Offline Handwritten Character Recognition Using MLPNN and PSO Algorithm

Volume 3, Issue 2, February – 2018 International Journal of Innovative Science and Research Technology
ISSN No:-2456 –2165
Offline Handwritten Character Recognition using

MLPNN and PSO Algorithm
Aarthna Maheshwari Khushboo Shah
Information Technology Information Technology
Mahakal Institute Of Technology Mahakal Institute Of Technology
Ujjain, India Ujjain, India
[email protected] [email protected]
Abstract—Classical techniques are not flexible and adaptive

for new handwriting constraints. Offline handwritten character II. RECOGNITION SYSTEM
recognition for English alphabets has been proposed in this
paper. The proposed system uses three layer feed forward neural A. Image Acquisition
network and optimized the weight using PSO algorithm. The The recognition system takes a scanned image of
proposed system works well with benchmark dataset from handwritten text as an input image. We have used benchmark
C.E.D.A.R. data set i.e. from CEDAR(Center for Excellence in Data
Analysis and Research) in Buffalo University, USA [2]. We
I. INTRODUCTION have used 30 samples of each characters in .png format.
Handwriting has been a method to convey and solidify B. Pre-processing
information from ages. Anything before handwritten history is
classified as pre-history. Handwritten text was codified into Pre-processing is a sequence of morphological operations
symbols which were combines using rules of linguistics to employed to enhance the features of an image. We have
form higher level entities. converted an RGB image to grey scale and employed grey
scale to binary format using binarization with global
Character recognition has been an active area of research in thresholding method [3]. We have used a global threshold of
the field of pattern recognition and machine intelligence. 0.5.
Accuracy of character recognition depend factors like
C. Normalization
A. Method of Data Acquisition:Offline and Online.
It is the process of equating binary array so as to extract the
B. Type of Text: Handwritten or Machine Printed. featured. We have cropped the image and reduces the size to fit
it into 80X50 array.[4] [5]
Offline data acquisition here refers to data acquired using
Optical scanning of handwritten or machine printed text. On- D. Thinning
line refers to data acquired by writing using a digital pen on a
digitizer screen. In online-data acquisition method, the data It thins objects to lines. It removes pixels so that an object
storage requirement is less as temporal data is stored i.e. the without holes shrinks to a minimally connected stroke, and an
consecutive co-ordinates of the trajectory are stored as a object with holes shrinks to a connected ring halfway between
function of time whereas in offline the complete scanned each hole and the outer boundary. This option preserves the
images is stored. Euler number.[6]
The reported accuracy of on-line character recognition E. Feature Extraction
techniques are higher than offline [1]. Mean method is used to extract the features of the character. It
takes the sum of all the 1s pixel and divides them with the
number of pixels in each box.
III. CLASSIFICATION AND RECOGNITION

It is the most important part of the recognition process. Its
accuracy depends on the quality of features. We have used
artificial neural network i.e. a feed forward neural network for
Figure 1 Image of Handwritten character from standard data recognition process. It is a supervised neural network with
set weights optimized using PSO algorithm.
Neural network provides robustness, fault tolerance and
parallel processing ability. It is a three layer feed forward
neural network architecture used to recognize handwritten
English characters.
IJISRT18FB79 www.ijisrt.com 124

ISSN No:-2456 –2165
nClass = 52; % number of classes wb = pso(fcn,nvars,[-1.*ones(1,nvars)],[ones(1,nvars)]);

