International Journal of Advanced Engineering Research and Technology (IJAERT) 119

Volume 2 Issue 4, July 2014, ISSN No.: 2348 8190

Feature Extraction and Classification of Handwritten Patterns using Neural

Networks
Dimple Bhasin*, Prof. Gulshan Goyal**, Dr. Maitreyee Dutta**
*CSE, Chandigarh Group of Colleges, Gharuan, Punjab
** CSE, Chandigarh Group of Colleges, Gharuan, Punjab
** CSE, NITTTR, Chandigarh

Keywords - Artificial Neural Network, Feature

Extraction, Filters Handwritten Pattern Recognition,
Pre-Processing, Thinning.

It is composed of a large number of highly

interconnected Processing elements (neurons) working
in unison to solve specific problems [2].
Advantages of neural network include:
Adaptive learning: An ability to learn how to do
tasks based on the data given for training or initial
experience.
Self-Organization: An ANN can create its own
organization or representation of the information it
receives during learning time.
Real Time Operation: ANN computations may be
carried out in parallel, and special hardware devices are
being designed and manufactured which take advantage
of this capability.
Fault Tolerance via Redundant Information
Coding: Partial destruction of a network leads to the
corresponding degradation of performance. However,
some network capabilities may be retained even with
major network damage [3].

1. INTRODUCTION

3. LITERATURE REVIEW

ABSTRACT
Handwritten Pattern recognition has been one of the
most captivating and interesting areas in the field of
Image processing and soft computing. The biggest
challenge faced by current technology is the recognition
of patterns or object accurately. To train machines for
the task of recognition is a tedious job due to variability
in patterns and different writing styles, yet they provide
useful solution thereby saving time and money. Neural
networks provide an efficient solution to the problem of
recognition. The paper presents a system that culminates
efficient steps of pre-processing along with the training
of neural network and hence providing better results of
recognition. The system aims to reduce processing time
while simultaneously improving recognition rates.

Machine simulation of human functions has been

cumbersome and challenging research area since the
advent of digital computers. Although extensive
improvements have been achieved yet humans
outperform even the most powerful computer in many
routine functions. Handwritten Pattern recognition is one
such area where machine simulation of human functions
is far from end [1]. Recognition of handwritten patterns
gets complicated due to numerous variations involved in
shape of patterns, and different writing styles that vary
from person to person. Neural networks with their
inherent learning ability provide efficient solution to
handwritten pattern recognition.
2. ARTIFICIAL NEURAL NETWORK
An Artificial Neural Network (ANN) is an information
processing paradigm that is inspired by the way
biological nervous systems, such as the brain, process
information. The key element of the ANN paradigm is
the novel structure of the information processing system.

Pattern recognition aims to classify patterns based on

either prior knowledge or on statistical information
extracted from patterns [5].It has been one of the
fascinating and challenging research areas. Tremendous
research has already been done and it is still being done
to get a more efficient and user friendly approach for the
problem. A brief review of various research papers is
laid down in following paragraphs.
In [4], a system is presented for recognition of
Devnagari handwriting using neural network. The entire
system consisted of 3 parts: Character Separation, Preprocessing comprising of Image binarization, Thinning
and Windowing. Character recognition and classification
is done using multi layer perceptrons trained by error
propagation algorithms. Also an important step is
deriving of input from character matrix. The presented
system is able to recognize most of handwritings. Also
the size of the dataset imposes a large impact of
recognition rate.

www.ijaert.org

International Journal of Advanced Engineering Research and Technology (IJAERT) 120

Volume 2 Issue 4, July 2014, ISSN No.: 2348 8190

In [5], three different soft computing techniques of feed

forward neural network namely evolutionary algorithm,
Back propagation algorithm and hybrid evolutionary
algorithm is used for recognition of handwritten English
alphabets. The techniques have been evaluated on the
basis of number of iterations and count of convergence
weight matrix. Also the algorithms are compared on the
basis of speed and computational cost. As a result of all
computations and comparisons hybrid evolutionary
algorithm is proven to be the best among the three.
In [6], a system for recognition of unconstrained offline
texts using hybrid Hidden Markov model/artificial
neural network is presented. The performance of system
is evaluated in terms of word error recognition rate and
obtained a 42 percent improvement compared to systems
adopted as baseline systems for comparisons.

