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

net/publication/297650409

Automatic Recognition and Counterfeit Detection of Ethiopian Paper Currency

Article in International Journal of Image, Graphics and Signal Processing · February 2016
DOI: 10.5815/ijigsp.2016.02.04

CITATIONS READS

16 4,165

2 authors, including:

Yaregal Assabie
Addis Ababa University
52 PUBLICATIONS 309 CITATIONS

SEE PROFILE

All content following this page was uploaded by Yaregal Assabie on 22 September 2016.

The user has requested enhancement of the downloaded file.

I.J. Image, Graphics and Signal Processing, 2016, 2, 28-36
Published Online February 2016 in MECS (http://www.mecs-press.org/)
DOI: 10.5815/ijigsp.2016.02.04

Automatic Recognition and Counterfeit Detection

of Ethiopian Paper Currency
Jegnaw Fentahun Zeggeye
Information Network Security Agency, Addis Ababa, Ethiopia
Email: [email protected]

Yaregal Assabie
Department of Computer Science, Addis Ababa University, Addis Ababa, Ethiopia
Email:[email protected]

Abstract—Currency recognition is a technology used to recognition is an image processing technology that is used
identify currencies of various countries. The use of to identify currency of various countries. In general,
automatic methods of currency recognition has been currency recognition systems usually have four major
increasing due its importance in many sectors such as processes: image acquisition, preprocessing, feature
vending machine, railway ticket counter, banking system, extraction and classification [1, 4, 19, 21]. Image
shopping mall, currency exchange service, etc. This paper acquisition is an activity of acquiring the currency image
describes the design of automatic recognition of in a digital form using a specific hardware designed for
Ethiopian currency. In this work, we propose hardware this purpose, usually scanner or camera. Image acquisition
and software solutions which take images of an Ethiopian is followed by preprocessing which is required to improve
currency from a scanner and camera as an input. We the currency image by suppressing unnecessary distortions
combined characteristic features of currency and local or enhancing major image features important for further
feature descriptors to design a four level classifier. The recognition processes. Feature extraction is a type of
design has a categorization component, which is dimensionality reduction that efficiently represents
responsible to denominate the currency notes into their interesting parts of the currency image as a compact
respective denomination and verification component feature vector. The types of features used for recognition
which is responsible to validate whether the currency is purpose play crucial role in currency recognition systems.
genuine or not. The system is tested using genuine There are mainly two types of features: structural feature
Ethiopian currencies, counterfeit Ethiopian currencies and statistical feature. Structural feature describes
and other countries’ currencies. The denomination geometrical and topological characteristics of pattern by
accuracy for genuine Ethiopian currency, counterfeit representing its global and local properties whereas
currencies and other countries’ currencies is found to be statistical feature describes characteristic measurements of
90.42%, 83.3% and 100% respectively. The verification the pattern. Some of the features commonly used for
accuracy of our system is 96.13%. currency recognition are Scale-Invariant Feature
Transform, Binary Robust Invariant Scalable Key points,
Index Terms—Ethiopian Currency, Currency Speeded Up Robust Feature, etc. [3, 7, 11, 16, 22].
Recognition, Counterfeit Detection. Classification is the final phase of the currecny
recognition process where the decision about a given
currency image is made based on the extracted feature sets.
I. INTRODUCTION The decision could be rejection or identification of the
currency image into one of the valid banknotes of a given
Currency is used almost everywhere in the world to
country.
facilitate business. Although electronic commerce is
The aim of currency recognition system is to help
emerging as an alternative means of making business
machines recognize different currencies and to help
without the use of paper currency, it is still an
people work with convenience and efficiency. An efficient
indispensable part of everyone’s daily routine. There are
currency recognition system is vital for automation in
various currencies all over the world, and each of them
many sectors such as vending machine, railway ticket
looks totally different and unique. This difference can be
counter, banking system, shopping mall, currency
due to the size of the paper currency, or color or pattern
exchange service, etc. Automatic machines capable of
drawn on the paper. Manual identification and counterfeit
recognizing banknotes are currently used in the developed
detection of currencies a difficult task requiring specially
part of the world in automatic dispensers of a number of
trained examiners which requires human labor and time.
different products ranging from cigarettes to bus tickets,
Since currency has great importance in facilitating the
as well as in many automatic banking operations [5, 13,
management of the economy of countries at large,
17]. Some other countries have been attempting to
automatic recognition of currency becomes a great area of
develop systems that recognize their respective currencies
interest for researchers and developers [17]. Currency

