An Iranian Cash Recognition Assistance System For Visually Impaireds

International Journal on Soft Computing, Artificial Intelligence and Applications (IJSCAI), Vol.4, No.1, February 2015
DOI :10.5121/ijscai.2015.4103 23
AN IRANIAN CASH RECOGNITION ASSISTANCE
SYSTEM FOR VISUALLY IMPAIREDS
Esmaiel Nojavani, Azade Rezaeezade and Amirhassan Monadjemi
Department of Computer Engineering, University of Isfahan, Isfahan, Iran
ABSTRACT
In economical societies of today, using cash is an inseparable aspect of human’s life. People use cash for
marketing, services, entertainments, bank operations and so on. This huge amount of contact with cash and
the necessity of knowing the monetary value of it caused one of the most challenging problems for visually
impaired people. In this paper we propose a mobile phone based approach to identify monetary value of a
picture taken from a banknote using some image processing and machine vision techniques. While the
developed approach is very fast, it can recognize the value of the banknote by an average accuracy rate of
about 97% and can overcome different challenges like rotation, scaling, collision, illumination changes,
perspective, and some others.
KEYWORDS
Cash Identification Using Mobile, Visually Impaired Assistant, Iranian Cash Identification, Mobile Phone
Extra Usage
1. INTRODUCTION
From the beginning of appearance of humans on the earth, we always gain lots of information
about our surrounding environment visually. As a result, our developed technologies mostly are
based on vision. Hence, human beings who have some kind of visually impairment always have
suffered from that. One of the most common innovations of humans that proof this assertion is
money. Visually impaired people have lots of difficulties to use money in daily transactions,
unless they use a person or device as an assistant to help them. In 2013, World Health
Organization approximates 285 million people are visually impaired worldwide, that 39 million
of them are blind. According to the Health Organization of Iran, about 600 thousands Iranian
have moderate or severe visual impairments that about a third of them are totally blind. This
statistics show the importance of developing some efficient methods to assist this people in cash
recognition. In this paper we develop a mobile phone based cash recognition system that helps
visually impaired people to identify the current Iranian cash easily and accurately.
Banknotes in different countries have texture, colour, size, and appearance differences [1], so
feature extraction and identification approaches that is used in one country usually is not usable in
other countries or does not work properly over there. Moreover, lots of previous work in concept
of cash recognition is restricted to specific standard conditions. For example many of the
developed approaches do not support rotation or noisy backgrounds [2-5], or in some other
approaches the whole banknote must be visible in taken picture [1, 6]. By considering visually
impaired limitations to take pictures with special physical characteristics, these approaches are
not user friendly enough.

24
In this study we consider many complexity and variety for taken pictures and as a result the
developed system can support rotation, scaling, complex or noisy background, camera angel
changes, collision and even the variation of illumination. Since the photo takers in proposed
methods are visually impaired, these complexities should be considered to result in an efficient
system. For explication we define this concept below and show some of them in
Figure 1.
• Rotation: the cash can have different rotation angle to the camera.
• Scaling: the cash can poses in different distances from the mobile camera.
• Complex background: the cash can be taken with any background except other cash.
• Perspective: the cash can poses in any camera angle.
• Collision: the taken picture can be from a part of cash. It should only contain the
digits of monetary value.
• Variation of illumination: cash picture can be taken in different illumination.
Figure 1. Some possible varieties in taking picture (perspective, rotation, scaling and different illumination)
In the developed approach, we used monetary value digits exist on cash to create a robust and
flexible cash recognition system. However, regarding to the different conditions of taking pictures
that was shown in
Figure 1, these digits are not in a constant place for all the pictures. So, first of all we should find
the place of these digits on the taken image. In this study, to localize the monetary value digits on
the image, we invent an algorithm that finds zeroes in the taken pictures. After finding the place
of zeroes in the image, we proceed to find the remained digit of monetary value that can be 1, 2 or
5, and try to identify this digit using a neural network. Besides that; we should count the number
of zeroes too. After these two steps we can identify the monetary value of that cash.
The contributions of this paper include: section 2 reviews the state of the art on banknote
recognition, section 3 explains the proposed approach for Iranian cash recognition, section 4
evaluates the recognition results and section 5 conclude the whole paper.
2. RELATED WORK
The existing cash recognition approaches in literature, mainly use image processing and neural
network techniques. Moreover, there are some developed devices which use physical
characteristics of banknotes like size or colour of them. The most considerable point about cash