4. RECOGNITION PROCESS
The whole recognition process for offline handwritten
patterns consists of following stages as shown in Fig. 1:
1) Image acquisition
2) Pre-processing
3) Feature Extraction
4) Training and Classification
4.1 Image Acquisition
The
images
used
are
obtained
from
www.mathwork.com. The available dataset comprises of
650 images of 26 English alphabets out of which 260
images have been sorted. Each image is of size 50x50
having bmp format. A sample of dataset is given in Fig.
2.

Image Acquistion

Pre-Processing

Feature Extraction

Training and
Classification

Fig. 2: Dataset of Handwritten Input images.

4.2 Pre-processing
Pre-processing consist of different steps to mould the
input image into a form suitable for processing. The
main objectives of pre-processing are:
1) Noise reduction
2) Normalization of the data
3) Compression in the amount of information to be
retained.
The presented Pre-processing approach onsists of:
1. Image Cropping
2. Noise reduction
3. Image Binarization
4. Cropping to edges
5. Image Complement
6. Normalization
7. Thinning
These steps involved in pre-processing are shown in
Figure 3.

Fig. 1: Schematic diagram of the Recognition process

www.ijaert.org

International Journal of Advanced Engineering Research and Technology (IJAERT) 121

Volume 2 Issue 4, July 2014, ISSN No.: 2348 8190

Image Cropping

Noise removal

Image binarization

Cropping to Edges
Fig.5: Different types of filters

Image Complement

4.2.2.1 Performance evaluation of filters

Normalization

Thinning
Fig. 3: Design of Pre-processing Approach for Offline
Handwritten Images

4.2.1 Image cropping

It involves removal of surplus pixels and blank spaces
from the image. It can be automated or done manually.
In the present work, cropping is done manually using
imcrop command present in MATLAB as shown in
Fig. 4.

Different filters that can be applied to the image are

evaluated in terms of:
1. Mean Square error (MSE)
2. Signal to noise ratio (SNR)
3. Peak signal to noise ratio (PSNR)
4. Root mean square error (RMSE)
Smaller the value of MSE, and larger the value of SNR
and PSNR, lesser will be the noise and higher the quality
of image.
According to Table 1, Wiener filters possess lowest
value of MSE and highest value of SNR and PSNR. It
reduces noise and improves the quality of images
effectively. Hence, Gaussian filter is used for filtering.
Table 1: Comparative analysis of different performance
evaluation parameters

4.2.2 Filtering
It aims to remove noise and spurious points, introduced
by uneven writing surface and poor sampling rate of data
acquisition device [5]. Main types of filtering methods
are shown in Fig.5.

a)

b)

Fig. 4: a) Original Image b) Cropped Image obtained using

imcrop command.

www.ijaert.org

International Journal of Advanced Engineering Research and Technology (IJAERT) 122

Volume 2 Issue 4, July 2014, ISSN No.: 2348 8190

4.2.3 Image binarization

Binarization involves conversion of a gray scale image
into a binary image where each pixel taking a value of 0
or 1.
4.2.4 Cropping to edges
It involves resizing the image in accordance with the
given window.

operate on a character stroke, which is only one pixel

wide [4] [7].
In the present work, thinning is done using fast parallel
thinning algorithms namely Zhang Suen algorithm and
Guo Hall algorithm.
4.2.6.1Performance evaluation of thinning algorithm
Thinning rate: The aim of image thinning is to obtain
the skeleton of the image. In order to evaluate the
thinning algorithm, the thinning rate (TR) is:
= 1

1
100%
0

Fig. 6: Image cropped to edges

4.2.5 Image complement

Image complement involves conversion of image into its

negative. This is done since it is pre-requisite steps for
thinning. An example of image complement is shown in
Fig. 7.

Where T0 is the number of object pixel in original

image, T1 is the number of object pixels in thinning
image.
Elapsed time or Execution time: It is defined as the
total time take to complete the task of thinning.
Generally, large TR and smaller value of execution time
indicates higher thinning degree [9]. A comparative
analysis is presented in Table 2.Since thinning rate
obtained by Guo Hall algorithm is greater, it can be used
for the thinning purposes.
Table 2: Comparative analysis of different thinning
algorithms.