Copyright © 2016 MECS I.J. Image, Graphics and Signal Processing, 2016, 2, 28-36
 Automatic Recognition and Counterfeit Detection of Ethiopian Paper Currency 29

[9, 10, 12, 14, 15, 18]. To the best of our knowledge, of US Dollars. The PCA algorithm is used to extract the
however, there is no research and development work main features of US banknotes and to reduce the size of
focused on recognition of Ethiopian paper currency which data. Linear Vector Quantization (LVQ) network is
is known as Birr. Thus, the purpose of this work is to applied for classification purpose. On the other hand,
design and develop a system that can identify Ethiopian Kagehiro et al. [13] take a number of discrete points
paper currency and categorize them in their denominations. selected from the entire image where the average of the
In addition, the proposed solution identifies counterfeit pixel at each point and its adjacent pixels is taken as the
currency from genuine Ethiopian currencies. The study observed value for each point. Then, the distance between
considers banknotes that are currently functional, i.e., 1 a template vectors and the feature vectors from the
Birr, 5 Birr, 10 Birr, 50 Birr and 100 Birr. observation points is used for classification. The authors
The remaining part of this paper is organized as follows. reported very low errors rates with high-speed processing.
Section II discusses related works. In Section III, we Aoba et al. [2] propose two types of neural networks for
present the proposed Ethiopian currency recognition recognition of Euro banknotes: a three-layered perceptron
system. Experimental results are reported in Section IV. and several Radial Basis Function (RBF) networks. They
Section V presents conclusion and future works. employ three-layered perceptron for classification of
banknotes whereas RBF network is used to reject invalid
data by estimating the probability distribution of the
II. RELATED WORK sample data. Since Euro banknotes have characteristic
features in IR images, the system uses infra-red (IR) and
The recognition of banknotes has been addressed in visible images as input data.
several ways depending on their types. Accordingly,
Verma et al. [18] propose an Indian currency
various methods have been proposed for recognition of recognition system based on the intrinsic feature of the
different banknotes. The commonly used methods include currency. The texture of the currency is extracted as
artificial neural network, currency characteristic intrinsic features fur currency recognition, and the class
comparison, principal component analysis, local feature discriminating capability of these features for Indian
descriptors comparison, hidden Markov models, Naï ve
currency is evaluated. On the other hand, Gogoi et al. [9]
Bayes classifier, etc. The selection of methods applied to classify Indian currency based on a set of unique features
recognize currencies depends on the characteristic features such as color, dimension and identification mark. Since
of the currencies.
each banknote has unique shape as identification mark,
Debnath et al. [6] used ensemble neural network to Fourier descriptor is used to extract the features. The,
recognize Bangladeshi currency where individual neural classification is made using artificial neural networks.
networks in an ensemble neural network are trained via
Liu et al. [15] develop real-time Chinese currency
negative correlation learning (NCL). NCL is used to recognition system using a combination of Haar wavelets
expertise the individuals on different parts or portion of and Support Vector Machine. The idea of this method is to
input patterns in an ensemble. The input patterns used in
extract the waveform features and transform them to the
this work are grayscale values of the currency image. The Support Vectors with fixed length. The waveform of a
decision of ensemble is computed using voting scheme banknote is a periodic signal that contains many features
where each network gives a vote for a certain class and
and details with varying scales. The Haar features are
classification is done by the winning neuron of that classified by making use of Support Vector Machine. The
network. The authors reported that they obtained authors reported that the recognition system is fast and
promising results even for noisy Bangladeshi currency
robust with high accuracy. Multiple-kernel Support Vector
images. Machines is also used to successfully detect counterfeit
Colors and texture features have been exploited for Taiwanese banknotes by minimizing false recognition
recognition of banknotes. Garcí a-Lamont et al. [7] rates [20]. In this method, each banknote is divided into
employ a method to recognize Mexican banknotes by partitions and each partition is associated with its own
using artificial vision, where the color and texture features
kernels. Then, the luminance histograms of the partitions
of the banknotes are extracted to classify them. The color are taken as the input of the system.
and texture features are extracted with the RGB space and The reviewed papers present different methods to
the Local Binary Patterns, respectively. The authors report
recognize banknotes and/or detect counterfeits. It can be
high denomination accuracies due to the fact Mexico, like seen that the banknotes of various countries have
many other countries, employs colors to recognize the characteristic features on which recognition systems rely.
denominations of the banknotes. Similarly, Yan et al. [19]
Accordingly, different classifiers are proposed based on
use color and texture features for recognition of US the types of features and the complexity of the problem.
banknotes in which the features are further processed and Due to such feature variations, none of the methods
classified by a Feed-forward Neural Network (FNN). The
guarantee us to work effectively across all currencies.
output of the system is a measurement of the similarity Thus, in this paper, we propose four different features
between existing samples and test banknote. (based on color and characteristic features) for recognition
Principal Component Analysis (PCA) is employed by
and counterfeit detection of Ethiopian banknotes.
Ahmadi et al. [1] for increasing the reliability of banknote
recognition machines. They define a method for validating
the reliability where they examine six different bill types