25
recognition is that many developed devices or methods for cash recognition in one country are not
usable in other countries. In this section we will introduce some of these devices and methods.
Money Talker [7] is a device that recognizes Australian bank notes electronically using the
reflection and transmission properties of light. This device utilizes the largely different colours
and patterns on Australian banknotes. Different colour lights are transmitted through the inserted
note and the corresponding sensors detect distinct ranges of values depending on the colour of the
note. Cash Test [8] is another device determines the value of banknote by a mechanical means
relying on the different lengths of each notes. The downfall of the Cash Test is that it does not
allow for the shrinking of notes with time nor the creases or rips that are common in our notes.
Moreover, while Cash Test is cheap and extremely portable by the reports of users, it is
inaccurate and difficult to use.
Kurzweil reader [9], iCare [10] (uses a wearable camera for imaging and a CPU chip for
computation) and some other devices have been developed too, each has their advantages and
disadvantages, but the common property of them is that the user most carry a device everywhere.
Some of them are bulky and expensive too. In [11], a method has been proposed that is a part of a
wearable aiding system for the visually impaired. A standard video camera has been used in this
system as the image collecting device. This method is based on identification of the geometrical
patterns that characterize the different denominations of banknotes and does not require training
or building the data base. This method is bound by the requirements of a wearable system. This
system require the flowing components: a wearable frame similar to conventional eyeglasses that
a camera are mounted on it, a portable computer which provide a user interface and can process
the image acquired by the camera, a set of headphone to announce the monetary value to the user.
In [1], Hassanpour and Farahabadi proposed a paper note recognition method which uses Hidden
Markov Model (HMM). This method models banknotes using texture characteristics including
size, colour and a texture-based feature from banknote and compares extracted vector with the
instants in a data base of paper cash. But the main purpose of this method is to distinguish
national banknotes from different countries and does not support visually impaired limitations. In
[12], Garcia-Lamont et.al introduced a recognition method for Mexican banknotes. This method
uses colour and texture features. For classification, the LVQ network and G statistic are
employed. The colour feature is modelled under the RGB colour space and the texture feature in
grey scale, is modelled with Local Binary Patterns method. This method originally developed for
automatic banknote detection is stable conditions. So do not support illumination variation. Also
this method cannot overcome the existence of background in the image and is not usable as
visually impaired assistance.
In [13], Lui proposed a background subtraction and perspective correction algorithm using Ada-
boost framework and trained it with 32 pairs of pixels to identify the bill values. This system uses
video recording and works on snapshots. This method does not support rotation and noisy
background and strongly relays on the white and straight edges of bill, so it is not usable for
Iranian paper cash recognition because they do not have this required margins usually. In [],
Stavros et.al proposed a simple approach to identify Euro, Zloty and US Dollar. The input of this
method is the RGB image of the corner of cash. The input image converts to a grey scale image.
Then an adaptive segmentation algorithm that traverse every line in the image and compute the
threshold by averaging the maxima and minima of every lines intensity, applies to convert the
grey scale image into a black and white image. After extracting the digits by vertical and scan-
lines, a unique pattern will be generate that can be used to identify the monetary value of cash.
The main issue of proposed approach in [13] is that the photo taker should take the picture in a
way that horizontal and vertical boundaries of the cash place in the correct direction.

