88 92

ISSN: 2277 – 9043
International Journal of Advanced Research in Computer Science and Electronics Engineering
Volume 1, Issue 5, July 2012

ANN Implementation for Classification of Noisy
Numeral Corrupted By Salt and Pepper Noise
Smita K. Chaudhari* G.A.Kulkarni
Dept. Of E& C Engg SSGBCOET,Bhusawal Dept. Of E& C Engg SSGBCOET,Bhusawal
smita.c20@gmail.com girish227252@rediffmail.com

 of as an "expert" in the category of information it has been given to
Abstract— Neural Network (NN) is information processing analyse.
paradigm that is inspired by the way biological nervous Neural networks take a different approach to problem solving
systems, such as the brain, process information. Neural than that of conventional computers. Conventional computers use
Networks are known to be capable of providing good an algorithmic approach i.e. the computer follows a set of
recognition rate in presence of noise. Neural Network with instructions in order to solve a problem. Unless the specific steps
various architectures and Training algorithms have successfully that the computer needs to follow are known the computer cannot
been applied for letter or character recognition [1]. Numerals solve the problem. That restricts the problem solving capability of
Recognition is one of the artificial intelligence applications conventional computers to problems that we already understand and
which provide an important fundamental for various advanced know how to solve. But computers would be so much more useful if
applications, including information retrieval and they could do things that we don't exactly know how to do. Neural
human-computer interaction applications. The neural networks networks process information in a similar way the human brain
are also able to extract meaningful features of the digits, such as does.
edges. Handwritten recognition is complex due to large Object recognition is the study of how machines can observe the
variation of handwritten style whereas printed character environment, learn to distinguish patterns of interest and make
recognition is also difficult due to increase number fonts. reasonable decisions about the categories of patterns. The
This paper uses hamming netwok to recognize noisy numerals. performance of a machine may be better than the performance
The proposed algorithm will design a system which associates of a human in a noisy environment due to the factors: human
every fundamental pattern with itself. That is, when presented performance degrades with increasing number of targets; where as
with xi as input, the system should produce xi at the output. In the performance of a machine does not depend on the size of the
addition, when presented with a noisy (corrupted) version of xi set of targets. The performance of a machine does not degrade
at the input, the system should also produce xi at the output. due to fatigue caused by prolonged effort. A knowledge based
The recognition results of the noisy numeral showed that the system is desirable for reliable, quick and accurate recognition of
network could recognize normal numerals with 100% accuracy, objects from noisy and partial input images [3].
numerals added with salt and pepper noise at average of 89%. The McCulloch and Pitts model was utilized in the development
of the first artificial neural network by Rosenblatt in 1959 [11]. This
network was based on a unit called the perceptron, which produces
an output scaled as 1 or -1 depending upon the weighted, linear
Index Terms— Character Recognition, Hamming Network,
combination of inputs.
Noisy Numeral, Salt & pepper Noise, Neural Network.
The optical character recognition system for hand printed
numerals of noisy and low-resolution measurement consists of the
two-stage feature extraction process. In the first stage a set of
I. INTRODUCTION primary features insensitive to the quality and format of a
Recently, neural network becomes more popular as a black-white bit pattern are extracted. In the second stage, a set of
technique to perform character recognition. It has been reported properties capable of discriminating the character classes is derived
that neural networks could produce high recognition accuracy. from primary features. The system is simple and reliable in that only
Neural networks are capable of providing good recognition at three kinds of primary features are needed to be detected. The
the present of noise that other methods normally fail. recognition is based on the decision tree which tests the logic
An Artificial Neural Network (ANN) is information processing statements of secondary features. [12]
paradigm that is inspired by the way biological nervous systems, The importance of using a hierarchical network is shown
such as the brain, process information. The key element of this in literature [16] Seong-Whan Lee finds a new scheme for off-line
paradigm is the novel structure of the information processing system. recognition of totally unconstrained handwritten numerals using a
It is composed of a large number of highly interconnected simple multilayer cluster neural network trained with the back
processing elements (neurons) working in unison to solve specific propagation algorithm which avoids the problem of finding local
problems. ANNs, like people, learn by example. An ANN is minima & improves the recognition rates [10]
configured for a specific application, such as pattern recognition or H. K. Kwan introduced multilayer recurrent neural
data classification, through a learning process. Learning in networks in the form of 3-layer bidirectional symmetrical and
biological systems involves adjustments to the synaptic connections asymmetrical associative memories are presented. The networks
that exist between the neurons. This is true of ANNs as well. Neural possess the features of both a multilayer feedforward neural network
networks, with their remarkable ability to derive meaning from and a bidirectional associative memory. These networks can have
complicated or imprecise data, can be used to extract patterns and two modes of recalling, namely, recalling by one pattern and
detect trends that are too complex to be noticed by either humans or recalling by a pattern pair in[12]
other computer techniques. A trained neural network can be thought Recognition of Noisy Numerals using Neural Network
by Mohd Yusoff Mashor and Siti Noraini Sulaiman.This paper
uses MLP network trained using Levenberg-Marquardt algorithm
to recognise noisy numerals. The recognition results of the noisy