code = 1:nClass; % code of each class
result = zeros(nClass*(nClass-1)/2,size(features,2)); % % Apply New Weights and Biases to Network
result of each NN for each case
net = setwb(net,wb');
k = 1;
end
hbar = waitbar(0, 'please wait...');
function mse = objective(x,net,inputs,targets)
for i = 1:nClass-1 % from each class
% Apply New Weights and Biases to Network
for j = i+1:nClass % to each class
net = setwb(net,x');
tmp = sim(nets{k},features); % test feature on NN
% Evaluate
idx = tmp(1,:) > tmp(2,:);
y = net(inputs);
result(k,idx) = code(i); % if true , feature belongs to ith
% Calculate MSE
class
result(k,~idx) = code(j); % else , feature belongs to jth mse = perform(net,targets,y);
class end
waitbar(k/size(nets,1),hbar); Here global constants C1 and C2 are used as prescribed by
k = k+1; PSO method. A population of 200 particles is created and
maximum 200 iterations can be applied to find the best
end candidate solution. These values of N and maxiteration are
purely based on hit and trial as in genetic algorithm. Mean
end Square Error method is used to reduce the error in the neural
close(hbar); network. Activation function used is sigmoidal. Gradient
descent with adaptive learning is used as learning algorithm.
result = mode(result,1); % maximum frequent class
The algorithm employed to recognize a character.
A. PSO Algorithm
Particle swarm optimization algorithm [7] [8] has been used
to optimize the weights employed to train the neural network.
It is a meta-heuristic algorithm but multiple training sessions
result into optimal candidates.
function [net,wb] = psonn(inputs,targets)
% net = gann(inputs,targets) returns a feed forward neural
% network, with optimized weights and bias values for
given input and
% target with Genetic algorithm.
Figure 2 Vertical Crop of Character C

% Creat a Feedforward Neural Network
net = feedforwardnet([6 6 6]);
% Configure for input and output
net = configure(net,inputs,targets);
% Calculate Number of Variables to Optimized
wb = getwb(net);
nvars = size(wb,1);
% Apply PSO
fcn = @(x) objective(x,net,inputs,targets);

Figure 3 Recognized character Result

ISSN No:-2456 –2165
IV. RECOGNITION RESULT AND CONCLUSION

The neural network with 52 classes output node of 6 training
algorithm gradient descent and mean square error function is
employed to achieve an accuracy of 83.8462%. The network
can be trained repeatedly to achieve even higher accuracy of
character recognition. The result is comparable with other
state of art methods of offline handwritten character
recognition.
V. ACKNOWLEDGEMENT
We are thankful to C.E.D.A.R., USA for their help as and
when required for the research.
REFERENCES
[1] S.N. Srihari "High Performance Reading Machines" Proc.

IEEEvol. 80 no. 7 pp. 1 120-1 132 1992.
[2] http://www.cedar.buffalo.edu/ilt/research.html
[3] P.K. Sahoo S. Soltani A.K.C. Wong Y.C. Chen "A Survey
of Thresholding Techniques" Computer Vision Graphics and
Image Processing vol. 41 pp. 233-260 1988.
[4] C. Downtown and C.G. Leedham, “Preprocessing and
presorting of envelop images for automatic sorting using
OCR”, Pattern Recognit., vol.23, no. 3-4, pp. 347-362, 1990.
[5] W. Guerfaii and R. Plamondon, “Normalizing and
restoring on-line handwriting,” Pattern Recognit., vol. 26, no.
3, pp. 418–431, 1993.
[6] J. M. Reinhardt and W. E. Higgins, “Comparison between
the morphological skeleton and morphological shape
decomposition,” IEEE Trans. Pattern Anal. Machine Intell.,
vol. 18, pp. 951–957, Sept. 1996.
[7] Kennedy J. (2011) Particle Swarm Optimization. In:
Sammut C., Webb G.I. (eds) Encyclopedia of Machine
Learning. Springer, Boston, MA
[8] http://www.particleswarm.info/

Offline Handwritten Character Recognition Using MLPNN and PSO Algorithm

Uploaded by

Copyright:

Available Formats

Offline Handwritten Character Recognition Using MLPNN and PSO Algorithm

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Offline Handwritten Character Recognition Using MLPNN and PSO Algorithm

Uploaded by

Copyright:

Available Formats

Volume 3, Issue 2, February – 2018 International Journal of Innovative Science and Research Technology

ISSN No:-2456 –2165

Offline Handwritten Character Recognition using

Abstract—Classical techniques are not flexible and adaptive

III. CLASSIFICATION AND RECOGNITION

IJISRT18FB79 www.ijisrt.com 124

nClass = 52; % number of classes wb = pso(fcn,nvars,[-1.*ones(1,nvars)],[ones(1,nvars)]);

Figure 2 Vertical Crop of Character C

fcn = @(x) objective(x,net,inputs,targets);

IJISRT18FB79 www.ijisrt.com 125

IV. RECOGNITION RESULT AND CONCLUSION

[1] S.N. Srihari "High Performance Reading Machines" Proc.

IJISRT18FB79 www.ijisrt.com 126

You might also like