New Approach of Preprocessing For Numeral Recognition

S. Benchaou et al Int. Journal of Engineering Research and Applications www.ijera.com
ISSN : 2248-9622, Vol. 4, Issue 7( Version 3), July 2014, pp.26-30
www.ijera.com 26 | P a g e
Image in color Transformation
into grey levels
Binarization Cropping Normalization
New Approach of Preprocessing For Numeral Recognition
S. Benchaou*, M. Nasri*, O. El Melhaoui*, B. Bouali**
*(Labo MATSI, EST, University Mohammed 1, Morocco)
** (Labo AGA, FS, University Mohammed 1, Morocco)
ABSTRACT
The present paper proposes a new approach of preprocessing for handwritten, printed and isolated numeral
characters. The new approach reduces the size of the input image of each numeral by discarding the redundant
information. This method reduces also the number of features of the attribute vector provided by the extraction
features method. Numeral recognition is carried out in this work through k nearest neighbors and multilayer
perceptron techniques. The simulations have obtained a good rate of recognition in fewer running time.
Keywords – features extraction, handwritten and printed numeral recognition, k nearest neighbors, multilayer
perceptron , preprocessing.
I. INTRODUCTION
For the last decade, the recognition of
handwritten characters has known a big progress in
numerous industrial applications, in particular the
numeral recognition which is used in several sectors
such as postal sorting, bank check reading, order form
processing, etc.
The recognition system requires three main steps
to predict the class of membership of an unknown
pattern, which are: preprocessing, features extraction
and classification.
The preprocessing phase consists of discarding
the imperfections and reducing the analyzed area.
The Numeral features extraction is a delicate
process and is crucial [1] for a good numeral
recognition. It consists of transforming the image into
an attribute vector which contains a set of
discriminated characteristics for recognition and also
reducing the amount of information supplied to the
system.
In the literature, several works have been
proposed for features extractions such as invariant
moments [2], Zernike moments [3], freeman coding
[4], Loci characteristics [5], etc.
The last step is the classification which consists
of partitioning a set of data entity into separate classes
according to a similarity criterion, different methods
are proposed in this context includes dynamic
programming, neural networks, support vector
machines, k nearest neighbors, k-means, etc.
In order to validate our contributions, we have
used in this work a database of 590 numerals, printed
and handwritten, provided by various categories of
writers.
In section 2 numeral features extraction method is
presented. The k nearest neighbors and the multilayer
perceptron techniques of classification are discussed
in section 3. The proposed system for numeral
recognition is presented in section 4. The result of
simulations and comparisons are introduced in section
5. Finally, we give a conclusion.
II. NUMERAL FEATURES EXTRACTION
Features extraction is a complex task. It knows a
great interest in multiple domains such as image
processing, segmentation, classification, etc.
After the preprocessing step, the preprocessed
image is represented by a matrix of pixels which can
be of very large size. So, it will be useful to represent
objects by characteristics containing the necessary
information. This operation is called features
extraction. In the literature, the feature extraction
methods are classified according to two categories:
statistical approach and structural approach. In our
work we are going to use the statistical approach
based on profile projection.
2.1 Profile projection
Profile projection is a statistical method
successfully used in recognition of handwritten
characters [6], etc.
The preprocessing stage consists of binarizing the
numeral input image which is presented in grey level.
The next stage is to only preserve the numeral
position in image by cropping it. Final stage is to fix
the size of cropped numeral image.
Fig. 1: Different steps for preprocessing of
numeral « 5 ».
This method calculates the number of pixels
(distance) between the left, bottom, right, top edge of
the image and the first black pixel met on this row or
column. The dimension of the obtained attribute
RESEARCH ARTICLE OPEN ACCESS