Copyright © 2016 MECS I.J. Image, Graphics and Signal Processing, 2016, 2, 28-36
30 Automatic Recognition and Counterfeit Detection of Ethiopian Paper Currency

and in English. These features are used to manually check

III. THE PROPOSED SYSTEM if the currencies are genuine and a digital camera captures
such features. Thus, recognition of Ethiopian currencies
A. System Architecture requires scanner and camera as hardware solutions to
We propose hardware and software solutions for capture characteristic features relevant for denomination
recognition of Ethiopian paper currency. The Ethiopian and verification purposes, respectively. Images of genuine
paper currency recognition system consists of three major 100 Birr banknotes captured with camera and scanner are
components: currency image acquisition, currency shown in Fig. 2.
denomination and currency verification. Currency image
acquisition is required to capture the images of the Thin golden strip Wide golden strip
currency using scanner and camera. Currency
denomination component is responsible for classifying the
paper currency into the respective denomination based on
the image of the currency captured by scanner whereas
currency verification checks whether specific paper
currency is genuine or counterfeit based on the image of
currency captured by a camera. Currency image
acquisition attributes hardware solution to the problem as
we use this component to attenuate characteristic features
of Ethiopian banknotes by making use of scanner and
camera. On the other hand, currency denomination and
verification components are software solutions to the
problem. The architecture of the proposed Ethiopian paper
currency recognition system is shown in Fig. 1.

Currency Image Acquisition

Scanner Camera

Preprocessing Preprocessing Fig.2. Banknote of 100 Birr captured with camera (top) and
scanner (bottom)

Feature Region of Interest C. Currency Denomination

Extraction Extraction
The denomination of a given test image into one of the
five Ethiopian banknotes (1, 5, 10, 50 and 100 Birr) is
Currency Verification
Classification Validation achieved by preprocessing the image, extracting
characteristic features from the image and classifying the
input image based on the features. Classification is made
Currency Currency by comparing the features of input image against standard
Denomination Verification sample images of banknotes stored as a database.

Result a) Preprocessing
In the preprocessing sub-component, we apply image
Fig.1. Architecture of the proposed system. re-sizing and enhancement with the purpose of preparing
the paper currency image to the main image processing
B. Currency Image Acquistion activities. All the five Ethiopian banknotes have different
The image acquisition component is responsible for sizes. However, size is not used as a feature for
acquiring images of paper currencies. Each of the classification in the proposed design and the size of the
Ethiopian banknotes have distinctive range of colors, sizes, image needs to be similar for comparing the correlation
looks, marks, texts and other characteristics that make coefficients of the color components among different
them unique for identification. Since these features are banknotes. Accordingly, the images of the banknotes are
visible under normal light conditions, they can be captured brought to a common size. On the other hand, the quality
by scanner. When seen against white light, Ethiopian of the images is degraded for old banknotes. To enhance
banknotes are also characterized by having thin golden the quality of such input images, we filter them using a
strip (for 5, 10, 50 and 100 Birr banknotes) and wide Gaussian filter of window 3x3.
golden strip (for 50 and 100 Birr banknotes) bearing the
name and sign of National Bank of Ethiopia in Amharic

Copyright © 2016 MECS I.J. Image, Graphics and Signal Processing, 2016, 2, 28-36
 Automatic Recognition and Counterfeit Detection of Ethiopian Paper Currency 31

b) Feature Extraction and Classification Table 1. Correlation Between Input And Standard Images

