Paper currency recognitionPaper Currency Recognition System
using Characteristics Extraction and Negatively Correlated NN
Ensemble Paper currency recognition are significant in many
applications. The requirements for an automatic banknote
recognition system offered many researchers to build up a robust
and dependable technique. Speed and precision of processing are two
vital factors in such systems. Of course, the precision may be much
significant than the speed. The designed system should have an
important precision in detecting torn or worn banknotes. The
currency recognition is one of the significant application domains
of artificial neural networks. This paper discusses the ENN for
currency recognition. NCL was used for the training of the network.
The use of NCL is to produce the diversity among the individual
networks in ensemble. The final decision of the network is taken
from voting among the individual NN.
Literature ReviewPresently ,there are a number of methods for
paper currency recognition: Using symmetrical masks technique for
recognizing paper currency in any direction. Other method:1. The
edges of patterns on a paper currency are spotted.2. In the next
step, paper currency is divided into N equal parts along vertical
vector.3. Then, for each edge in these parts the number of pixels
is added and fed to a three-layer, back propagation neural
network.4. In this process, to conquer the problem of recognizing
dirty worn banknotes, the following linear function is used as a
pre-processor:
f(x) = Fax + Fb (1)
where x is the given (input) image in gray scale, f(x) is the
resultant image;and Fa= 3 , Fb = -128 and N =50
other method use infrared or ultraviolet spectra may be used for
discriminating between genuine and counterfeits notes. Most of
paper currency recognition techniques use a single multilayer
feed-forward NN for the recognition. These uses edge detection
technique for feature extraction. This reduces the network size.
For new notes feature extraction from edge detection is simple. But
for the noisy notes it is very difficult. If a network takes a
false classification it will be not practical. So a single network
is not reliable enough. Therefore ENN is presented in this paper to
solve this problem.
Characteristics extraction Size The first phase of recognition
in the algorithm considers size of the banknote. The edges of
banknotes are generally worn and torn due to circulations. Hence,
its size is reduced, or even is increased slightly in rejoining the
torn banknotes. the size condition in the decision tree is
presented as: | x x0 |< dx & | y y0 |< dy (2) Where x0
and y0 are size of the testing paper currency, and x and y are size
of the reference paper currency.dx and dy are vertical and
horizontal directions.
ColorImage of the banknote is transformed to an image in gray
scale]. Then the gray scale level is reduced to have a significant
judgment about the background color.
TextureFor recognizing the template, Markov chain concept is
used in representing random phenomenon. A random process {xk, k =
0, 1, 2....} is called a Markov chain if the possibility value in
state xn+1 depends on just the possible value in state xn , that
is:
P(xn+1 = | xn = , xn-1 = n-1 ,..,x0 = 0 ) = P(xn+1 = |xn = )
This possibility can be shown by Pij. The state space of a
Markov chain can be shown in a matrix that is:
P=
where n is the number of states in the chain.
Steps for paper currency recognition :1. Banknote Size is
calculated. If its size satisfies equation (2) it is considered as
a possible true banknote.2. The banknote image histogram is
calculated.3. The transition matrices (Nx and Ny) are calculated
then, the main diagonal elements of the matrices (namely Dx and Dy)
are taken out as a feature for distinguishing between different
denominations.4. The paper currency under observation is assigned
to a denomination class if the Euclidean distances between the main
diagonal elements of its transition matrices (Dx and Dy) and the
main diagonal elements of the corresponding matrices of the
reference banknote (DRx and DRy ) are smaller than a predefined
value.5. At the end, the computed histogram in stage 2 is compared
with the histogram of the winner class in stage 4. If the Euclidian
distance between the two histograms is larger than the predefined
value, the banknote is assigned to an unknown class.
An approach using negative correlation learning NCL is used for
the training of the network. The use of NCL is to produce the
diversity among the individual networks in ensemble.
Assume a training set S of size N. S = {(x (l), d(l), x(2),
d(2)),..(x(N), d(N))} Where x is the input vector and d is the
desired result. Consider approximating d by forming an ensemble
whose result F(n) is the average in the component NN result
Fi(n)
(5)
Where M and n refer to the number of NN in ensemble and training
pattern, respectively. The error function Ei of the network i in
NCL is given by the following eq (6).