vector is twice the sum of the number of rows and
columns associated to the image of the numeral.
Fig. 2: The four profile projections of numeral «4»
III. CLASSIFICATION METHODS
3.1 K nearest neighbors
K nearest neighbors (KNN) is a widely used
method for data classification. Proposed in 1967 by
Cover [7], it has been widely used in handwritten
numerals recognition for its simplicity and its
robustness [8].
KNN is a method which was inspired from the
closest neighbor rule. It is based on computing the
distance between the test sample and the different
learning data samples and then attributes the sample
to the k nearest neighbors.
3.2 Multilayer perceptron
Neural networks are one branch of artificial
intelligence. Artificial neural network is defined as a
computer system who has been inspired from a long
study of human brain and how the human neurons
function [9].
Artificial neural networks in general and
multilayer perceptron (MLP) in particular have been
widely used for data classification, pattern recognition
(characters, voice, signal, etc.) [10][11], prediction
[12], etc. The MLP is characterized by its ability to
learn and gradually improve its performance through a
learning process.
MLPs are forward propagation networks where
the two closest layers are fully connected. The MLP
structure contains an input, an output and a certain
number of hidden layers.
We have been limited, in this work to one hidden
layer.
Learning is a phase where the behavior of the
network is modified by modifying the synaptic
weights until a desired output pattern is obtained.
In this work, we have used the gradient
backpropagation algorithm, the objective is to
minimize the squared error between the desired and
computed output of the MLP. Algorithm 1 shows the
different steps of the gradient backpropagation
algorithm.
Algorithm 1 : Gradient backpropagation
1. Randomly initialize the synaptic weights
between -1 and 1.
2. Randomly apply a realization vector of an
object to the input layer and its corresponding
known output to the output layer.
3. Compute the network output and error E
between computed and desired outputs.
4. Adjust the weights by the gradient method:
( )
( ) ( )
( 1) ( )
r
r r
W
E
W t W t


  
η is the learning rate, which is in general a value
between 0.1 and 0.9, r = 1,2.
5. Go to 2 as long as the network does not show
satisfactory performances.
We have a structure of three layers:
- The input layer has n1 neurons that we call nek,
1≤k≤n1,
- The hidden and output layers contain n2 and n3
neurons that we respectively call ncj and nsi, where
1≤j≤n2, 1≤i≤n3.
(1)
k z is the realization of the attribute vector
component for a numeral image I, with k=1,2,….n1.
( ) (1) (1)
jk W  w , j=1….n2, k=1…..n1, (1)
jk w are the
synaptic weights connecting neurons in the input to
the neurons in the hidden layers.
( ) (2) (2)
ij W  w , i=1….n3, j=1…..n2,
(2)
ij w are the
synaptic weights connecting neurons in the hidden to
the neurons in the output layers. The output of the
neuron j, ncj of the hidden layer is:
( ) (2) (2)
z j  f y j
(1)
Where 


1
1
(2) (1) (1)
n
k
y j wjk zk (2)
And (2)
1
1
( ) (2)
j j y
e
f y


 (3)
For j=1,2,…,n2, f is the activation function which
we choose to be of sigmoid type.
The output of the neuron i , nsi of the output layer
is:
( ) (3) (3)
zi  f yi
(4)
Where 


2
1
(3) (2) (2)
n
j
i ij j y w z (5)
For i= 1,2,…,n3.
Notice that the superscripts (1), (2) and (3)
represent respectively input, hidden and output layers.
 Optimization of the (2)
ij w :

Binarization Cropping
Windowing Normalization
In order to be able to update (2)
ij w synaptic weights
for different indices i and j, consider the network
error to be:
2
3
1
(3) ( ( ))
2
1
E(t) z di t
n
i
i   

(6)
Where di (t) is the known desired output at neuron
nsi. t represents the current iteration. Differentiating
E(t) with respect to (2)
ij w gives:
(2)
(3)
(3)
(2)
( ( ) ( ))
( )
ij
i
i i
ij w
z
z t d t
w
E t


 


(7)
Where ( ) (1 ( )) (2) (3) (3)
(2)
(3)
j i i
ij
i z f y f y
w
z
   


(8)
The synaptic weight
(2)
ij w may now be updated.
( 1) ( ) ( ) ( ) (2) (2) (3) (2) w t w t t z t ij ij i j    (9)
For i=1….n3 and j =1….n2
Where
( ) ( ( ) ( )) ( ) (1 ( )) (3) (3) (3) (3)
i i i i i  t  z t  d t f y   f y
(10)
 Optimization of the
(1)
jk w
The derivation in this case is similar to that
presented previously, so we go to the updating of the
weights according to:
( 1) ( ) ( ) ( ) (1) (1) (2) (1) w t w t t z t jk jk j k    (11)
For j =1….n2 and k= 1…n1
Where
( )
( ( ))
( ) ( ) ( )
(2)
(2)
1
(2) (3) (2)
3
y t
f y t
t t w t
j
j
n
i
j i ij