Feature extraction and classification are the main parts

Standard currency images
of currency denomination component. Four different
feature sets are selected for the classification of the Correlation
banknotes. These feature sets are dominant color of the 1 Birr 5 Birr 10 Birr 50 Birr 100 Birr
image, the distribution of the dominant color, the hue
R G B R G B R G B R G B R G B
value of the image and Speeded Up Robust Feature
(SURF) of the banknotes. Based on these features, the
R x x x

Dominant color
classification of input images into their respective

of input image
denomination is carried out. The classification process is
applied after each feature is extracted. Because of this, the G x x
classifier is called a four level classifier. At each stage, a
currency image can be a candidate to be classified into one B x
of the five Ethiopian banknotes but needs to pass through
the four phases as discussed below.
Phase 1 (Dominant color): The image is acquired in The correlation coefficient (CCB) between the dominant
RGB format and it has three major components; Red (R), color values of an image A and the corresponding color
Green (G) and Blue (B). The composition of these three values of a candidate banknote B is computed as:
components gives the specific color of the currency image
[8]. In each image either red or green or blue is the  ( A mn  A)( Bmn  B ) (1)
dominating color, i.e., the maximum pixel occupancy in CCB  m n

 2  2
  ( Amn  A)   ( Bmn  B ) 
the image. The five Ethiopian banknotes have distinct
colors for human vision system. Each color of the  m n  m n 
banknotes is different because the combination of the
primary colors (Red, Green and Blue) is different in each where m and n are pixel coordinates, Amn is the pixel value
currency. In addition to this, one of the three primary of the dominant color of the image, and A is mean of A.
colors has greater pixel occupancy in the image. To find Empirically, it is found that new Ethiopian banknotes have
the maximum pixel occupancy, the respective color correlation coefficient of more than 1.0 and degraded
components of the pixels are added for the entire image, Ethiopian banknotes have more than 0.3. Therefore, to
which will give three results in accordance with the three include degraded Ethiopian currencies a threshold value
primary colors. Among the three results, the one with (T1) of0.3 was taken. In order for the test image to
maximum pixel sum is considered to be the dominant. We continue as a candidate into the respective class or
empirically made dominant color analysis on samples of denomination, the correlation coefficient needs to be
Ethiopian banknotes and it was revealed that: greater than this threshold value. Otherwise, the input
image would be considered as counterfeit or non-
 Red is dominant color for 1 Birr, 10 Birr and 50 Birr; Ethiopian banknote.
 Green is dominant color for 5 Birr and 100 Birr; and Phase 3 (Hue value): The HSV color model, also
 Blue is dominant color for 5 Birr. called HSB (Hue, Saturation, Brightness), defines a color
space in terms of three constituent components: hue,
Accordingly, the input image is classified based on the saturation and value. Hue is the color type (such as red,
analysis of the dominant color observed in the image. At magenta, blue, cyan, green or yellow); saturation refers to
this level we are still classifying the input image into the intensity of specific hue; and value refers to the
intermediate results and a final decision is yet to be made brightness of the color [8]. As hue varies from 0 to 1.0, the
after passing through the remaining phases. corresponding colors vary from red through yellow, green,
Phase 2 (Distribution of dominant color): The cyan, blue, magenta, and back to red, so that there are
distribution of the dominant color of the input image actually red values both at 0 and 1.0. As saturation varies
needs to be compared with the corresponding dominant from 0 to 1.0, the corresponding colors (hues) vary from
colors of candidate banknotes identified in Phase 1. This is unsaturated (shades of grey) to fully saturated (no white
achieved by computing the correlation coefficient (CC) of component). As value, or brightness, varies from 0 to 1.0,
the corresponding color values. For example, if the the corresponding colors become increasingly brighter [8].
dominant color of the image is red, then we compute the Among the three components, hue component is selected
correlation between red component of the input image and to compare the different currency images because the
the red components of 1 Birr, 10 Birr and 50 Birr, other two components could be easily affected by the
separately. Table 1 shows the dominant color correlations currency degradation and light effect while the image is
to be computed between input image and the standard scanned or captured. In addition to that, each image has a
sample banknotes. specific hue value.

