Optical character recognition performance analysis of sif and ldf based ocr

Dhinaharan Nagamalai et al. (Eds) : ACITY, WiMoN, CSIA, AIAA, DPPR, NECO, InWeS - 2014
pp. 319–328, 2014. © CS & IT-CSCP 2014 DOI : 10.5121/csit.2014.4532
OPTICAL CHARACTER RECOGNITION
PERFORMANCE ANALYSIS OF SIF AND
LDF BASED OCR
Pradeep Kumar Jena1
, Charulata Palai2
, Lopamudra Sahoo3
and Anshupa Patel4
1
Department of MCA, National Institute of Science and Technology,
Berhampur, Odisha, India
pradeep1_nist@ yahoo.com
2
Department of CSE, National Institute of Science and Technology,
Berhampur, Odisha, India
charulatapalai@ gmail.com
3,4
National Institute of Science and Technology, Berhampur, Odisha, India
06lopamudra@gmail.com
anshupa@gmail.com
ABSTRACT
The Optical Character Recognition (OCR) is becoming popular areas of research under pattern
recognition and smart device applications. It requires the intelligence like human brain to
recognize the various handwritten characters. Artificial Neural Network (ANN) is used to
gather the information required to recognize the characters adaptively. This paper presents a
performance analysis of character recognition by two different methods (1) compressed Lower
Dimension Feature(LDF) matrix with a perceptron network, (2) Scale Invariant Feature (SIF)
matrix with a Back Propagation Neural network (BPN). A GUI based OCR system is developed
using Matlab. The results are shown for the English alphabets and numeric. This is observed
that the perceptron network converges faster, where as the BPN can handle the complex script
recognition when the training set is enriched.
KEYWORDS
Character recognition, perceptron network, back-propagation neural network, scale invariant
feature, low dimensional feature.
1. INTRODUCTION
Automatic character recognition is a well-accepted area of research under pattern recognition. In
handwritten character recognition, characters are written by different individuals that vary
drastically from person to person due to variation in the writing style, its size and orientation of
characters[1]. This makes the system difficult to recognize the characters. Artificial Neural
Network (ANN) helps to solve the problem of identifying handwritten characters in an automated
manner. ANN is an adaptive computational model which is activated by set of pixels of a
specific character as features, processing of the similar and divergence information available in
the features used to recognize the character.

320 Computer Science & Information Technology (CS & IT)
Various methods have been developed for the recognition of handwritten characters, such as the
compressed Column-wise Segmentation of Image Matrix (CSIM) [2] using Neural Network. In
this method, the image matrix is compressed and segmented column-wise then training and
testing using a neural network for different character is performed. The Multi-scale Technique
(MST)[2][3] used for high resolution character sets.
In case of feature based method, Scale Invariant features of characters such as height, width,
centroid, number of bounded regions and textual features such as histogram information are used
for character recognition. The features are feed to a Neural Network [4] for training and testing
purpose. Hand written character recognition using Row-wise Segmentation Technique (RST)
approach used to find out common features along the rows of same characters written in different
hand writing styles and segmenting the matrix into separate rows and finding common rows
among different hand writing styles[5]. Block-wise segmentation technique is also used for the
character recognition [6] by matching the similarity among blocks of the characters.
The complex character such as Hindi, Oriya and Bangla character recognition [7][8] is more
challenging at it produces very alike feature matrices for different characters due to their
structural complexity. Hybrid methods are also applied to recognize the hand written characters.
One such method is a prototype learning/matching method that can be combined with support
vector machines (SVM) in pattern recognition [9].
It is observed that the finding the ideal feature set for particular language and normalizing the
feature matrix is not easy, it requires substantial amount of processor time. In this work we have
done a comparative study on character recognition using feature matrix with a Back Propagation
Neural network (BPN) verses the character recognition using reduced dimensional block
matrix(8x8) with a Perceptron Neural network for English hand written characters i.e. the upper
case, lower case alphabet as well as digits.
2. METHODOLOGY
The sample handwritten characters are collected form ten different persons using black gel pen
i.e. 10 samples of each letter each having different style. These blocks of characters were
digitized using a scanner. Then each character is extracted from the scanned image automatically
and saved with an appropriate name.