(6)
(7)
Where Ei(n) is the value of the error function of the network i
for the nth training pattern. The first term of (7) is the
empirical risk function of the network i. In the second term, Pi is
a correlation penalty function is given by eq (8).
(8)
The partial derivative of Ei(n) with respective to the output
network i on the nth training pattern is
= = = (1-) (Fi(n) d(n)) + (F(n) d(n)) The NCL is a simple
extension to the standard Back-propagation algorithm [8]. In fact,
the only alteration that is needed is to compute an extra term of
the form for the ith network. During the training process, the
entire ensemble interacts with each other through their penalty
terms in the error functions. Each network i minimizes not only the
difference between Fi(n) and d(n) , but also the difference between
F(n) & d(n). That is, negative correlation learning considers
errors what all other networks have learned while training a
network.
Comparative study of different paper currency and coin
recognition method
Currency has great importance in day to day life so currency
recognition is a great area of interest .We can conclude that image
processing is the most popular effective method of currency
recognition .Image processing based currency recognition technique
consists of:1. Image acquisition (using cameras or scanners).2.
Pre-processing (features extracting).3. Recognition of
currency.
Currency can be of two types:1. coin currency.2. paper
currency.
Coin currency recognition methodCoin recognition by method to
designed a neural network (NN) by using a genetic algorithm(GA) and
simulated annealing.(2000) Effective in a small number of input
signals. Small size neural network is developed. Low cost. Accuracy
is 99.68%.
Coin-o-omatic(2006) Designed to perform reliable classification
of heterogeneous coin collection. Uses combination of coin
photographs and sensor information in classification. Perform
automatic classification of coin in :1. Segmentation.2. Feature
extraction (using edge angle-distance distribution).3.
Pre-selection.4. Classification (nearest-neighbor).5. Verification.
Accuracy is 72%.
Image abstraction and spiral decomposition based system (2007)
Obtain abstract image (considering strong edges) from the original
image. Features extraction (spiral decomposition method) Spiral
distribution of pixels is the key concept that enables the system
to recognize the similarity between a full color multicomponent
coin images. No cost in image segmentation.
Image based approach using Gabor wavelet (2009) Extract features
for local texture representation. Divide image into small section
(using concentric ring structure). Statistics of Gabor coefficients
within each section is concatenated into a feature vector for whole
image. Matching between two coin image (via Euclidean distance and
nearest neighbor). Accuracy of 74.27%.Paper currency
recognition
Paper currency recognition for euro using three layer perception
and radial basis function (RBF) (2003) Used three layer perception
for classification and RBF for validation. RBF network has a
potential to reject invalid data.
Currency recognition using ensemble neural network (ENN) for
TAKA (bangladesi currency) (2010) Neural network in ENN is in fact
a classifier trained via negative correlation learning (NCL). The
currency image converted to gray scale and then compressed, each
compressed pixel is an input to the network. ENN is useful in
different types of currency.
Block LBP(local binary pattern) for characteristics extraction
in paper currency recognition (2010) Is improved version of LBP.
Works in two phases:1. Model creating : Preparing template.
Features extracting.2. The verification High recognition speed.
High classification accuracy.
Side invariance paper currency recognition based on matching
input note image with database of note image (2012)Overall process
are: Image acquisition and segmentation. Dimension matching.
Template matching. Decision making.
Recent developments in paper currency recognition system
Main steps in any currency recognition are:Matching
algorithmFeature extractingCurrency note localization(edge
detection and segmentationImage aquistion
Final output decision
Image acquisition: getting currency image by digital camera or
scanner. Edge detection: identifying the points at which the image
brightness changes sharply. Image segmentation: dividing the image
into its constituent regions or object. Feature extraction: one of
challenging tasks, identify the unique and distinguishing features
of each denomination under condition like old ,torn and worn notes.
Matching algorithm: classifies the currency notes.Potential
applications1. Assisting blind and visually impaired people.2.