Copyright © 2016 MECS I.J. Image, Graphics and Signal Processing, 2016, 2, 28-36
32 Automatic Recognition and Counterfeit Detection of Ethiopian Paper Currency

In the proposed system, the hue value of the test image either of the features are counted and the ratio of these
is compared with each class of currency image’s hue value. two values is taken. An empirical analysis revealed
Since the test image is classified in the second that the match percent for genuine Ethiopian
classification step, the denomination of the test image is currencies is found to be more than 30. Accordingly,
known at this stage. Therefore, the hue value of the test we set a threshold value (T3) of 30 for match percent
image is compared against the hue value of a standard (MP). Thus, if the match percent between an input
image of its class. Empirically, it is found that the hue image and a standard image of the same denomination
difference (HD) of a genuine Ethiopian paper currency is greater than 30, it is classified as an Ethiopian
and the standard Ethiopian currency of its type is not more currency with the identified denomination. Otherwise,
than 75. Therefore, a threshold value (T2) of 75 is used to it is non-Ethiopian currency or counterfeit.
continue analyzing the input image or reject as counterfeit.
If the hue difference between the test image and the
standard image is less than 75, that test image is still
considered as a candidate of an Ethiopian currency in that
specific denomination. Otherwise, it is considered as a
counterfeit or non-Ethiopian paper currency.
Phase 4 (SURF matching): In this phase, the SURF
features of the input image are extracted for a final
decision making on the classes of denominations. The
feature extraction and classification process is
accomplished using the following five steps.

i. Finding the interest points or SURF points: We are

required to compute interest points called SURF points.
The SURF approximates the second order Gaussian
derivate with box filters (mean or average filter),
which is able to be calculated fast through integral
images. The localization of interest point is determined
by the determinant of Hessian matrix. So, interest
points are finally localized in scale space and image
space by using non-maximum suppression in their
3×3×3 neighborhood [3, 11]. Thousands of SURF
points are detected in each of the five banknotes.
ii. Selecting strongest points: The strongest points need
to be selected to make the system computationally
efficient. Constructing feature vectors and matching
thousands of feature points is expensive. Thus, feature
points having strong metric or more relevant features
need to be selected. By taking the first 100 strongest
points, we are able to detect some of the major features
which can differentiate the five Ethiopian currencies.
Among these features included in the first strongest
points are letter and number denomination values
written in English and Amharic, alphanumeric note
number and currency identification sign for visual
impaired people. Figure 3 shows 100 strongest SURF
points detected for each Ethiopian banknote.
iii. Constructing feature vectors or descriptors: In this
step, the feature vectors or descriptors from the
selected strongest points are constructed and
descriptors from a region around each interest point
are extracted.
iv. Match features: After the feature vectors are
constructed, the relatedness of the feature vector of an
input currency image and a standard currency image of
that specific denomination is checked. An indexed
structure that contains matching features is built and
stored for further processing.
v. Calculating match percentage: To calculate the match
percentage, the matched points and the valid points in Fig.3. SURF points detected on Ethiopian banknotes (from top to
bottom: 1, 5, 10, 50 and 100 Birr.)

Copyright © 2016 MECS I.J. Image, Graphics and Signal Processing, 2016, 2, 28-36
 Automatic Recognition and Counterfeit Detection of Ethiopian Paper Currency 33

The four levels of feature extraction and classification paper denominations. In this process, non-Ethiopian
processes of currency denomination component are currencies or counterfeit banknotes with major deviation
illustrated in Fig. 4. Once the final decision is made on the from genuine Ethiopian banknotes are rejected as
denomination of a given input image, verification process counterfeits. However, it is likely that banknotes produced
will continue to check if the classified image is genuine or with the purpose of counterfeiting Ethiopian currencies
counterfeit. will be classified into one of the five denominations. Thus,
results of currency denomination component are further
D. Currency Verification processed to check if the denomination is genuine or not.
The task of currency denomination component is to In Ethiopian paper currencies, there are few features that
classify an input image into one of the five Ethiopian distinguish the genuine currency from the counterfeit. The

Preprocessed currency image

Key Feature Extraction
Get Dominant Color
Feature Extraction

Classification Classification
No Dominant color=B? No Dominant color=G?
Yes Yes

Feature Extraction
Get Correlation of R Get Correlation of B Get Correlation of G

Classification
No Yes
CC1 > CC50?