88
All Rights Reserved © 2012 IJARCSEE

ISSN: 2277 – 9043

numeral showed that the network could recognize normal indicate the correlation between the prototype patterns and the input
numerals, blended numerals[15]. matrix.
The neurons will compete each other to determine a winner. When
II. BACKGROUND & TERMINOLOGY. the processes are finished, there will be only one neuron with
nonzero output. This neuron indicates the prototype pattern that is
The Hamming network method was developed by a closest to the input.
mathematician, Richard W. Hamming. He has many contributions The processes in the recurrent layer will be divided into iterations.
not only in the mathematical field, but also for computer science and When one iteration is finished, a function will check whether there
telecommunication [5]. He was also the founder and has been the is only one nonzero output. If available so , the process in this layer
president of Association for Computing Machinery. Hamming will be stopped, and the process will continue to generate the output.
network method is developed to solve pattern recognition problems In Figure D is the function to check whether there is only one
which use binary format, such as a matrix with only two possible nonzero output. W2 is the weight matrix for this layer with the
values, 0 and 1. In the Hamming network, there is a matrix which dimension of S x S. The iteration number will be given as t, and it
stores the patterns of all objects, called the prototype data matrix. will be added by one until the iteration stopped. The activation
The patterns will not be learned by the system, but rather to be function which is used is the positive linear transfer function
stored as a matrix data. The matrix will be used to define the output (poslin). This function is linear for positive values and zero for
of the network. The objective of the Hamming network is to decide negative values.
which prototype matrix is closest to the input matrix. It calculates
the similarities between the prototype matrix of all objects and the
input. C. Salt and Pepper Noise.
It is designed explicitly to solve binary pattern recognition Salt and pepper noise is an impulse type of noise, which is also
problems. It has both feed forward and recurrent layer. The number referred to as intensity spikes. This is caused generally due to errors
of neuron in the first layer is the same as the number of neurons in in data transmission. It has only two possible values, a and b. The
the second layer. The objective of the hamming network is to probability of each is typically less than 0.1. The corrupted pixels
decide which prototype vector is closest to the input vector. are set alternatively to the minimum or to the maximum value,
This decision is indicated by the output of the recurrent layer. When giving the image a “salt and pepper” like appearance. Unaffected
the network converges, there will be only one nonzero output. This pixels remain unchanged. For an 8-bit image, the typical value for
indicates the prototype pattern that is closest to the input vector. pepper noise is 0 and for salt noise 255. The salt and pepper noise is
generally caused by malfunctioning of pixel elements in the camera
sensors, faulty memory locations, or timing errors in the digitization
process. The probability density function for this type of noise is
shown in Figure 2. Salt and pepper noise with a variance of 0.05 is
shown in Image 3