26
In [14] Hassanuzzaman and Yang proposed a component based framework using SURF features.
The proposed method achieved 100% accuracy for US dollars through various conditions. While
the proposed framework is a robust and effective approach and is available to overcome
challenges like image rotation, noisy background, collision, scaling and illumination changes, it is
not a suitable method for Iranian banknote recognition. This disproportion refers to the texture
difference of Iranian currency tissues and US dollars. For proving this unfitness we have applied
this approach to Iranian banknotes and checked it by extracting different components. But the
output accuracy was extremely low which has convinced us that SURF is not a suitable approach
for Iranian cash recognition.
In [15] Grijalva et.al proposed a method that has been tested on different Nokia smartphone. The
reported accuracy rate is about 99.8% and the speed of processing reported as 7 frames per
second. The proposed approach does not support rotation, perspective, existence of background
and partial detection. In the proposed approach Grijalva et.al used digital image processing
concepts like Eigen-face and PCA.
As we can see, while all these methods have their advantages, we cannot use them in a user
friendly and accurate manner for Iranian cash recognition. So we are going to develop a user
friendly method which is portable with no difficulties, have enough accuracy, and considering
impaired people limitation for taking pictures in the next sections.
3. THE PROPOSED APPROACH
In this section we will explain the developed system completely. As mentioned earlier, in this
method we should provide a picture of the cash that contain the monetary digits. The taken
picture must be from the side that contains Iranian digits. For satisfying this condition, photo
taker should use face detection technology of his mobile phone. In front side of all Iranian cash,
there is a face of Imam Khomeiny. So if the user is taking picture from the front side his mobile
will detect Imam’s face. But if he is taking picture from the back side his mobile would not detect
any face, so he will understand that the cash must be turned back. We assume this picture is taken
by a mobile phone camera, so it does not have high resolution and is taken by a visually impaired
person, so can have any position. Then this picture will be analysed by machine vision and image
processing tools to extract the individual features of the cash and to recognize the value of that
cash. After that, the identified monetary value will be revealed to the user.
Feature extraction should be done by approaches that are robust to different environmental
conditions that reviewed earlier including rotation, scaling, noisy background, collision,
perspective, camera different angle, variance in illumination and any other possible situation. For
this process we will have two main steps. The firs one is localization of monetary value digits.
This step is carried out by means of some image processing methods and some innovation in
interpreting appeared concepts. Second step is recognition of the monetary value of cash. In this
step we will use neural network to identify the nonzero digit and identify the value of that cash. In
the following we will explain these two steps.
The mentioned face detection process, localization step for finding zeroes and counting them and
identifying nonzero digits are shown in Figure 2.

27
Figure 2. Summary of main steps in proposed approach
3.1. Localization of the Monetary Value
In the first step, to find the place of monetary value digits, the RGB image of the cash should be
changed to a grayscale one; it is because all the next processes for feature extraction and cash
identification are supposed to be done on a mobile device, so it should not be time and CPU
consuming. We apply a Wiener filter on the grayscale image to reduce the effect of oldness of
cash or some noises on the image. This grayscale image then will be converted to a binary image.
Since photo taker can use any mobile phone camera with any quality, and regarding to that the
user can take picture in variant illumination and with different background, for conversion of
grayscale image to a valid binary one we use a local adaptive approach for determining the
optimum threshold. The used approach is explained in [16]. In Figure 3(a) a sample of original
picture taken by a mobile phone camera is shown. The binary image after grayscale conversion
and using local adaptive threshold is shown in Figure 3(b). As you see, in this picture we have a
lot of black and white regions. We refer to these white regions as components.
In some Iranian cash, like 20000, 50000 and 100000 Riyals, because of existence of some huge
white components near the zeroes, there is a possibility that after binarization, these zeroes and
big components stick to each other. Moreover, we have this problem for zeroes in 100000 Riyals.
For example in Figure 3(c), a bigger region belongs to Imam’s face is stuck to zeroes. So, for
separation of these two regions, we use a 3×3 median filter to delete some small components on
image and to fill some small holes created on zeroes.

