Artificial Neural Network Application For A Data Set

In this study, the following transactions are performed for

determining if the banknotes are counterfeit or not using the data
set of 1,372 units of banknote samples.
 As testing data, 206 units (15%) of data have been
chosen randomly.
 As validation data, 206 units (15%) of data have been
chosen randomly.
 As training data, the remaining 960 units (70%) of data
have been chosen.
In this study, a feed forward network structure that contains an
input layer, a hidden layer and an output layer (Fig. 1) was used.
After the ANN structure was designed, the data obtained in the
Figure 2. Regression of Training Data Set
experimental study were normalized in the 0-1 value set using Eq.
1 in order to improve the characteristics of the training. The Back The regression graph of the testing data set is shown in Figure.3.
Propagation algorithm was used in the training procedure.
Different transfer functions (Purelin, Tansig, Logsig etc.) were
used and tried in the neurons in the hidden and output layers and
(Tansig) was selected as the transfer function that yielded the best
result.
x  xmin
xnorm  (1)
xmax  xmin

The training data set was used to determine ANN neuron and bias
weight values. Training was repeated to obtain the lowest level of
error by changing the number of neurons and the epoch number.
Then, the trained algorithm was applied on the test data set.
The network has been trained by giving different values trough the
hidden layer and the results have been observed. It has been seen
that better results were obtained using a network with a hidden
layer of 10 neurons.
At the end of these procedures, the network structure that yielded
the best classification is given in Table 1.
Figure 3. Regression of Training Data Set

The regression graph of the validation data set is shown in Fig. 4.

IJISAE, 2016,4(1), 16-19 International Journal of Intelligent Systems and Applications in Engineering
 different than the data used for training process.

Dataset Results/ ANN Predicts

1

0,8

0,6

0,4

0,2

1
14
27
40
53
66
79
92
105
118
131
144
157
170
183
196
ANN Target_Val
Figure 4. Regression of Training Data Set

The regression graph of the output is shown in Figure 5. Figure 7. Validation Result / ANN Predict

Finally, the graph of the values obtained from the remaining data
which is the estimation of ANN is shown in Figure.8.
Dataset Results/ ANN Predicts

0,8

0,6

0,4

0,2

691
1
70
139
208
277
346
415
484
553
622

760
829
898
Figure 5. Regression of Output Set ANN Target_train

3. Result and Discussion Figure 8. Train Result / ANN Predict

The comparison of experimental measurement values for BC test As seen in the figures, the estimation results and dataset results are
data set and ANN estimation values are shown in Figure.6. All the almost overlapping. The deviation between experimental and
randomly selected data used for test is different than the data used estimated results is very small and negligible for any BC
for training process. performance.

Dataset Result/ ANN Predict 4. Conclusions

In this study, it has been concluded that ANN can be used in the
1
clustering procedure and better results can be achieved. According
0,8 to the results, it has been observed that 99% and above
0,6 classification accuracy can be achieved
0,4
ACKNOWLEDGEMENTS
0,2
Contributors gratefully acknowledge support of their work by the
0
Institute of Agrophysics of the Polish Academy of Sciences in
92
1
14
27
40
53
66
79

105
118
131
144
157
170
183
196

Lublin and [Link] .

ANN Target_Test
5. References
Figure 6. Data Set Result / ANN Predict [1] [Link]
[2] [Link]
The comparison of experimental measurement values for BC [3] Yüksel Özbay, “EKG Aritmilerini Hızlı Tanıma”, Doktora Tezi,
1999
validation data set and ANN estimation values are shown in [4] Ömer KELEŞOĞLU, Cevdet Emin EKİNCİ, Adem FIRAT, The
Figure.7. All the randomly selected data used for validation is Using Of Artificial Neural Networks In Insulation Computations,

International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2016, 4(1), 16-19 |18
 Journal of Engineering and Natural Sciences Mühendislik ve Fen [8] Lorenz, A., Blum, M., Ermert, H., & Senge, Th. (1997).
Bilimleri Dergisi Sigma/2005-3 Comparison of Different Neuro-Fuzzy Classification Systems for
[5] Çolak,C , Çolak M.C, Atıcı M.A, “Ateroskleroz’un Tahmini İçin the Detection of Prostate Cancer in Ultrasonic Images.
Bir Yapay Snir Ağı” 2005. Ultrasonics Symposium, 2, 1201-1204.
[Link] [9] Ronco, A.L., & Fernandez, R. (1999). Improving
[6] Yao X. Evolving Artificial Networks, Proceeding of the Iee Ultrasonographic Diagnosis of Prostate Cancer with Neural
1999;87:1423-44. Networks. Ultrasound in Med. & Biol., vol. 25, no. 5, pp. 729–
[7] Cinar, M., Engin, M., Engin, E.Z., & Ates, Y.Z. (2009). Early 733.
Prostate Cancer Diagnosis by Using Artificial Neural Networks.
Expert Systems with Applications, 6357–6361.

IJISAE, 2016,4(1), 16-19 International Journal of Intelligent Systems and Applications in Engineering