Distinguishing original note from counterfeit currency.3. Automatic
selling goods.4. Banking service and applications.Related work
Image acquisition : is the creation of digital image. Image
pre-processing: Enhance some image features1. Image adjusting:
reducing the image size.2. Image smoothening: by applying mask on
the image higher is the size of mask, more is the smoothing. Edge
Detection: is the fundamental tool in feature detection and
extraction, reflects sharp intensity in colors and identifies
object boundaries using Sobel,Prewitt,Robert and Canny. Canny is
more powerful as it can detect true weak edges. Boundary
Subtraction: detect and recognize the note, black pixels touching
the boundary of the image were regarded as background, as note had
a white background. Feature Extraction: includes feature of serial
numbers of currency notes,effects on design and performance of the
classifier. Evaluation algorithm: after getting features of
currencies which then will be recognized by effective recognition
system called classifier, one of the most common techniques is
Artificial Neural Network (ANN). The Neural Network (NN) consists
of three layers: input layers, hidden layers and output layers.
Acquired EGB image convert to gray scale. Edge detection done on
whole gray scale. Paper currency characteristics are cropped and
segmented, and then extracted. Comparing these with the original
pre-stored image in the system. If matching it is genuine otherwise
it is counterfeit. The neural network evaluate the hue and
saturation threshold of input image ,if the Neural Network
threshold is less than current image threshold it is genuine
otherwise it is counterfeit.
Other work (2011) They used image histogram based on plenitude
of different colors of notes calculated and compared with reference
note. Also they used Markov chain concept to model texture of paper
currency as a random process. Finally they used Ensemble Neural
Network (ENN) with negative correlation in classification. ENN has
better performance than single network.Discussion Artificial Neural
Network based currency classification is the most frequently used
method like Feed Forward Network, Back Propagation Neural Network,
RBF network and ENN. There are also some models developed by
researchers like Markov chain.
A Novel Paper Currency Recognition using Fourier Mellin
Transform, Hidden Markov Model and Support Vector Machine
The needs for an automatic banknote recognition system
encouraged many researchers to develop fast, accurate, reliable and
robust technique which also be able to adapt to high noise.this
task can be:
1. Extracting features from paper currency images that vary from
each denomination.2. Using these features in an intelligent system
for recognition.
Recognition systems should be able to recognize paper currency
from each side and torn or worn banknotes.these systems depends on
currency note characteristics.This paper presents a new paper
currency recognition technique that independent to the number of
paper currency classes using texture characteristics of paper
currency. Texture modeled using Hidden Markov Model(HMM).
Existing paper currency methods: Recognition paper currency in
any direction using symmetrical masks: Compute sum of non-masked
pixels value and fed to a neural network. Uses two sensors one for
front and the other for back of paper currency, but only the front
image is the criterion for decision. Other research: Detect edge of
patterns on paper currency. Divide paper currency into N equal
parts vertically. For each edge of parts number of pixels counted
and fed to a three-layer, back propagation neural network.
Overcoming the dirty worn banknotes by function:
f(x) = Fax+Fd
where x is the input image in gray scale, f(x) is the output
image and selecting Fa=3, Fd=-128, N=50
Wiener filterIt is a filter used to reduce the effect of dirt
through improving image lightness.Wiener filter estimate local mean
and variance around each pixel as:
where is the N-by-M local neighborhood of each pixel in the
image then,a pixel-wise wiener filter using above estimation has
been created:
Where v2 is the noise variance. if the noise variance not given
wiener uses the average of all estimated variance.
(a)
(b)
(c)
(d)Fig 1:Using the Wiener filter to reduce the dirt from the
worn banknote. a)Original image before filtering, b)Obtained image
after filtering. c)Worn image before filtering d)Worn image after
filtering.
Fourier-mellin TransformIt is a powerful tool for image
recognition as: its resulting spectrum is invariant in rotation,
translation and scale. The Fourier Transform (FT) itself is
translation invariant and its conversion to log-polar coordinates
converts the scale and rotation differences to vertical and
horizontal offsets that can be measured. A second FFT, called the
Mellin transform gives a transform-space image that is invariant to
translation, rotation and scale . For an input image, I[m,n], the
Fourier-Mellin transform is defined as below:
(a)
(b)
(c) Fig 2:Using the Fourier-Mellin transform to reduce the
effect of rotation of the banknote: a) Original image, b) Rotated
image c) Output image after applying Fourier-Mellin transform.