28
Figure 3. (a) A sample of original taken picture, (b) Binary image after applying local adaptive threshold (c)
Corresponding image after applying median filter, erosion and dilation operators. (d) Corresponding image
after elimination of regions that have a length to width ratio greater than 2 or number of white pixel to
boundary rectangle area pixel ratio greater than two.
In continuous, we are going to eliminate regions which cannot be a batch of zeroes because of
their shapes. In Persian orthography the width and length of zeroes must be approximately the
same. It means that the width and length of boundary rectangle of zeroes on cash or on the taken
picture must be approximately equal. So we can omit regions which the ratio of their boundary
rectangle length to its width is more than 2. So, components in a region with too different width
and length can be omitted with a high reliability. Moreover, while the shape of component are
irregular, there are some component that the ratio of length and width of their boundary rectangle
is less than 2, but if we calculate the ratio of their white pixel to the boundary rectangle area it is
greater than 2. It means that these components cannot be zero too. So we can omit them again.
The resulted image is shown in Figure 3(d).
While the existing zeroes in the image of cash are near to each other, if we apply closing
morphology operator [17] to the image, these zeroes and some other components will stick to
each other again. The result image is shown in Figure 4(a).
The lowest number of zeroes on current Iranian cash in is 3. So after applying closing operator,
we can delete connected regions which are made up of less than 3 components. This will limit the
number of questioned regions for being a batch of zeroes seriously. Resulted image is shown in
Figure 4(b).We should notice that it is not correct to omit regions with more than 5 components
(5 is the number of zeroes in 100000 Riyals) because some nonzero components might be
connected to the zeroes too.

29
Figure 4. (a) Corresponding image after using closing operator. (b) Elimination of regions with less than 3
components. (c) Corresponding image after elimination of regions and components which are not belong to
an acceptable line. (d) Keeping the region with the least difference between its components.
According to the existence of zeroes on a unique line, we can find regions with more than 2
components on a communal line and omit other regions and then in remained regions omit
components which are not on the corresponding line. For this purpose, we extract a line in each
connected region crossing gravity centres of any two components of that region. For example, if
we have a region with 3 connected components, we should establish equation for 2 lines. The first
one must cross gravity centre of the first and second components. The second one must cross
gravity centre of the second and third components. After that, for each line we should test if it
crosses other components of the connected regions or not. In the previous example, firstly we
should test that the first line crosses the third component or not and then we should check if the
second line crosses the first component or not. Notice that in this step existence of only one pixel
of the component on that line is sufficient. For each connected regions we keep the line that have
the most components on it. In addition, for each connected regions any other components which
are not on the established line should be omitted because they cannot be zero. The result is shown
in Figure 4(c).
As it is shown in Figure 4(c), it is possible that the algorithm keeps more than one line. Now we
should expand a solution to find the right region that contains zeroes. For this purpose we notice
the size of zeroes that should approximately be the same. So, for each region we get the size of all
components. After comparing the size of all components in each region, we keep only the region
that its components have the least difference and omit the other lines and regions. The result is
shown in Figure 4(d).

30
Figure 5. Obtaining the rotation angle
Surprisingly we can compute the rotation angle of the cash by means of corresponding line of the
last remained region which should be the batch of zeroes. It is just enough to obtain the angle of
the line that crosses the centre of first and last zeroes as it is shown in
Figure 5. We can compute the angle from (1).
In this equation, y0 is central width of first zero, y1 is central width of last zero, x0 is central length
of first zero and x1 is central length of last zero. Now if we rotate the image of Figure 3(b) by
angle, the rotated image in Figure 6 will be generated. Then, we can easily count the number of
zeroes. This count will be used for identifying the monetary value of image.
Figure 6. Corresponding image after rotation by angle
3.2. Recognition of Monetary Value
Till now, we could find place of zeroes and the rotation angle. Now we can fine the nonzero digit
using place of zeroes. In this section we will discuss the features extracted for remained digit. To