 

  (12)
IV. PROPOSED SYSTEM
We have used a database of 590 Arabic numerals,
provided by different writers. A sample of the
database is shown in figure 3. The database is divided
into two sets, one set of 400 numerals is used for
learning and the remaining 190 numerals are used for
the test stage.
Fig. 3: A sample of handwritten and printed
numerals
The proposed system consists of adding a step of
windowing in the preprocessing stage, in order to
reduce the size of the attribute vector, which
decreases the running time.
The proposed system contains three main steps,
preprocessing, features extraction and classification.
The full system is shown in figure 4.
Fig. 4: Scheme for numeral recognition
The preprocessing stage, in our case, is so fast in
computing time. It consists of binarizing the numeral
input image which is presented in grey level. Then,
we preserve only the numeral position in image by
cropping it. The next stage is to normalize the image
in a predefined size. Finally, we have applied the
windowing stage. This operation consists of dividing
the numeral image into small windows. For every
window we calculate the average of pixel grey level.
This operation reduces widely the size of the image
and preserves the useful information for recognition
of every input image. So, the dimension of attribute
vectors applied to the system will be reduced too.
The three last steps (cropping, normalization and
windowing) reduce also widely the computing time
of processing.
After preprocessing, features extraction is carried
out by profile projection method. KNN and MLP
methods are used for classification task.
V. EXPERIMENTAL RESULTS AND
COMPARATIVE STUDY
In the first test, the simulations of numeral
recognition are carried out without using windowing
stage in the preprocessing phase. Several experiments
were carried out to determine the recognition rate
according to the size of normalization and the
number of neurons in the hidden layer of the
multilayer perceptron.
Table 1 shows that the couple son and nc have
an effect on recognition rate. In fact, with the couple
(son,nc)= (40x40,40) we have achieved the
maximum recognition rate equal to 86.8%. The
attribute vector obtained contains 160 elements for
each numeral.
Image
in
grey
level
Features extraction :
Profile Projection
Classification :
- Multilayer Perceptron (MLP)
- K nearest neighbors (KNN)
Input Image

Table 1: Numeral recognition rate according to the size of normalization and the number of neurons in the hidden layer
Size of normalization (son)
Profile projection
Number of primitives
Number of neurons in the hidden layer (nc)
Recognition rate %
20x20
80
15
77.89
20
82
30
81.05
30x30
120
25
82.63
35
84
45
83.68
40x40
160
30
78.9
40
86.8
55
84
50x50
200
50
84.73
80
85.26
140
83.16
Figure 5 illustrates the evolution of recognition rate according to the number of neurons in the hidden layer, using the size 40x40 as size of normalization.
Fig. 5: Evolution of recognition rate according to the number of neurons in the hidden layer A normalization size of 40x40 obtains good recognition rate performance of isolated numerals. This size will be preserved for application of our proposed system. In the classifier, the number of neurons in the input layer is too high. It is equal to 160 which weakened the performance of the multilayer perceptron and increased the time of processing. For this reason, and in order to reduce the number of input neurons of the numeral recognition system, we have proposed the windowing step which consists to split the numeral image into equal areas of m pixels. For each area, the average of pixel grey level is calculated. This operation reduces the size of input image, attribute vector dimension, error rate and time processing. In the second test, we have fixed the size of normalization on 40x40. Then, we have applied our approach of windowing in preprocessing phase. Table 2 illustrates the numeral recognition rate using the proposed system after several simulations according to the number of hidden neurons. Table 2: Numeral recognition rate of proposed system
Table 2 highlights the fact that with the operation of windowing, the size of the attribute vector is reduced to 120 and 80 while it was initially 160. The time processing using the K nearest neighbors and multilayer perceptron is reduced too. We also notice that the choice of the size of windows is essential. The window of two pixels preserves the same recognition rate. It is equal to 86.8%. But the window of four pixels improves the recognition rate to 87.9%, this justifies that the redundant attributes are partially eliminated.
VI. CONCLUSION
In this work we have presented a new approach for numerals recognition. In the preprocessing stage we have added the step of windowing. It consists to split the numeral image into equal areas of m pixels. For each area, the average of pixels is calculated. Features extraction is carried out by profile projection. Multilayer perceptron method and K nearest neighbors are used for classification. The simulations have obtained good results; a clear reduction of attributes vector size and time processing with a good increase of recognition rate are obtained. That confirms the performance of our proposed system for numeral recognition.
REFERENCES
[1] R.G.1. Benne, B.V.1. Dhandra and M. Hangarge, “Tri-scripts handwritten numeral recognition: a novel approach”. Advances in
Size of Window-ing
Number of primitives
Profile projection
K nearest neighbors
Recognit-ion rate (%)
Time process-ing (min)
Recognit-ion rate (%)
Time processing (s)
Without windowing
160
86.8
13.58
86.8
31
Window of 2 pixels
120
86.8
13.5
86.8
24
Window of 4 pixels
80
87.9
4
87.9
19