Fig.1 Hamming Network

A. Feedforward Layer Fig 2 . PDF for salt and pepper noise

Feedforward layer calculates the correlation between each
patterns of the prototype matrix and the input matrix (figure 1). The
calculation results will be processed to generate the output neurons
for this layer.
As shown in the figure 1, the layer has the input matrix from p,
which has the dimension as R x 1. This input matrix goes to the
weight matrix (W1) with the dimension of S x R. The net of this
layer (n1) will be the sum of the W1p and the bias input b. The
weight matrix of W1 will be the matrix of the prototype data which
include the patterns of all objects. The element of the bias b will be
given as the number of R. The transfer function which is used in this Fig 3. Salt & Pepper Noise.
layer is the linear transfer function (purelin). This function will not
change the value so the output of this feedforward layer (a1) will be
given as: a1 = purelin (W1p + b1) The output neurons of this layer
will be used as the initial input for the recurrent layer. III. LITERATURE SURVEY.
B.Recurrent Layer The neural networks were also able to extract meaningful
features of the digits, such as edges. The cascade correlation
The recurrent layer is also called as a competitive layer. In this network was the least successful, possibly because the network was
layer, there is a neuron for each prototype pattern. The neurons are committing itself to poor results on in training when few hidden
initialized with the output neurons of the feedforward layer, which units were present. It was found that an elaborate conjugate gradient
minimization technique yielded little improvement in generalization