31
find the nonzero digit, we start moving to right and left from the most right zero and the most left
one equally. We will stop when we find the first component. This component certainly is our
nonzero digit. For extracting this digit from the image, we use the length of crossed line from first
zero to the last. In all cash the nonzero digit will be fixed in a square by size of L that L is
obtained from (2):
One of the vertical sides of square is very close to last zero and the centre of this side is
intersecting by the corresponding line crosses zeroes. The other vertical side of square is parallel
with the first one and has distance of size L from it. According to this we can extract the nonzero
digit of monetary value. Depending on whether we find the component in left side of the zeroes
or in right side of them, the nonzero digit will have 0 or 180 degree rotation. If the rotation angle
for nonzero digit is 0 degree, extracting the nonzero digit is finished, while if the rotation angle of
nonzero digit is 180 degree, we should rotate the extracted square 180 degree. Figure 7 shows
this operation.
Figure 7. A photo after rotation with angle has straight angle so we should rotate it 180 degree.
Figure 8 shows an extracted digit using this approach. Probably we will have some small
components in the extracted square but we remove them by keeping the largest connected
component and delete remainders. So we will maintain only the nonzero digit.
Figure 8. Extracted nonzero digit
Now it is time to identify the number exist in extracted square. As we mentioned earlier,
remained digit always is 1, 2 or 5. So for identification of any digit remained after finding it, we
must extract features that can classify these digits accurately, so we choose features as below:

32
1. Ratio of digit pixels to boundary rectangle pixels.
2. Ratio of height of digit to its width (it is computed as ratio of length to height of
boundary rectangle for each digit)
3. Ratio of left length to right length and ratio of top width to bottom width.
4. Horizontal and Vertical Symmetry that compare corresponding features in top and down
half and right and left half.
For classification we used a Multi-Layer Perceptron (MLP) neural network with one hidden layer.
For obtaining the best number of neurons in hidden layer, the network has been trained and tested
with different number of hidden nodes and the best result has been achieved by 20 neurons in one
hidden layer.
A summary of what we should do to identify the monetary value of a cash exist in taken photo is
available in Figure 9. In the last stage we have the count of zeroes and identified the nonzero
digit, so we can combine them and announce the monetary value to the user.
Taken
photo from
front side
Binarization Preprocessing
Localizing &
Counting
Zeroes
Finding the
Rotation angle
Fixing the
Rotation
Extracting the
Nonzero Digit
Fixing180
degree angle if
needed
Using MLP
for Nonzero
Digit
Identified
Monetary
Value
Figure 9. Total stage of recognition in phase 2 and 3
4. EVALUATION OF THE PROPOSED METHOD
In this section we will demonstrate the accuracy and performance of our method. Moreover we
will recommend a protocol to improve the accuracy of proposed method. As we mentioned
earlier, proposed approach is a two-phase method:
1. Localization of monetary digits and counting the zeroes
2. Identifying nonzero digit of the monetary value
So for computing the accuracy of this method we must use conditional probability. It is because
of dependence of second phase accuracy to the first phase. We suppose that the accuracy of
correct localization and counting number of zeroes is a% and accuracy of identifying remained
digit correctly is b%, so total accuracy of proposed method can be compute via the conditional
probability of (3):

