Вы находитесь на странице: 1из 6

Currency Recognition System for Visually Impaired:

Egyptian Banknote as a Study Case


Noura A. Semary1, Sondos M. Fadl2, Magda S. Essa3, Ahmed F. Gad4
#
Faculty of Computers and Information, Menoufia University,
Shebeen El-Kom, Menoufia, Egypt
1
noura.samri@ci.menofia.edu.eg
2
sondos.magdy@ci.menofia.edu.e
3
eng_magda_mcse@yahoo.com
4
ahmed.f.gad@gmail.com

Abstract One of the most important problems facing visual In this work, camera-based Egyptian paper currency is
impaired people is money recognition especially for paper trained to be recognized using very simple image processing
currency. In this paper we present a simple system currency utilities what makes the processing time is very short with
recognition system applied on Egyptian banknote. Our proposed acceptable accuracy. The proposed systems has the ability to
system is based on simple image processing utilities that insure treat papers captured partially and under different lighting
performing the process as fast and robust as possible. The basic conditions.
techniques utilized in our proposed system include image
foreground segmentation, histogram enhancement, region of The rest of the paper is organized as follows; Section 2
interest (ROI) extraction and finally template matching based on presents the literature works related to paper money
the cross-correlation between the captured image and our data recognition research works and Mobile applications. Section 3
set. The experimental results demonstrate that the proposed explains the details of the proposed system. Section 4 presents
method can recognize Egyptian paper money with high quality the experimental results while Section 5 concludes the paper.
reaches 89% and short time.
II. RELATED WORK
KeywordsMoney Recognition, Visual impaired, currency
reader
In this section, related literature works will be investigated.
AS we have mentioned before, most works consider the whole
I. INTRODUCTION money paper (Scanned paper) what is not the case of our
interest. Accordingly, camera-based works are discussed in this
The World Health Organization estimated the number of section too.
visually impaired at the global level, based on the latest studies,
that there are about 285 million people who suffer from visual A. Scanner-based systems
impairment worldwide, of whom 246 million people have Trupti and Bawane [2] proposed a system that scans the
visual impairment and 39 million people are totally blind. It whole paper for recognition. They considered the size as one of
also estimated that the Eastern Mediterranean Region the three characteristics of paper currencies in addition to color
constitute 12.6% of the proportion of blindness in the world and texture. Image histogram and the plenitude of different
[1]. colors in a paper currency is calculated. The method discussed
One of the main problems faced by people with visual in this paper can be used for recognizing paper currencies from
disabilities is the inability to recognize the paper currencies due different countries. It also represents a currency recognition
to the similarity of paper texture and size between the different system using ensemble neural network (ENN). The individual
categories. Hence, the role of technology is to develop a neural networks in an ENN are skilled via negative correlation
solution to resolve this crisis to make blind people feel safety learning. The purpose of using negative correlation learning is
and confidence in the financial dealings. to skill the individuals in an ensemble on different parts or
portion of input patterns. They used ENN to identify new, old
There are two trends in Money recognition research field; and noisy currencies. The Ensemble network is used for the
Scanner-based and Camera-based. Scanner-based systems categorization of different types of currency. It minimizes the
assumes capturing the whole paper (like scanner). Such chances of misclassification than a single network and
systems are suitable for the Machinery of money counters. ensemble network with independent training. They didnt make
While camera-based systems assumes capturing the paper by a any experiments.
camera which may capture a part of the paper. Most related
works in literature deal with the scanner-based type [2-5]. For Yaseri et al.[3] used Fourier-Mellin transform for invariant
visual impairment usage, its supposed to enable users to rotation, translation and scale of the input image. The image is
capture any part of the paper by their smart phone and let the segmented and Markovian characteristics of each segment have
system recognize it and tell the currency value. been utilized to construct a feature vectors. These vectors are
then fed into SVM classifier for paper currency recognition.
This technique considered the whole paper currency was
applied to both sides of the banknotes, and it could result morphological shape filters. Two levels classification were
98.7% accuracy in recognizing paper currencies. Due to using used for different categories of patches (portrait, denomination,
the texture characteristic, the system recognizes the banknotes tactile marks), and the votes are combined through an ensemble
in any direction. decider.
Ahmed and Mirfa [4] developed an intelligent system for Moreover, many applications are found in google play for
Pakistani paper currency that could recognize the currency note currency recognition without any reference on their exact
accurately. In their paper, they have taken samples domain of scientific contributions. IDEAL Currency Identifier [9] is an
five different Pakistani paper currency notes (Rs. 10, 20, 50, Android mobile application that allows visually impaired
100, 1000). They scanned total 100 currency notes (whole persons to identify U.S. dollars. It provides several facilities to
currency paper), 20 from each sample of selected domain for make recognizing the notes easier. It periodically captures
feature extraction of these images using a software. The frames using the device camera and process them in short time
proposed system uses the different features of the currency for to notify the user by the recognized note. Using vibration
recognition. Experimental results are presented which show patterns the user can know when an image frame was captured.
that this scheme can recognize currently available 8 notes of Text-to-speech (TTS) is used to tell the user by the value of the
Pakistans Currency (Rs. 10, 20, 50, 100, 500, 1000 etc.) paper note which provides easy understanding and integration
successfully with an average accuracy of 98.57%. between the user and the application. Automatic camera focus