Fig. 7: Image Complement

4.2.6 Normalization
It aims to obtain a standardized data. In the present works
binarized and complemented image is normalized to a size of
25x25 as shown in Fig. 8.

a)

b)

Fig.8 a) Cropped image b) Resized image 25x25

4.2.6 Thinning
Thinning is a morphological operation that removes
selected foreground pixels from binary images and is
particularly used for skeletonization [8]. Thinning can
remove irregularities in letters and in turn, makes the
recognition algorithm simpler because they only have to

www.ijaert.org

International Journal of Advanced Engineering Research and Technology (IJAERT) 123

Volume 2 Issue 4, July 2014, ISSN No.: 2348 8190

4.3 Feature Extraction

Feature extraction is done in order to extract relevant
information by reducing the dimension of the data. In
this step a pattern is analyzed and a set of features are
selected that can uniquely identify the pattern [21]. In the
present system, the thinned image is converted into a
5x7 matrix. This matrix contains a set of intensity
values. These values represent the minimum features
required to identify the pattern. This set of features is
further used as an input vector to the network.

4.4 Training and Classification

4.4.1 Neural Network
The network receives the 35 intensity values as a 35
input vector. It is then required to identify the alphabet
by responding with a 26-element output vector. In order
to operate accurately, the network needs to respond with
a 1 in the position of the letter being presented to the
network. All other values should be kept 0. The network
should be able to handle noise. The network does not
receive the intensity values as perfect Boolean values i.e.
0 and 1 only [4].
4.4.2 Architecture
A feed forward network is created that requires 35
neurons in the input layer and 26 neurons in the output
layer. The value of number of neurons in hidden layer is
chosen arbitrarily . The function newff creates a
feedforward network. The network is 2 layer logsigmoid/linear networks. The network is trained using
Levenberg-Marquardt algorithm. This algorithm is the
fastest among different back propagation algorithms.
The Created network is shown in Fig. 9 and training of
neural network is shown in Fig. 10.

Fig. 9: Neural network architecture

Fig. 10: Neural Network training

www.ijaert.org

International Journal of Advanced Engineering Research and Technology (IJAERT) 124

Volume 2 Issue 4, July 2014, ISSN No.: 2348 8190

5. RESULTS

Alphabets

The network training parameters opted is shown in Table

3. The system is analyzed on the basis of the number of
neurons in the hidden layer, Mean Square error and the
accuracy. Accuracy defines the rate of recognition.

A
B
C
D
E
F
G
H
I
J
K
L
M
N
O
P
Q
R
S
T
U
V
W
X
Y
Z

Feedforward neural network

parameters
Input nodes
35
Output nodes
Epochs
Learning rate

26

Training algorithm

LevenbergMarquardt
algorithm
10

Hidden layer
neurons
Performance goal

1000
0.75

0.01

Table 3: Feedforward Neural network training parameters

Number
of
Variance
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10

Number
of
Success
6
9
8
9
6
9
9
7
6
9
7
5
10
7
8
9
6
10
9
8
7
7
8
10
6
8

Accuracy
(%)
60
90
80
90
60
90
90
70
60
90
70
50
100
70
80
90
60
100
90
80
70
70
80
100
60
80

Table 4: Results

6. CONCLUSION

Fig.11: Performance Graph for a single character

The results are tabulated in Table 4. An accuracy of

80-100% was obtained for more than half of the
samples.

Offline handwritten patterns are difficult to recognize

due to high variability in writing styles and fonts from
person to person. The system presented is efficiently
working with the given network parameters and is able
to recognize a good amount of handwritings. The use of
an efficient pre-processing approach has improved the
recognition problem to a great extend. Thinning rate
imposed a great impact on the recognition rate. More the
thinning rate better is the recognition. Also recognition
rate is improved with the greater number of neurons in
the hidden layer. The only limitation is the number of
samples taken. However, the system yields good results
for the given samples and hence can be efficiently used
for different handwritings in different languages.

www.ijaert.org

International Journal of Advanced Engineering Research and Technology (IJAERT) 125