89

ISSN: 2277 – 9043

performance and resulted in six times longer training time than H. K. Kwan introduced multilayer recurrent neural
ordinary backpropagation. The importance of using a hierarchical networks in the form of 3-layer bidirectional symmetrical and
network is shown in literature [9] Seong-Whan Lee finds a new asymmetrical associative memories are presented. The networks
scheme for off-line recognition of totally unconstrained handwritten possess the features of both a multilayer feedforward neural network
numerals using a simple multilayer cluster neural network trained and a bidirectional associative memory. These networks can have
with the back propagation algorithm which avoids the problem of two modes of recalling, namely, recalling by one pattern and
finding local minima & improves the recognition rates [10] recalling by a pattern pair in[12].
Le Cun et al. [19] achieved excellent results with a back Recognition of totally unconstrained handwritten numeral
propagation network using size-normalized images as direct input. strings. The system is built upon a number of components, named, a
Their solution consists of a network architecture which is highly presegmentation module, an isolated numeral recognizer, a
constrained and specifically designed for the task. There are four segmentation-free module and a merging module. Presegmentation
internal layers, two layers made of independent groups of feature consists in dividing continuous numeral string image into groups of
extractors and two layers which perform averaging/sub-sampling. numerals, each of which represents an integer number of numerals.
The last internal layer is fully connected to the ten-element output, For each group, the actual number of numerals and their identity are
but all other connections are local and use shared weights. In total, then determined by a cascade of two recognition-based tests:
there are 4,635 units and 98,442 connections but only 2,578 isolated numeral and segmentation-free. The last one is able to
independent parameters. recognize a numeral group of any length. All results from all groups
A modified quadratic classifier based scheme was used for the are eventually merged yielding the final interpretation of the input
recognition of off-line handwritten numerals of six popular Indian numeral string. The concept of dummy symbol in order to overcome
scripts such as Devnagari, Bangla, Telugu, Oriya, Kannada and the problem o f noisy parts that cannot be eliminated by standard
Tamil scripts. The features used in the classifier are obtained from filtering algorithms.[13]
the directional information of the numerals. For feature
computation, the bounding box of a numeral is segmented into IV. DESIGN AND IMPLEMENTATION OF THE
blocks and the directional features are computed in each of the SYSTEM.
blocks. These blocks are then down sampled by a Gaussian filter
and the features obtained from the down sampled blocks are fed to a
modified quadratic classifier for recognition.[20] The system designed in this paper associates every
Amit Choudhary analyzes the performance of back-propagation fundamental pattern with itself. That is, when presented with xi as
feed-forward algorithm using various different activation functions input, the system should produce xi at the output. In addition, when
for the neurons of hidden and output layers. For sample creation, presented with a noisy (corrupted) version of xi at the input, the
250 numerals were gathered form 35 people. After binarization, system should also produce xi at the output. The system which is
these numerals were clubbed together to form training patterns for developed is a system that gets an input of digit, process it through
the neural network. Network was trained to learn its behavior by the network, and generates the result. The digit which are used in the
adjusting the connection strengths at every iteration. The conjugate development are limited to printed digit from 1 to 9. The system has
gradient descent of each presented training pattern was calculated to some prototype data that consists of the pattern of digits, from 1 to
identify the minima on the error surface for each training pattern. 9. This prototype data is used as the weight matrix for the process in
Experiments were performed by selecting different combinations of the feedforward layer of the Hamming network. The system is built
two activation functions out of the three activation functions using the MATLAB and the images are processed using the
„logsig‟, „tansig‟ and „purelin‟ for the neurons of the hidden and Microsoft Paint.
output layers and the results revealed that the percentage The type of the image file is bitmap (.bmp) . The image is read &
recognition accuracy of the neural network was observed to be converted into 64×64 matrix form. This matrix is converted to 8×8
optimum when „tansig‟-„tansig‟ combination of activation functions matrix to reduce the computations. Since two dimensional input
was used for neurons of hidden and output layers.[16] can‟t be given to neural network then it is converted to 64×1
Handwritten Numeral recognition plays a vital role in column vector and this column vector is the prototype pattern The
postal automation services especially in countries like India where system will have a function that simulates the Hamming network.
multiple languages and scripts are used. Because of intermixing of The function will act as the network and process the input data to
these languages; it is very difficult to understand the script in which generate the output. The output neuron of the network indicates the
the pin code is written. Objective of this paper is to resolve this result of the recognition process.
problem through Multilayer feed-forward back-propagation In the feedforward layer, p is the input matrix. It will be the matrix
algorithm using two hidden layer. This work has been tested on five of the input image which size is 8 x 8. Therefore, the input p will be
different popular Indian scripts namely Devnagri, English, Urdu, a matrix of 64 x 1. The R number is 64, which is the number of input
Tamil and Telugu. Network was trained to learn its behavior by neuron, while S is the number of the output neuron for this network,
adjusting the connection strengths on every iteration. The resultant which is 9. The weight matrix W1 will be generated using the
of each presented training pattern was calculated to identify the prototype data. It will take the prototype data matrix of the 9 digits,
minima on the error surface for each training pattern. Experiments so the weight matrix will be a matrix of 9 x 64. The bias b will be a
were performed on samples by using two hidden layers and as the matrix of 9 x 1.
number of hidden layers is increased, more accuracy is achieved in In the recurrent layer, there are 9 output neurons which represent
large number of epochs.[17] the number of digits.
The recognition of machine printed and handwritten The salt & pepper noise having different density is added in the
numerals has been the subject of much attention in pattern image by using the MATLAB function & then it is processed &
recognition because of its number of applications such as bank recognized by using the designed system.
check processing, interpretation of ID numbers, vehicle registration
numbers and pin codes for mail sorting. Promising feature V. PERFORMANCE ANALYSIS.
extraction methods have been identified in the literature for
recognition of characters and numerals of many different scripts. The task was to design a system which associates every
i
These include template matching, projection histograms, geometric fundamental pattern with itself. That is, when presented with x as
i
moments, Zernike moments, contour profile, Fourier descriptors, input, the system should produce x at the output. In addition, when
i
and unitary transforms. A brief review of these feature extraction presented with a noisy (corrupted) version of x at the input, the
i
methods is found in [21] system should also produce x at the output.

90

ISSN: 2277 – 9043

Let the Hamming distance between two binary vectors x and
y (of the same dimension) be denoted as d(x, y). The design phase 100
of the Hamming memory involves simply storing all the patterns of
the fundamental memory set. In the recall phase, for a given input 80

%Accuracy
N
memory key x є (0, 1) , the retrieved pattern is obtained as follows
(1) Compute the Hamming distances dk = d(x, xk), k = 1, 2, 60
……., m.
(2) Select the minimum such distance dk = min {d1, d2, 40
…..dm} % Accuracy
k 20
(3) Output the fundamental memory y = x (closest match)
(4) Input: storage patterns for Hamming network. 0
(5) Input prototype images for digits 1-9 from .bmp format.
(6) Example: p = 64*64 matrix of prototype input image of 1 2 3 4 5 6 7 8 9
digit 1.
Input Pattern