33
For testing proposed approach we collect a data set containing 3500 image of current Iranian
cashes in different situations and by a vast amount of variety (different environmental conditions
we mentioned in the introduction). We took these pictures by different mobile phone cameras
which have different resolution from 3 to 8 mega pixels. The rate of accurate localization of
zeroes and accurate count of them (first phase) for different cashes are shown in Table 1.
Table 1. Accuracy of localizing monetary value and counting number of zeroes
Cash Accuracy
1000 Riyals 96.8%
2000 Riyals 91.6%
5000 Riyals 96%
10000 Riyals 98.4%
20000 Riyals 93.6%
50000 Riyals 95.2%
100000 Riyals 89.4%
Average 94.43%
As we mentioned earlier in 100000 Riyals cashes, the zeroes are near each other. Even in some
cases they are stick to each other. Moreover, as you can see in
Figure 10 in the case of 100000 Riyals cashes, there is an edge around zeroes. So, taking picture
from 100000 Riyals cannot support very tough situations. For example, when the taken picture of
100000 Riyals is from a big distance of the cash or when the perspective is very large, the chance
of correct recognition will decrease. After binarization, this edge creates some noises around the
zeroes that cause to incorrect localization of zeroes. The decrease in detection of zeroes place in
the case of 100000 Riyals in Table 1 is because of this issues. In continues we will recommend a
protocol which can reduce this challenge and increase the chance of correct detection of 100000
Riyals.
Figure 10: 100000 Riyal’s zeroes has stuck to each other
As we mentioned earlier, we used MLP neural network with one hidden layer for identifying
nonzero digit of monetary value. Number of neurons in the hidden layer is 20. Number of neurons

34
in the input layer is 4 in accordance with the number of extracted features and the number of
neurons in the output layer is 3 in accordance with the number of outputs including 1(1), 2(2) and
5(5), so we consider one neuron for each outputs. Transfer function for hidden layer is ‘logistic’
too.
Table 2. Accuracy of nonzero digits identification
Nonzero
Digit
Accuracy in
Training Set
Accuracy in
Test Set
1(1) 99.40% 98.83%
2(2) 99.69% 99.28%
5(5) 100% 100%
Average 99.7% 99.37%
For training and test of neural network, we divide created data set, which include 3500 instances,
into 2 distinct data sets randomly. 70% of samples should be in training data set and reminded
30% should be in test data set. Rate of correct identification or accuracy in this phase is shown in
Table 2.
Now by means of Table 1 and
Table 2 and by using multiplying rule in equation (3), total accuracy will be obtained as what is
shown in Table 3.
Table 3. Total accuracy for Iranian cash recognition
Cash Total Accuracy
1000 Riyals 95.67%
2000 Riyals 90.94%
5000 Riyals 96%
10000 Riyals 97.25%
20000 Riyals 92.93%
50000 Riyals 95.2%
100000 Riyals 88.35%
Average 93.76%
The obtained average accuracy is one of the best accuracies in field of Iranian cash recognition,
but it is not sufficient for a reliable detection, mostly about 100000 Riyals. So, in order to
increase the accuracy of proposed approach, we can do the photography operation with simple
limitations. As result we recommend a very simple protocol for taking picture. This protocol is
very easy to understand and utilize by visually impaired people but it can improve the results of
detection noticeably. The recommended protocol is as follow:
• As far as possible, refuse acceptance of worn out cashes. These cashes can be identified
simply by touching the surface and a visually impaired person can do it easily.
• After identifying the face of Imam Khomeiny on the cash in first phase, tie the cash in
front side and put your finger on the tied edge.
• Put the mobile’s camera on the middle of cash.
• Distance the camera slowly from the cash. You can count to three; so you will not
distance the camera from the cash too much.
• Take a photo.