Hidden Markov Model (HMM) By using the hidden Markov Model the
texture of a banknote is modeled as a random process. A random
process is called a Markov chain if the possibility value in state
depends on just the possible value in state that is:
P=
Where n is the number of states in a chain. In a discrete time
Markov chain, the possibility value of different states in the
matrix is computed as follow:
Where is the number of transitions from state i to state j.
Considering the above equation , matrix can be multiplied by the
factor
In order to obtain the below equation, this matrix is used to
differentiate between textures in different denominations. We can
scan the banknotes from top to bottom and from left to right to
obtain the transition matrix across the row (Nx) and across the
column (Ny)
N= =
An image is recognized by the value of its pixels at different
places, the way that adjacent pixels vary can also be used to
distinguish different images. Considering a paper currency image as
an matrix that is shown in table1, the value of each pixel like ij
can be considered as one state.It is clear that using the lower
numbers of gray scales leads to a lower computational load.
Therefore, in this paper the main diagonal values of the matrix N
have been used. Each element of the main diagonal value represents
the number of times that the corresponding value repeated in
adjacent pixels. The transition matrix (N) for a 5 euro banknote
shown in Figure 3 that has been quantized in 11 level gray scales
is shown in Table 1. In this table, for example, the 39 represents
the number of times the values of adjacent pixels are between 0 and
22. Our investigations indicate that the main diagonal values are
sufficient to distinguish different denominations. The main
diagonal values of seven different euro banknotes are illustrated
in Table 2. As it shows, all of the banknotes are distinguishable
using the main diagonal value. It should be noted that the dominant
colors of the banknote is recognizable using the main diagonal
values. Our investigations indicate that using the main diagonal
values is robust against worn banknotes. To show the robustness of
this feature, the main diagonal values of clean and dirty 5 euro
banknote are shown in table3. Because of the monotonic impurity in
dirty banknotes, the effect of the impurity on the main diagonal
values is negligible.
Fig 3: Image of 5 Euro banknote
Table 1. Transition matrix for a 5 Euro banknote45 19 16 8 7 5 5
4 4 0 0
35 402 238 50 38 25 9 16 21 4 0
9 235 871 356 114 90 48 25 15 9 5
7 50 337 1553 459 183 116 39 33 17 5
3 29 150 556 1938 700 231 100 74 35 12
2 26 59 157 876 7077 1726 240 125 70 28
5 19 49 51 222 1860 3406 467 213 126 26
6 52 26 26 75 256 486 793 675 331 39
1 5 13 23 46 107 198 634 6726 2098 60
0 1 15 17 48 67 198 409 1978 3769 80
0 0 3 2 5 16 21 38 67 103 49
5euro 45 402 871 1553 1938 7077 3406 793 6726 3769 49
10euro 0 35 609 585 512 2284 2646 6804 12852 2760 5
20euro 9 451 1359 461 1007 4821 2025 5372 11825 245 8
50euro 0 261 985 777 2122 4753 6051 6806 19935 16754 19969
100euro 0 119 1405 910 1430 2503 5725 5648 8204 1427 34
200euro 0 83 811 768 678 913 1605 3287 5160 15485 90
500euro 0 98 1255 876 1429 2615 8082 2457 2412 8895 63
Table 2. Transition matrix for 7 types Euro banknotesOriginal 5
euro 45 402 871 1553 1938 7077 3406 793 6726 3769 49
Dirty 5 euro 39 398 813 1456 1892 6981 3355 721 6451 3689 24
Table 3. Transition matrix for original 5 Euro and dirty 5
euro
Finally from the four paper I have read I found:
Currency denomination recognition is one the active research
topics at present , and this wide interest is due to: Monetary
transaction is an integral part of our day to day activities. blind
people particularly suffer in monetary transactions. They are not
able to effectively distinguish between various denominations and
are often deceived by other people. Also there are many
applications that depends on currency recognition like:1. Assisting
visually impaired people-2. Distinguishing original note from
counterfeit currency-3. Automatic selling-goods-4. Banking
Applications- Paper currency are coins or paper. paper is
significant rather than coins as they become old early.
Methodology of any currency recognition system the main steps
:
Final output decisionMatching algorithmFeature
extractingCurrency note localization(edge detection and
segmentationImage aquistion
Most of matching algorithm used Artificial Neural Network
specially Ensemble Neural Network as its better performance than
single Neural Network.Also the most common model technique is the
Markov Chain.