Computational Research, ISSN: 0975–3273, Vol 1, Issue 2, pp: 47-51,2009.
[2] Anuja P. Nagare. “License Plate Character Recognition System using Neural Network”. International Journal of Computer Applications (0975-8887), Vol.25, N°.10, July 2011.
[3] N. Abu Bakar, S. Shamsuddin, and A. Ali. “An integrated formulation of zernike representation in character images”. In Nicolás García-Pedrajas, Francisco Herrera, Colin Fyfe, José Benítez, and Moonis Ali, editors, Trends in Applied Intelligent Systems, volume 6098 of Lecture Notes in Computer Science, pages 359–368. Springer Berlin / Heidelberg, 2010.
[4] X. Dupré, “Contributions à la reconnaissance de l'écriture cursive à l'aide de modèles de Markov”. Thèse, Université René Descartes-Paris 5, 2003.
[5] R.Ebrahim, M.R.Moradian, A.Esmkhani, F.M.Jafarlou. “Recognition of Persian handwritten digits using characterization Loci and mixture of experts”. International Journal of Digital Content.Technology and its Applications, Vol.3,N°.3, 2009 .
[6] D. Gokana, “Contribution à la Reconnaissance Automatique de Caractères Manuscrits. Application à la Lecture Optique de Caractères sur Supports Mobiles”. Thèse de Doctorat : Université de Paris Sud, Centre d’Orsay, 1986.
[7] R. GIL-PITA, X. YAO, “Evolving edited k- nearest neighbor classifiers”. International Journal of Neural Systems, Vol.18, N°.6, pp 459–467,2008.
[8] I. Kuncheva, “Editing for the k-nearest neighbors rule by a genetic algorithm”. Elsevier, Pattern Recognition Letters 16, pp: 809-814,1995.
[9] A. Adnan and S.Sameer, “Optical character recognition : Neural network analysis of hand-printed characters”. In SSPR ’98/SPR ’98 : Proceedings of the Joint IAPR International Workshops on Advances in Pattern Recognition, London, UK, Springer- Verlag, 1998.
[10] J.P.Pinto, “Multilayer perceptron based hierarchical acoustic modeling for automatic speech recognition”. Thèse, Ecole Polytechnique fédérale de Lausanne, Suisse, 2010.
[11] L.Oukhellou, “Paramètrisation et classification de signaux en contrôle non destructif. Application à la reconnaissance des défauts de rails par courant de foucault”. Thèse, Université Paris XI Orsay, 1997.
[12] H.Bouziane, B.Messabih, A.Chouarfia, “Prédiction de la structure 2D des protéines par les réseaux de neurones”. Communication of IBIMA, Vol.6, 2008.

New Approach of Preprocessing For Numeral Recognition

More Related Content

What's hot (17)

Viewers also liked (20)

Similar to New Approach of Preprocessing For Numeral Recognition (20)

Recently uploaded (20)

New Approach of Preprocessing For Numeral Recognition