Fig. 5. Reconstruction Efficiency for Salt & Pepper Noise
Reconstruction is done by Hopfield network which gives
which gives maximum % accuracy for digit 1. Figure 5
shows the graph of % accuracy of reconstruction and input
pattern
Fig. 4. Prototype Images
Scale data and display as image to use the full colormap. Colormap VI. CONCLUSION.
(gray) sets the current figure‟s colormap to gray. The values are in
the range from 0 to 1. A colormap matrix may have any number of Pattern recognition can be done in normal computers and
rows, but it must have exactly 3 columns. Each row is interpreted as neural networks. Computers use conventional arithmetic algorithms
a color, with the first element specifying the intensity of Red light, to detect whether the given pattern matches an existing one. It will
the second Green light, and the third Blue. Color intensity can be say either yes or no. It does not tolerate noisy patterns. On the other
specified on the interval 0.0 to 1.0. For example, [0 0 0] is black, [1 hand, neural networks can tolerate noise and, if trained properly,
1 1] is white, [1 0 0] is pure Red, [.5 .5 .5] is gray, and [127/255 1 will respond correctly for unknown patterns. Neural networks
212/255] is aquamarine. Resizes a matrix map image to an 8*8 constructed with the proper architecture and trained correctly with
matrix to reduce computations. i.e. Convert and compression of good data give amazing results, not only in pattern recognition but
image. also in other scientific and commercial applications.
Example: p2 = resize (p, [8, 8]). The model hamming is used for image pattern classification
Table 1: Classification Efficiency/ Output digit for salt & paper noise this algorithm supply the prototype images in the model
memory and then use the memory later to identify the stored
Input Noise Density/ Recognised output % patterns; when distorted input is given as input to the model .
Efficiency of both models varies according to the noise. The
Pattern 0.01 0.05 0.1 0.5 1 Accuracy
hamming network could recognize the input numerals added with
1 1 1 1 1 4 80 salt and pepper noise at average of 89% . The developed system can
be used in car plate recognition. In future we can consider alphabets
2 2 2 2 2 3 80
for recognition.
3 3 3 3 3 1 80
REFERENCES
4 4 4 4 4 1 80
[1] Hagan M.T., Demuth H. B., Beale M., “Neural Network Design”,
5 5 5 5 5 4 80 Thomson Learning Vikas Publishing House, 1996.
[2] R C.Gonzalez, R. E. Woods, “Digital Image Processing”, Pearson
6 6 6 6 1 4 60 Education, Inc. and Dorling Kindersley Publishing, Inc.2008.
7 7 7 7 7 9 80 [3] S Jayaraman, S Esakkirajan, T Veerakumar, “Digital Image
Processing”, Tata McGraw Hill Education, 2009
8 8 8 8 8 4 80 [4] Earl, Gose, Richard Johnsonbaugh, Steve Jost, “Pattern recognition
and Image analysis”, Asoke K. Ghosh, Prentice Hall, 1997.
9 9 9 9 9 4 80 [5] “Statistical Pattern Recognition: A Review”, Anil K. Jain,Fellow,
IEEE, Robert P.W.Duin, and Jianchang Mao, Senior Member, IEEE
Transactions On Pattern Analysis And Machine Intelligence, Vol. 22,
Classification efficiency of the system is different for each digit for No. 1, January 2000.
the increased noise density. It gives maximum 80% accuracy for [6] Zurada, J. M., “Introduction to Artificial Neural Systems”, Jaico
numeral 1 to 5 and 7 to 9 publishing House, Mumbai, 2002.
[7] The IEEE website. [Online]. Available: http://www.ieee.org/
“PDCA12-70 data sheet,” Opto Speed SA, Mezzovico, Switzerland.
[8 ] McCulloch, W.S., and Pitts, W. (1943), "A Logical Calculus of the
Ideas Immanent in Nervous Activity," Bulletin of Mathematical
Biophysics, 5, 115-133.
[9] Brion IC. Dolenko and Howard C. Card, “Handwritten Digit Feature
Extraction and Classification Using Neural Networks", „CCECE’,
1993, IEEE 0-7803-1443, PP 88-91.