is one of the advantages of this application as it can adjust the
Fatemeh Daraee and Saeed Mozaffari [5] proposed a brightness of the camera to match the existing environment to
method for worn out Farsi money notes recognition using their efficiently recognize the currency notes under different
texture content and wavelets transform. First, with the help of illumination conditions.
face detection algorithm, the obverse of the money note is
separated from its reverse side. Then, central part of the money This application uses template matching image processing
note, containing texture is extracted. Finally, wavelets technique to detect and recognizes the note. Using the
transform is applied to this region of interest (ROI) to extract templates, it can compare them by specific regions in the
some features. Some distance measures are utilized to classify captured frame. It normally uses two templates of width and
the input money note into predefined groups according to height of 100 pixels for front and back of each note. Thus it can
minimum distance. To increase accuracy of the system, in the provide information whether the back or front of the note is
post processing step, they used their test amount of the money shown. It also has the ability to resize these templates in
note and template matching technique. The experimental different sizes to match the captured frame size. But
results have shown that system performance is 80% for eroded unfortunately there are some limitations to work with notes
money notes recognition. with poor quality and low lightning. It has false results with
worn, wrinkled notes.
Most discussed works dedicate their systems for Money
counting applications rather than visual impairment issue. LookTel Money Reader [10] is a mobile application that
operates on iOS to help people identify and count their notes. It
B. Camera-based systems can help fully blind by notifying them by the amount of
Very limited applications serves visual impaired peoples in identified note using TTS. Also it can help visually impaired
this issue. In this subsection we will investigate a number of people by presenting the amount of recognized note using a big
mobile applications and devices that are dedicated for helping font appearing at the middle of the screen.
them to stand alone in different tasks such as the ability to
recognize money in different currencies. It has a large database that supports twenty one currencies
such as US Dollar, Euro, United Arab Emirates Dirham,
Suriya et. al. [6] have proposed a mobile application starts Japanese Yen, Kuwaiti Dinar. Also it can work in different
with segmentation stage for removal of clutter. They have used languages such as English, Spanish, French, Japanese. It has
a visual Bag of Words (BoW) method. The methods used work little limitations on the location of the device when capturing
well on noisy images captured from a mobile phone. They have an image and works instantly in real-time.
applied their system on paper bills of Indian National Rupee
and reported 96.7% with SIFT descriptor. One of its features is that users can open a session and
during it the application can count the amount of recognized
Anas [7] has proposed a non-parametric approach where a currencies to give a total at the end of the session. It was also
nonparametric model is obtained by averaging aligned samples developed to work with Apple laptop and desktop computers.
of a banknote. Then, the coefficients of determination are
calculated between an unknown banknote and all the models. ViaOptaDaily [11] is an Android application that provides
Finally, a discriminant analysis is used to assign a model to the an integrated environment that includes six main services for
unknown banknote. The proposed method was applied blind people. First one is to help blinds to know the weather. It
efficiently to recognize three kinds of Saudi banknotes with can provide weather information about current and next three
error rate 10%. days and notifying the blind by them. It has a magnifying glass
that allow capturing an image using the device camera that
In [8] the authors have proposed a currency recognition contains text. After capturing the image, it allow to zoom in.
system built on the bionic eyeglass. Relevant shapes are Also it can read the text in the image and present it in an
extracted from the image flow of the Hungarian banknote audible form to the user. Third feature is a simple timer that
shown to the mobile camera using adaptive thresholding and count down and notifying the person when it fires. Fourth
feature is to allow the person to assign a specific sound to a Image
specific contact and thus providing higher flexibility to make a acquisition
call. The fifth service is the color recognizer that periodically
captures image using the camera, detect the main color in the
image and notify the user about the color using sound. The
final feature is the money recognizer. Using this service it can Pre-processing
work with three currencies. Euro, U.S. Dollar and Saudi
Arabian Riyal.
Generally, this application has a number of useful services.
Although it is compatible with different languages such as Segmentation
English, French, Arabic and Spanish, it is complicated in the
interaction with one that cannot see because its interface
contains many icons and there is no voice commands make the
process easier.
Histogram
Regarding the field of this paper to recognize Egyptian equalization
money, there are a number of applications that recognize the
Egyptian money.
Floos Mubser is the first Egyptian Android application that ROI extraction
recognizes the Egyptian currency. It can efficiently recognize
many of the Egyptian notes under high lighting. It has a
number of features such as using camera flash lighting to
provide recognition of the currency under poor lighting. It can
work with the devices with little processing power by capturing Template
frames with different qualities. Also it can adjust camera Matching
contrast.
(a) (b)
But unfortunately it lakes flexibility as all of these features
are manual that adds a heavy burden of the blind or the visually Fig. 1. (a) System Block Diagram (b) visual results for 2 samples
impaired person to adjust them and thus adding complexity in
the process. A. Image acquisition.
The image is captured by digital camera as RGB image
Another Egyptian application is Floosy [12]. It is Android and is converted to gray scale version by intensity equation 1.
mobile application that has similar algorithms with Floos
Mubser application. it uses image color features rather than
   