No No No Yes
CC10 > CC50? CC1 > CC10? CC100 > CC5?
Yes Yes

CC50> T1? No CC10> T1? No No CC5> T1? No No

CC1> T1? CC100> T1?
Yes Yes Yes Yes Yes

Feature Extraction
Get Hue Difference Get Hue Difference Get Hue Difference Get Hue Difference Get Hue Difference

Classification

HD< T2? No HD< T2? No No HD< T2? No No

HD< T2? HD< T2?
Yes Yes Yes Yes Yes

Feature Extraction
SURF Matching SURF Matching SURF Matching SURF Matching SURF Matching

Classification
No No No No No
MP> T3? MP > T3? MP> T3? MP> T3? MP> T3?
Yes Yes Yes Yes Yes

50 Birr 10 Birr 1 Birr 5 Birr 100 Birr

Counterfeit
To Currency Verification Component
Fig.4. Flowchart of feature extraction and currency classification.

Copyright © 2016 MECS I.J. Image, Graphics and Signal Processing, 2016, 2, 28-36
34 Automatic Recognition and Counterfeit Detection of Ethiopian Paper Currency

most outstanding feature in all denominations, except 1 The objective of binarization is to mark pixels that belong
Birr banknote, is the broken thin golden strip, which to true foreground regions with a single intensity and
becomes solid line when it is seen from the direction of background regions with different intensities. Through
background light. In addition to this thin golden strip, in experiment, it is found that for genuine Ethiopian currency,
50 Birr and 100 Birr, there is a wider golden strip which a region with pixel value less than 30 is noise. Therefore,
bears the name and sign of National Bank of Ethiopia in binarization is applied with a pixel intensity threshold of
Amharic and in English. These characteristic features are 30. This operation considers pixel intensities less than 30
captured by a digital camera which helps to apply as noise and therefore removes those pixels.
verification of the classified banknotes. In addition to the Morphological area opening: Morphological area
preprocessing step, the verification component has two opening removes from a binary image all connected
major processes: region of interest (ROI) extraction and components (objects) that have less than a given threshold
validation. However, the techniques we apply on these of pixels, producing another binary image. It is assumed
processes differ due to the difference in the features of that those objects which have less area than the given
banknotes. threshold of pixels are noises that need to be removed
from the ROI. In the proposed system, an area opening by
a) Verification by the Thin Strip
100-pixel object is done.
The thin golden strip exists in all Ethiopian banknotes, Morphological closing: Morphological closing of an
except 1 Birr banknote. The 1 Birr banknote is assumed to image is dilation followed by erosion, using the same
be genuine if it is correctly classified in the currency structuring element for both operations. In the proposed
denomination component. For the rest of paper banknotes, system, a morphological closing with square structural
ROI is now considered to be the thin golden strip. The element of size 5 symmetric with respect to the
localization of region of interest and validation processes neighborhood centre is applied. This size of the structural
are done using the following steps. element is set through experiment.
Counting the number of objects: If the number of
ROI localization: The position of the thin strip is not in objects is one (representing wide golden strip), it is a
a specific location for currencies in a specific genuine currency. Otherwise, it is a counterfeit.
denomination. Therefore, the existence of the thin strip is
checked by taking a strip iteratively.
Binarization: For better processing, the RGB is IV. EXPERIMENT
converted to HSV color format. From the three HSV
components, only saturation and value components are A. Dataset Collection
usually exposed to noise. And therefore binarization is For evaluating and testing the proposed prototype,
applied on these two components of the image with 0.3 different denomination of Ethiopian and non-Ethiopian
value and 0.4 saturation thresholds. This threshold value is currencies are collected. Genuine sixty 100 Birr notes,
set through experiment. sixty 50 Birr notes, forty 10 Birr notes, forty 5 Birr notes,
Morphological closing: A morphological closing with forty 1 Birr notes are collected. In these notes new, a little
a line structural element symmetric with respect to the old, very old and worn out currency notes are included. To
neighborhood center is applied. test the verification capability of the system, counterfeit
Morphological area opening: A morphological area currency notes are collected. Among these counterfeit
opening is applied with 30 pixel area to remove objects notes, previous Ethiopian genuine notes are also included.
less than 30 pixel area. The value 30 is set through In addition to these currencies, other countries’ currencies
experiment. are also collected, namely US dollar, Rand, Dirham, and
Counting the number of objects: The final Shilling. Both scanner and camera images of these
verification step is to count the number of objects in the currencies are acquired and stored for evaluating and
localized image. If the number of objects is one testing the system. The scanner used for capturing images
(representing thin golden strip), it is a genuine currency. is HP Scan jet G2710. Each currency is scanned with a
Otherwise, it is a counterfeit. resolution of 200dpi, size 1122 x 570 pixel and saved in a
JPEG image. From the scanned image, a standard image is
b) Verification by the Wide Strip selected from each denomination and stored as a reference
for comparing the test currency image. The camera used
In addition to the thin golden strip, 50 Birr and 100 Birr to capture the images is Canon PowerShot SD1400 IS
banknotes have wide strips that would help for further Digital ELPH with 14.1 Mega pixels. The images of the
verification. The verification involves the following steps. currency are taken with a background white light, so that
the broken thin golden strip to be solid line.
ROI localization: The position of the wide strip is
similar and, as a result, the ROI is be localized by using B. Implementation
pixel coordinates.
For this work, MATLAB tool is used as a development
Binarization: For better processing, the RGB is
environment. The SURF extraction and matching are
converted to grayscale format and the grayscale image is
implemented and included in MATLAB ToolBox.
converted to binary image by using an optimal threshold.
Toshiba laptop of the following specification is used for