91

ISSN: 2277 – 9043

[10] Seong-Whan Lee, “off-line recognition of totally unconstrained
handwritten numerals using a simple multilayer cluster neural
network”, IEEE transactions on pattern analysis and machine
intelligence, vol. 18, no. 6, June 1996, 648-652
[11] Rosenblatt, F. (1959), "The Perceptron: A Probabilistic Model for
Information Storage and Organization in the Brain," Psychological
Review 65:386-408.
[12] “Recognition of handprinted numerals by two-stage feature extraction”,
IEEE transactions on systems science and cybernetics, April 1970
[13] “Handwritten alphabet recognition using hamming network”, Arnold
Aribowo, Samuel Lukas, Handy, Seminar National Aplikasi
Teknologi Informasion, 2007 (SNATI 2007) ISSN: 1907-5022
Yogyakarta, 16 June 2007
[14] A Two-Level Hamming Network for High Performance Associative
Memory”, by Nobuhiko Ikeda, Paul Watta, Metin Artiklar and
Mohamad H. Hassoun.
[15] “Recognition of Noisy Numerals using Neural Network”, Mohd
Yusoff Mashor and Sitz Noraini Sulaiman Centre for Electronic
Intelligent System (CELIS), School of Electrical and Electronic
Engineering, University Sains Malaysia Engineering Campus, 14300
Nibong Tebal Pulau Pinang, Malaysia.
[16] Amit Choudhary, Rahul Rishi, Savita Ahlawat, “Performance
Analysis of Feed Forward MLP with various Activation Functions
for Handwritten Numerals Recognition” IEEE, Volume 5, 2010,pp
852-856,
[17] Stuti Asthana, Farha Haneef, Rakesh K Bhujade, “Handwritten
Multiscript Numeral Recognition using Artificial Neural Networks”
IJSCE 2231-2307, Volume-1, Issue-1, March 2011 pp
[18] Leah Bar, Nir Sochen, and Nahum Kiryati “Image eblurring in the
Presence of Salt-and-Pepper Noise”IEEE IMAGE PROCESSING, VOL.
16, NO. 4, APRIL 2007
[19] Y. Le Cun, et al., ”Constrained Neural Network for Unconstrained
Handwritten Digit Recognition,” Proc. of First Int. Workshop on
Frontiers in Handwriting Recognition, Montreal, Canada, 1990, pp.
145-154,.
[20] U. Pal, T. Wakabayashi and F. Kimura, “Handwritten numeral
recognition of six popular scripts,” Ninth International conference on
Document Analysis and Recognition ICDAR 07, Vol.2, pp.749-753,
2007.
[21] Øivind Due Trier, Anil K. Jain and Torfinn Taxt, Feature Extraction
Methods for Character Recognition- A
survey, Pattern Recognition, Volume 29,
Issue 4, April 1996, pp 641-662.

Smita K. Chaudhari is presently Pursuing
ME in Electronics & Communication
Engineering from SSGB College of Engg. &
Technology, North Maharashta University-
Jalgaon, Maharashtra, India. She received the
BE degree from Godavari College of Engg.&
Technology, North Maharashta University,
Jalgaon. Her interested area is Image
processing, Neural Network,VLSI.

G.A.Kulkarni is presently Associate Professor &
Head of Electronics & Communication Engg.
Department SSGB College of Engg. &
Technology, affiliated to North Maharashtra
University- Jalgaon, Maharashtra, India.. He
received the M.E degree from the Dr.B.A.M.
University Aurangabad and presently he is
persuing PhD degree from Dr.B.A.M. University
Aurangabad. His research interests include
communication systems & electromagnetic engg.,
Neural network.

92

88 92

More Related Content

What's hot (20)

Similar to 88 92 (20)

More from Ijarcsee Journal (20)

Recently uploaded (20)

88 92