template matching to identify the currency. For each Egyptian
note, it extracts one color features stored in its database.
Automatically and periodically it can capture frames using the Where I is intensity value, R is red channel value, G is
device camera. By extracting the feature from each note, it Green channel value and B is Blue channel value.
compares it with its pre-calculated features and based on B. Pre-processing.
similarity measures it can find the nearest note it can be. But In order to remove noises on the image and improve the
because it uses color features, the lighting is a key problem. It quality, Gaussian blurring is used. The equation of a Gaussian
behaves differently under different illumination conditions and function in two dimensions is presented in equation 2.
fails or mistakenly identify notes. This application has a
number of features such as using the flash light and adjusting

camera focus. These features can be manually configured and 

  

thus lacing flexibility.
In this work, simple system is proposed and tested in Where x is the distance from the origin in the horizontal
Windows and Android platforms. The proposed system axis, y is the distance from the origin in the vertical axis, and
considers capturing any part of the money paper in a specific is the standard deviation of the Gaussian distribution [13].
direction. Simple image processing utilities are used to
recognize the currency value using template matching. Next C. Segmentation
section discuss the proposed system in details. In order to extract the currency paper from the background,
the image is converting to binary version. Binary image is the
III. PROPOSED APPROACH image that only contains two possible intensity values; usually
In this work, we propose a general framework for black (0) and white (1) as in the equation 3:
identifying paper banknotes. There are 5 main stages in the
system, as presented in Figure 1. 
   

Otsu thresholding method is using for adapting the
threshold value th After that, label connected components is
determined and morphology holes filling is performed as
following [14] :
1. Run-length encode the input image.
2. Scan the runs, assigning preliminary labels and
recording label equivalences in a local equivalence
table.
3. Resolve the equivalence classes.
4. Re-label the runs based on the resolved
equivalence classes.
D. Histogram equalization.
Histogram equalization is used to adjust the contrast based
on the image histogram, modifying the brightness of the image
and make the image look more clear.
E. ROI Extraction
First, the extracted mony paper is resized to the dataset
height. So, if the currency paper size is M N ,then the
extracted paper region will be resized by M N/M. Then a
block of size r c is selected for time acceleration (In this
paper, the top right corner is selected).
F. Template Matching
After the region of interest ROI (block) , correlation is
performed to measure the similarity between images in the
database and the ROI part. Cross-correlation function (4) has Fig. 2. The Used Dataset
its maximum value at the position of the found ROI in the
Money saved image. The original image and the template
image do not need to be of the same size.


    

Where is the image, is the mean of the template and


is the mean of in the region under the template. Fig. 3. Some testset samples

IV. EXPERIMENTAL RESULTS


In this section, we first introduce the experiment procedure
and then, discuss the results to prove the robustness and
sensitivity of our algorithm.

A. Experiment Procedure
First, the experiments were carried out on the Matlab Fig. 4. Currency recognition application at work under Samsung tablet GT-
R2012a running on 4 GB RAM and 2.30 GHZ processor N800 testing one piece of 5 EGP from different views
machine. An then transferred to Android platform. The aim of
using Matlab is to validate the proposed image processing The test images have been captured using popular mobile
framework and compare the accuracy and the time of running phone cameras, with different resolutions, viz. VGA, 1.3
such systems on both platforms. For the Android application, megapixels (MP), 2 MP and 5 MP, all on default settings. For
OpenCV library is utilized for all image processing functions Matlab version, all the images has been captured offline with a
and all the codes were written in Java. limitation of having both the horizontal edges of the paper in
the image. While for the Android system an automatic snapshot
Egyptian paper money of values 1, 5, 10, 20, 50, 100 and
function is used to detect the upper and lower horizontal edges
200 have been scanned for both sides and saved in a dataset.
before automatic capturing. Hough line detection transform is
Figure 2 presents the images used in our dataset while Figure 3
utilized to obtain the horizontal edges.
presents some test samples. In this paper, 120 test images have
been examined, 20 for each class.
Since the android application is designed for visual The accuracy of the system reached 89% under the Matlab
impaired users, there is no icons or buttons and no manual system.
configurations wanted. The application starts reading the
database of the saved currency banknotes then start capturing
the fronted scene. Once both horizontal edges are detected a
vibration effect is occurred to notify the user of correct money
detection. After a while, a sound pronounces the correct value
in Arabic is hear. Figure 5 shows the application running on
Samsun GT-N800 tablet testing 5-pounds mony paper.
B. Visual Result
Figure 5 shows samples of the system steps illustrated in
Matlab. row shows input images, row shows gray
images, row shows Gaussian blur images, row shows
binary images, row shows object extraction images, row
shows the extracted ROI block and row shows the
recognition results.