Copyright © 2016 MECS I.J. Image, Graphics and Signal Processing, 2016, 2, 28-36
 Automatic Recognition and Counterfeit Detection of Ethiopian Paper Currency 35

the research: Windows 7 32-bit operating system. The REFERENCES

hardware component comprises of Intel(R) Core(TM) 2 [1] A. Ahmadi, S. Omatu, T. Kosaka and T. Fujinaka (2004).
Duo CPU of 2.1 GHz, 3GB memory, and 500GB hard “A reliable method for classification of bank notes using
disk. artificial neural networks”. Artificial Life and Robotics,
C. Test Results 8(2): 133–139.
[2] M. Aoba, T. Kikuchi and Y. Takefuji (2003). “Euro
The accuracy of the denomination or denomination rate banknote recognition system using a three-layered
is measured by the ratio of number of accurately perceptron and RBF networks”. Transactions on
denominated currencies and the total number of currencies Mathematical Modeling and Its Applications, 44:99–109.
[3] H. Bay, A. Ess , T. Tuytelaars, L. Gool (2008). “Speeded-
tested. And the accuracy of the verification or the
Up Robust Features (SURF)”, Computer Vision and
verification rate is measured as the ratio of number of Image Understanding: Similarity Matching in Computer
accurately verified currencies and the total number of Vision and Multimedia, 110(3):346–359.
accurately denominated currencies. Both the [4] S. Chae, J. Kim, and S. Pan (2009). “A Study on the
denomination rate (DR) and verification rate (VR) are Korean Banknote Recognition Using RGB and UV
calculated for each denomination of Ethiopian currency as Information,” in Communication and Networking, Series
follows. of Communications in Computer and Information Science,
56:477–484.
Correctly Denominated Currency Images [5] T. H. Chia and M. J. Levene (2009). “Detection of
DR  *100
Total Test Currency Images counterfeit U.S. paper money using intrinsic fluorescence
(2) lifetime,” Optics Express, 17(24): 22054–22061.
[6] K. Debnath, S. Ahmed and Md. Shahjahan (2010). “A
Correctly Verified Currency Images
VR  *100 paper currency recognition system using Negatively
Correctly Denominated Currency Images Correlated Neural Network Ensemble”, Journal of
(3)
Multimedia, 5(6): 560-567.
Accordingly, test results show that the average [7] F. Garcia-Lamont, J. Cervantes, A. Lopez (2012).
denomination rate is 90.42% with an average processing “Recognition of Mexican Banknotes via their colour and
texture features”, Expert Systems with Applications, Vol.
time of 1.68 second. In the experiment, it was observed
39.
that the denomination rate for relatively new currencies of [8] L. Georgieva, T. Dimitrova, N. Angelov (2005). “RGB
all denominations is 100%. But for very old notes, the and HSV Colour Models in Colour Identification of
denomination rate is between 80% and 90%. On the other Digital Traumas Images”, International Conference on
hand, 83.3% of 50 Birr and 100 Birr counterfeit bank Computer Systems and Technologies, Vol. 12, No.1.
notes were classified according to their denomination. [9] M. Gogoi, S. Ali and S. Mukherjee (2014). “Automatic
However, no counterfeit banknote passed the verification Indian Currency Denomination Recognition Based on
component and we achieved 100% accuracy for rejecting Artificial Neural Network”. Proceedings of the 2nd
counterfeit banknotes. The system was also tested with International Conference on Signal Processing and
Integrated Networks, Noida, IEEE.
Dollar, Rand, Shillings, and Dirham. All of these foreign
[10] H. Gou, X. Li, X. Li, and J. Yi (2011). “A Reliable
currencies were rejected by the system, which is the other Classification Method for Paper Currency Based on LVQ
strength of our approach. Neural Network,” in Advances in Computer Science and
Education Applications, 202:243–247.
[11] F. Hasanuzzaman, X. Yang, Y. Tian (2011). “Robust and
V. CONCLUSION AND FUTURE WORKS Effective Component-based Banknote Recognition by
SURF Features”, Proceedings of the 20th Annual
We described a system that automatically classifies a Conference on Wireless and Optical Communications,
given input image into one of the five Ethiopian paper Newark, NJ, pp. 1-6.
currencies and also checks if the banknotes are genuine or [12] N. Jahangir and A. Chowdhury (2007). “Bangladeshi
not. Although, to our knowledge, this is the first attempt Banknote Recognition by Neural Network with Axis
on Ethiopian currency recognition, we have achieved very Symmetrical Masks”, In Proc. of 10th International
promising results that could encourage applying the Conference on Computer and Information Technology,
9:1-5.
system in real-world environment. The system can
[13] T. Kagehiro, H. Nagayoshi and H. Sako (2006). “A
stimulate the economy by facilitating business hierarchical classification method for US bank notes”.
transactions in a more convenient way. Test results Transactions on Information and Systems, E89D
showed that the system is effective in rejecting counterfeit (7):2061–2067.
and foreign banknotes. By tuning the parameters and [14] W. Kavinda, S. Dhammika (2013). “Bank notes
adjusting thresholds, we can even obtain an optimal result. recognition device for Sri Lankan vision impaired
Thus, future work is directed towards improving the community”, In Proc. of 8th International Conference
performance of the system and installing it in real Computer Science & Education, Colombo, pp. 609-612.
environment.