Volume 2 Issue 4, July 2014, ISSN No.: 2348 8190

REFERENCES
[1] Arica N., Yarman-Vural F.T. (2001), An Overview of
Character Recognition Focused on Off-Line Handwriting
IEEE transactions on systems, man, and cyberneticspart c:
applications and reviews, vol. 31, no. 2, pp.216-233.
[2] Jos C. Principe, Neil R. Euliano, Curt W. Lefebvre
Neural and Adaptive Systems: Fundamentals Through
Simulations, ISBN 0-471-35167-9.
[3] Ni D.X. (2007) Applications of neural network to
character recognition Proceedings of student/faculty research
day,CSIC Pace University.
[4] Rajput K.Y., Mishra S. Recognition and editing of
devnagari handwriting using neural network Proceedings of
SPIT-IEEE Colloquium and International Conference, vol. 1,
pp. 66-70.
[5]Shrivastava S.,Singh M.P. (2010)
Performance
evaluation of feed forward network with soft computing
techniques for handwritten English alphabets Applied soft
computing ,Elsevier, vol.11 pp. 1156-1182.
[6] Espana-Boquera S., Castro-Bleda M.J., Gorbe-Moya J.,
Zamora-Martinez F. (2011) Improving offline handwritten
text recognition with hybrid HMM/ANN models IEEE
transactions on pattern analysis and machine intelligence
vol.33, no.4, pp. 767-779.
[7] Indira B., shalini M., Murthy Raman M.V., Shaik M.S.
(2012) Classification and Recognition of Printed Hindi
Characters Using Artificial Neural Networks I.J. Image,
Graphics and Signal Processing, vol. 6, pp. 15-21.
[8] Shenbagavadivu S., Dr. Devi M.R.(2013) An
investigation of noise removing techniques used in spatial
domain image processing International Journal of Computer
Science and Mobile Computing Vol.2 Issue. 7, pp. 378-384.
[9] Dancheng Xu, Bailiang Li,Nijholt A. (2009) A novel
approach based on PCNNs Template for fingerprint Image
Thinning, Eight IEEE/ACIS International Conference on
computer and Information Science, pp. 115-119.
[10] Goyal G., Dr. Dutta M., Er. Girdhar A. (2010) A Parallel
Thinning Algorithm for Numeral Pattern images in BMP
Format.
[11] R.Gonzalez and R.E. Woods, Digital Image Processing,
Prentice Hall, 2011.
[12] Rani R., Kaur K. (2013) Experiment analysis of
different texture based features of image using simplified
Gabor Gaussian Wavelet transform International Journal of
Engineering and Advanced Technology (IJEAT), Vol. 2, pp.
365-368.
[13] K.purnima, T V Sampath Kumar, Lossless Image
Compression Using Traditional and Lifting Based Wavelet
Transform International Journal of Innovative Research and
Studies. ISSN 2319 -9725.
[14] Saeed K., Tabe Dzki M., Rybnik M., Adamski M. (2010)
K3M: A Universal algorithm for image skeletonization and a
review of thinning techniques International Journal of
Applied Mathematics and Computer Science, Vol. 20, No.
2,pp. 317335.
[15] Xu D., Li. B., Nijholt A. (2009) A novel Approach
Based on PCNNs Template for Fingerprint Image Thinning

Eight IEEE/ACIS International Conference on Computer and

Information Science, pp. 115-119.
[16] Ahmed P. (1995) A neural network based dedicated
thinning method Elsevier, pp. 585-590.
[17] Shang L., Yi Z. (2007) A class of binary images
thinning using twp PCNNs Elsevier, pp. 1096-1101.
[18] Hallale S.B., Salunke G.D. (2013) Offline and
handwritten digit recognition using neural network
International Journal of Advanced Research in Electrical,
Electronics and Instrumentation Engineering Vol. 2, Issue 9.
[19] Yadav D., Sanchez-Cuadrado S., Morato J. Optical
character recognition for Hindi language using a neural
network approach (March 2013) J. Inf. Process Syst. vol. 9,
no.1 pp 117-138.
[20] Kumar H. and Kaur P. (2011) A Comparative Study of
Iterative Thinning Algorithms for BMP Images
(IJCSIT) International Journal of Computer Science and
Information Technologies, Vol. 2 (5), pp. 2375-2379.
[21] Vipin, Dass R., Rajni (2013) Character recognition
using neural network, International journal of advanced
trends in computer science and engineering, pp. 62-67.

www.ijaert.org