C. System Evaluation
The system has been evaluated under the Matlab system
using accuracy measure.

    

Where True indicates both true positive and true negative


results while false indicates the opposite. Table 1 shows the
accuracy of our system calculated for our 120 test images (20
sample for each class). The average accuracy was 89%. The
least accuracy results for the 50 pounds class. Its supposed that
this values resulted due to the fact that in the 50 pounds paper,
the top right corner has very few details what may cause the
fault ratio. The average running time for Matlab system is 10
seconds while in android system is 12 seconds.

TABLE I. EGYPTIAN CURRENCY CLASSIFICATION RESULTS


5 10 20 50 100 200
pound pound pound pound pound pound
Quantity 20 20 20 20 20 20

True 18 17 20 16 19 17
False 2 3 0 4 1 3 Fig. 5. Samples of Matlab System Results
Accuracy 90 85 100 80 95 85
(%) REFERENCES
V. CONCLUSION [1] D. Pascolini, , and S. P. Mariotti. "Global estimates of visual
impairment: 2010." British Journal of Ophthalmology (2011):
In this paper, simple Egyptian currency recognition system bjophthalmol-2011.
has been proposed. The proposed system starts with capturing
still image. Simple image processing techniques like [2] A. Ms Trupti Pathrabe, and B. Dr NG Bawane. "Paper currency
thresholding, noise removal, histogram equalization and recognition system using characteristics extraction and Negatively
segmentation are used to extract the ROI and facilitate the Correlated NN Ensemble." International Journal of Latest Trends
template matching procedure. Correlation based template in Computing (E-ISSN: 2045-5364) vol. 1, no. 2 (2010), pp: 121-
matching is used after that to find the ROI in the dataset 124.
images. The system has been written in Matlab and OpenCV
library under Android platform as well. The Matlab system [3] A. Yaseri, and S. M. Anisheh,. "A Novel Paper Currency
considers offline captured images while the Android one was Recognition using Fourier Mellin Transform, Hidden Markov
designed to match visual impaired users. Live video capturing Model and Support Vector Machine". International Journal of
is considered for the Android application. Automatic horizontal Computer Applications, vol. 61, no. 7 (2013), pp: 17-22.
edges has proposed to match our template matching system.
[4] A. Ali, , and M. Manzoor. "Recognition System for Pakistani and Design (ECCTD), 2011 20th European Conference on, 2011,
Paper Currency." World Applied Sciences Journal vol. 28, no. 12 pp. 841-844
(2013), pp: 2069-2075.
[9] http://www.wirelessrerc.org/node/160
[5] F. Daraee,, and S. Mozaffari. "Eroded money notes recognition
using wavelet transform." In 6th Iranian Machine Vision and [10] http://www.looktel.com/moneyreader
Image Processing (MVIP), 2010, pp. 1-5. IEEE, 2010.
[11] http://viaopta-apps.com/ViaOptaDaily.html
[6] S. Singh, S. Choudhury, K. Vishal and C. V. Jawahar, Currency
Recognition on Mobile Phones, 22nd International Conference on [12]https://play.google.com/store/apps/details?id=icansee.money.eg
Pattern Recognition (ICPR), Sweden, (24 August 2014), pp: 2661- &hl=en
2666, IEEE
[13] M. S. Nixon and A. S. Aguado. Feature Extraction and Image
[7] A. M. Ali Fattouh, A Non-Parametric Approach for Paper Processing. Academic Press, 2008.
Currency Recognition, International Journal of Computer Science
and Software Engineering (IJCSSE), vol. 4, no. 5, (May 2015), [14] R. M. Haralick, , and L. G. Shapiro, Computer and Robot
pp: 121-125. Vision, Volume I, Addison-Wesley, 1992, pp. 28-48.

[8] Z. Solymr, A. Stubendek, M. Radvnyi, and K. Karacs,


"Banknote recognition for visually impaired," in Circuit Theory

Вам также может понравиться