35
By using the above protocol, the taken photos have little variety. Moreover, taken photos include
instances that their appearances do not cause to false detection. Besides, the monetary value
would not be covered by user’s hand. By removing the photos which have one of these problems
and generating a dataset with 3300 samples, the accuracy of proposed method in different phases
and totally is like what is shown in Table 4, Table 5 and Table 6.
Table 4. Accuracy of localizing monetary value and counting number of zeroes after using the
protocol
Cash Accuracy
1000 Riyals 93.4%
2000 Riyals 96.3%
5000 Riyals 98.9%
10000 Riyals 99.6%
20000 Riyals 97.2%
50000 Riyals 98.5%
100000 Riyals 94.3%
Average 96.89%
Table 5. Accuracy of nonzero digits identification after using the protocol
Nonzero
Digit
Accuracy in
Training Set
Accuracy in
Test Set
1(1) 99.24% 98.86%
2(2) 100% 100%
5(5) 100% 100%
Average 99.75% 99.62%
Table 6. Total accuracy for Iranian cash recognition after using the protocol
Cash Total Accuracy
1000 Riyals 93.4%
2000 Riyals 96.3%
5000 Riyals 98.9%
10000 Riyals 99.6%
20000 Riyals 97.2%
50000 Riyals 98.5%
100000 Riyals 94.3%
Average 96.89%
5. CONCLUSIONS
In this paper we presented a fast and accurate method for recognizing monetary value of Iranian
cashes. This proposed method is based on image processing and machine learning approaches. In
the proposed approach, as input we need a picture taken by mobile phone camera. So the
resolution can be very low. We take into account the limitation of visually impaired people for
taking pictures. So this proposed approach is robust to rotation of input picture, different scale of
the taken picture, different perspective of camera, variation of illumination, noisy background and
collision. Moreover, this method is robust to different design of cashes. All this robustness is

36
because of special component we focus on; all the processes in this method are on monetary value
of cashes.
In the first step of our method we fine the place of monetary value on the cash. This is down with
a string of image processing procedures and operators. After finding the place of monetary value
or more specially the place of zeroes we count the number of them, discover the rotation angle of
input picture and extract the nonzero digit of monetary value. In second step we identify the
nonzero digit by means of a MLP neural network with one hidden layer and 20 neurons in that
layer. Outputs of this neural network are 1, 2 or 5 digits based on the input image. Based on
accuracy of this step and using the proposed protocol, we could achieve total accuracy around
97%.
While the accuracy of this work is high in the field of Iranian cash recognition, and while it is the
only implemented method of Iranian cash recognition for visually impaired people, the accuracy
of it is not enough for a reliable detection. So one of the most important planes for future work is
increasing the accuracy by means of machine vision approaches or testing other methods for
nonzero digit identification. Besides that, because of the nature of proposed method that finds the
place of monetary value and recognizes it, this method has the ability to be extended for other
uses. For example it can be combined with object detection approaches to recognise price of
objects in shopping or it can be extended to identify the written monetary values on bills. Of
course this extension needs some modification in proposed approach.
ACKNOWLEDGEMENTS
The authors would like to thank Dr. Sayed Amir Hassan Monadjemi, all the survey and study
participant.
REFERENCES
[1] Hassanpour, H., & Farahabadi, P. (2009) "Using Hidden Markov Models for paper currency
recognition," Expert Systems with Applications, Vol. 36, No. 6, pp 10105-10111.
[2] Frosini, A., Gori, M. & Priami, P. (1996) "A neural network-based model for paper currency
recognition and verification," IEEE Transactions on Neural Networks, Vol. 7, No. 6, pp 1482-1490.
[3] Kosaka, T. & Omatu, S. (1999) "Bill money classification by competitive learning," in Proceedings of
IEEE Midnight-Sun Workshop on Soft Computing Methods in Industrial Applications.
[4] Kosaka, T., Omatu, S. & Fujinaka, T. (2001) "Bill classification by using the LVQ method," in
Proceedings of IEEE International Conference on Systems, Man, and Cybernetics.
[5] Lee, J., Jeon, S. & Kim, I. (2004) "Distinctive point extraction and recognition algorithm for various
kinds of euro banknotes," International Journal of Control Automation and Systems, Vol. 2, No. 2, pp
201-206.
[6] Jahangir, N., & Chowdhury, A. (2007) "Bangladeshi banknote recognition by neural network with
axis symmetrical masks," in Proceedings of IEEE 10th International Conference on Computer and
Information Technology, pp 1-5.
[7] Hinwood, A., Preston, P., Suaning, G. & Lovell, N. H. (2006), "Banknote recognition for the vision
impared," in Proceedings of Australasian Physical Engineering Sciences in Medicine, Vol. 29, No. 2,
pp 229-233.
[8] Reserve Bank of Australia, "How the RBA assists people with a vision impairment to differentiate
notes," http://www.rba.gov.au/CurrencyNotes/vision_impaired.html, accessed 7/6/2010.
[9] K-NFB reader website, Available from: http://www.knfbreader.com/, accessed: 11/18/2014.
[10] Krishna, S., Little, G., Black, J. & Panchanathan, S. (2005) "A wearable face recognition system for
individuals with visual impairments," in Proceedings of the 7th internationa ACM SIGACCESS
conference on Computers & accessibility, pp 106–113.