Copyright © 2016 MECS I.J. Image, Graphics and Signal Processing, 2016, 2, 28-36
 36 Automatic Recognition and Counterfeit Detection of Ethiopian Paper Currency

[15] Y. Liu, L. Zheng and Y. Huang (2014). “Haar-SVM for

Real-Time Banknotes Recognition”. Journal of
Information and Computational Sciences, 11(12): 4031-
4039.
[16] A. Sargano, M. Sarfraz, N. Haq (2013). “Robust features
and paper currency recognition system”, Proceedings of
the 6th International Conference on Information
Technology.
[17] F. Takeda and S. Omatu (1995). “High-speed paper
currency recognition by neural networks”, IEEE
Transaction on Neural Networks, 6(1):73-77.
[18] K. Verma, B. Singh and A. Agarwal (2011). “Indian
Currency Recognition Based on Texture Analysis”.
Proceedings of the Nirma University International
Conference on Engineering (NUiCONE), Gujarat, IEEE.
[19] W. Yan, J. Chambers and A. Garhwal (2015). “An
Empirical Approach for Currency Identification”,
Multimedia Tools and Applications, 74: 4723-4733.
[20] C. Yeh, W. Su and S. Lee (2011). “Employing multiple-
kernel support vector machines for counterfeit banknote
recognition”. Applied Soft Computing, 11: 1439-1447.
[21] E. Zhang, B. Jiang, J. Duan, and Z. Bian, “Research on
Paper Currency Recognition by Neural Networks”, In
Proceedings of the 2nd International Conference on
Machine Learning and Cybernetics, Vol. 4, 2003.
[22] X. Zhu, M. Ren, “A Recognition Method of RMB
Numbers Based on Character Features”, In Proc. of 2nd
International conference on Information, Electronics and
Computer, 2014.

Authors’ Profiles

Jegnaw Fentahun Zeggeye received Bachelor Degree in

Physics, Bachelor Degree in Computer Science and Master’s
Degree in Computer Science from Addis Ababa University,
Ethiopia. His research interest includes pattern recognition and
digital image processing.

Yaregal Assabie received his PhD in Electrical Engineering

from Chalmers University of Technology, Gothenburg, Sweden.
He received Master’s Degree in Information Science and
Bachelor Degree in Computer Science from Addis Ababa
University, Ethiopia. He is currently working as an Assistant
Professor at the Department of Computer Science, Addis Ababa
University. His research interests are natural language
processing, pattern recognition and digital image processing.

Copyright © 2016 MECS I.J. Image, Graphics and Signal Processing, 2016, 2, 28-36

View publication stats