37
[11] Domínguez, A.R., Lara-Alvarez, C. & Bayro, E. (2014) "Automated Banknote Identification Method
for the Visually Impaired," in Progress in Pattern Recognition, Image Analysis, Computer Vision, and
Applications, Springer, pp 572-579.
[12] García-Lamont, F., Cervantes, J. & López, A. (2012) "Recognition of Mexican banknotes via their
color and texture features," Expert Systems with Applications, Vol. 39, No. 10, pp 9651-9660.
[13] Liu, X. (2008 ) "A camera phone based currency reader for the visually impaired," in ACM Proc.
ASSETS’08, Halifax, Nova Scotia, Canada, pp. 305-306.
[14] Hasanuzzaman, F. M., Yang X. & Tian, Y. (2009), "Robust and Effective Component-based
Banknote Recognition by SURF Features," in 20th Annual Wireless and Optical Communications
Conference (WOCC), pp 10105–10111.
[15] Grijalva, F., Rodriguez, J. C., Larco, J. & Orozco, L. (2010) "Smartphone recognition of the US
banknote's denomination, for visually impaired people," in IEEE ANDESCON, pp 1-6.
[16] Singh, T. R., Sudipta R., Singh, O. I., Sinam, T. & Singh, Kh. M. (2011) "A New Local Adaptive
Thresholding Technique in Binarization," IJCSI International Journal of Computer Science Issues,
Vol. 8, Issue 6, No 2.
[17] Serra, J. (1986) "Introduction to Mathematical Morphology," Computer Vision, Graphics and Image
Processing, Vol. 35, No. 3, pp 283-305.
AUTHOR
E. Nojavani born in 1987 in Bonab, Iran. He received his B.Sc. in computer software
from Urmia University, Urmia, Iran, in 2010. Currently he is a M.Sc. student in
Artificial Intelligence at the Department of Computer Engineering, University of
Isfahan, Isfahan, Iran. His research interest contains OCR, DIP, and Artificial
Intelligence.
S. A. Monadjemi is born in 1968 in Isfahan, Iran. He received his B.Sc. in electrical/
computer engineering from Isfahan University of Technology, Isfahan, Iran in 1992,
and his M.Sc. in computer Engineering, machine intelligence from Shiraz University,
Shiraz, Iran in 1995 and his PhD in computer engineering, processing and pattern
recognition, from Bristol University, Bristol, England in 2004. His research interests
are DIP, Machine Vision, Pattern Recognition, Artificial Intelligence, and Training
through Computer. Dr. Monadjemi is currently an Asst. Professor in Department of
Computer Engineering, University of Isfahan, Isfahan, Iran.
A. Rezaeezade born in 1988 in Aligudarz, Iran. She received her B.Sc. in computer
hardware from Isfahan University of Technology, Isfahan, Iran, in 2010. Currently
she is a M.Sc. student in Artificial Intelligence at the Department of Computer
Engineering, University of Isfahan, Isfahan, Iran. Her research interest contains OCR,
DIP, and Artificial Intelligence.

An Iranian Cash Recognition Assistance System For Visually Impaireds

More Related Content

What's hot (19)

Similar to An Iranian Cash Recognition Assistance System For Visually Impaireds (20)

More from IJSCAI Journal (20)

Recently uploaded (20)

An Iranian Cash Recognition Assistance System For Visually Impaireds