Computer Science & Information Technology (CS & IT)
Figure 1:
2.1. Image Pre-processing
The individual character image is
character. These involve various tasks such as (1) Binarization to reproduce the image with 0
(black) or 1(white), (2) Thinning to remove the thickness artefact of the pen used for writing
characters, (3) image dilation to restore the continuity of the image pixels. Figure
inverted binarized data sheet of the set of handwritten characters.
2.2. Approach-1: Low dimensional feature based recognition
The images are resized into uniq
dimension. They are converted to a reduced dimensional image matrix of size 8x8. It preserves
only the highly significant features of the character that are used for the character recognition.
The input vector i,e. a [64x1] matrix is prepared from the 8x8 image. It is
and testing of perceptron neural network.
Computer Science & Information Technology (CS & IT)
Figure 1: The Proposed Model for Analysis
e individual character image is pre-processed to produce a skeletal template of the handwritten
aracters, (3) image dilation to restore the continuity of the image pixels. Figure
inverted binarized data sheet of the set of handwritten characters.
1: Low dimensional feature based recognition
The images are resized into unique standard since the sample character images are different in
hey are converted to a reduced dimensional image matrix of size 8x8. It preserves
The input vector i,e. a [64x1] matrix is prepared from the 8x8 image. It is used for the training
and testing of perceptron neural network.
321
processed to produce a skeletal template of the handwritten
aracters, (3) image dilation to restore the continuity of the image pixels. Figure-2 show the
images are different in
hey are converted to a reduced dimensional image matrix of size 8x8. It preserves
used for the training

Figure 2: Processed image with bounding box
2.3. Rescaling of image matrix
Case-1: When the original image matrix is a multiple of 8.
i) Input image matrix
ii) Dimension of the original image is divided by 8 i.e. 64 x 32 will be 8 x 4
iii) Original matrix is split into ‘n’ uniform blocks of new dimension
iv) A uniform block is assigned to ‘1’ if the number of 1’s is greater than or equal to the number
of 0’s. Otherwise, a uniform block is assigned to ‘0’.
Case-2: When the original image matrix is not a multiple of 8.
i) Input image matrix
ii) Dimension of the original image is first converted to the nearest multiple of 8 by appending
dummy (zeros) rows and columns i.e. 60 x 50 will be 64 x 56
iii) Dimension of the revised image is divided by 8
iv) Revised matrix is split into ‘n’ uniform blocks of new dimension
v) A uniform block is assigned to ‘1’ if the number of 1’s is greater than or equal to the
number of 0’s. Otherwise, a uniform block is assigned to ‘0’.
In this way images of different dimensions are resized into an 8 x 8 binary matrix. The columns
of 8 x 8 matrixes are stored in a single column matrix one after other. Likewise, 10 samples of
each 26 characters are considered and transformed it into column of size 64 each. As a result of a
training set of 64 x 260 matrix is obtained. Accordingly the target set of 5 x 260 is generated.

Computer Science & Information Technology (CS & IT) 323
2.4. Perceptron based ANN training
An Artificial Neural Network (ANN) is an adaptive computational system, it follows perceptron
learning technique. The input layer consists of 64 neurons that represent one character as input,
the hidden layer consists of 32 neuron, where as the output consists of 7 neurons that represents
pattern of 0 and 1 which maps to an individual character. Each neuron is connected to other
neurons by a link associated with weights. The weight contains information about the input,
which is updated during each epoch.
2.5. Approach-2: Scale Invariant feature based recognition
Aspect Ratio: Height and width of character is obtained. The ratio of height and width remain
approximately same for same person for the different characters.
‫ܴܣ‬ =
‫ܮ‬
ܹ
Where, AR=Aspect Ratio
L=Length of Character
W=Width of Character
Occupancy Ratio: This feature is the ratio of number of pixels which belong to the character to
the total pixels in the character image. This feature provides information about character density.
Number of Horizontal Lines: It’s the number of horizontal lines in a character. It’s found out
using a 3x3 horizontal template matrix.
Number of Vertical Lines: It’s the number of vertical lines in a character. It’s found out using a
3x3 vertical template matrix.
Number of Diagonal Lines: It’s the number of diagonal-1 lines in a character. It’s found out
using a 3x3 diagonal-1 and diagonal-2 template matrix.
Number of Bounded Regions: It’s the number of bounded areas found within a character image.
Number of End Points: End points are defined as those pixels, which have only one neighbor in
its eight way neighborhood. Figure-3 the two end points of the image.
Figure 3: End point in an image
Vertical Center of Gravity: Vertical centre of gravity shows the vertical location of the
character image. Vertical centre of gravity of image is calculated as follow
ܿ‫݃݋‬ሺ‫ݒ‬ሻ =
∑ ‫.ݕ‬ ܰ‫ݕ‬
∑ ܰ‫ݕ‬
Where, Ny: the number of black pixels in each horizontal line Ly with vertical coordinate y,
Horizontal Center of Gravity: Horizontal centre of gravity shows the horizontal location of the
character image. Horizontal centre of gravity of image calculated as follow:

ܿ‫݃݋‬ሺℎሻ =
∑ ‫.ݔ‬ ܰ‫ݔ‬
∑ ܰ‫ݔ‬
Where, Nx: the number of black pixels in each vertical line Lx with horizontal coordinate x
2.6. BPN based ANN Training
Back-propagation neural network is mostly used for the handwritten character recognition since it
support the real values as input to the network. It is a multi-layer feed-forward network which
consists of an input layer, hidden layer and output layer. The normalized values of the scale-
invariant feature matrix are given as the input for the training, the output is pattern of 0’s and 1’s
which maps to an individual character.
2.7. Testing and Recognition
Separate test sets are prepared for testing purpose. After the training process is completed, the
test pattern is fed to the neural network to check the learning ability of the trained net. The output
of the simulation is compared with the specified target set. The character is recognized by
selective thresholding technique.
3. RESULT ANALYSIS
The OCR system is developed using MATLAB. The experimental results are shown below which
is carried out to recognize the “A” and “5” as the inputs. The Figure-4 represents the normalized
values of the scale invariant features of the English Alpha-numerics. Figure-5 indicates the
performance analysis of BPN network as trained with normalized feature values. The Figure-6
shows the interface to load the test image and to select the training type such as BPN based
training. Figure-7 indicates the recognition of “A”, The Figure-8 shows the LDF matrix, Figure-9
represents the perceptron based training and testing. The recognition of “5” is shown in Figure-
10.
Figure 4: Scale Invariant normalized feature matrix for ‘A’-‘z’ & ‘0’-‘9’

Figure 5: BPN based training scale invariant features
Figure 6: User Interface for Load Image

Figure 7: Recognition of character ‘A’ using BPN
Figure 8: Reduced dimensional feature matrix[64x1] for ‘A’-‘z’ & ‘0’-‘9’

The target matrix of
‘A’ is
[ 0
0
0
0
1 ]
The output matrix of ‘A’ is
Figure 9: Perception based training & testing
Figure 10: Recognition of character ‘5’ using perceptron network
4. CONCLUSION
In this work it has been observed that finding the reduced dimensional feature matrix of an image
is easy in comparison with the scale invariant feature matrix. The training performance of the SIF
matrix is much faster and reliable. The performance of the network depends upon selection of the
features into the SIF matrix. The feature selection is more challenging in case of structurally
complex scripts such as Bangla, Hindi and Oriya. It has been observed that using the SIF features,
the English alphabets and numeric’s matches with an accuracy of 95% on average matching while
the LDF match accuracy varies from 78% to 96% for different characters.
The work may be further extended to test the regional language scripts. It can be tested with
different reduced dimensional matrix of different dimensions.

REFERENCES
[1] Vijay Patil and Sanjay Shimpi, “Handwritten English Character Recognition Using Neural Network”,
Elixir Comp. Sci. & Engg. 41 (2011) 5587-5591, November 2011.
[2] Velappa Ganapathy, and Kok Leong Liew, “Handwritten Character Recognition Using Multi-scale
Neural Network Training Technique”, World Academy of Science, Engineering and Technology 39
2008.
[3] Rakesh Kumar Mandal, N R Manna, “Hand Written English Character Recognition Using Column-
Wise Segmentation of Image Matrix (CSIM)”, Issue 5, Volume 11, May 2012.
[4] Rakesh Kumar Mandal, N R Manna, “Hand Written English Character Recognition Using Row-Wise
Segmentation (RST)”, International Symposium on Devices MEMS, Intelligent Systems &
Communication (ISDMISC) 2011.
[5] Apash Roy, N. R. Manna, “Handwritten Character Recognition Using Block wise Segmentation
Technique (BST) in Neural Network”.
[6] Soumya Mishra, Debashish Nanda, SanghamitraMohanty, “Oriya character recognition using Neural
Networks”, Special Issue of IJCCT Vol. 2 Issue 2, 3, 4; 2010 for International Conference [ICCT-
2010], 3rd-5th December 2010.
[7] B. Kumar, N. Kumar, C. Palai,.P. K. Jena, S. Chattopadhyay, “Optical Character Recognition using
Ant Miner Algorithm: A Case Study on Oriya Character Recognition”, International Journal of
Computer Applications (0975 – 8887) Volume 61– No.3, January 2013.
[8] Fu Chang, Chin-Chin Linand Chun-Jen Chen,Institute of Information Science, Academia Sinica ,
Taipei, Taiwan Dept. of Electrical Engineering, National Taipei University of Technology, Taipei,
Taiwan”Applying A Hybrid Method To Handwritten Character Recognition”.
AUTHORS
Pradeep Kumar Jena Working as an Associate Professor in the Dept. of MCA at National
Institute of Science and Technology, Berhampur. His domain of research includes Biometric
AuthenticationImage Processing, Pattern Recognition and Data Mining.
Charulata Palai Working as an Assistant Professor in the Dept. of CSE at National Institute
of Science and Technology, Berhampur. Her domain of research includes Soft
ComputingImage Processing, Pattern Recognition.
Lopamudra Sahoo is a BTech Final Year student of NIST, Berhampur.
Anshupa Patel is a BTech Final Year student of NIST, Berhampur.

Optical character recognition performance analysis of sif and ldf based ocr

More Related Content

What's hot (20)

Viewers also liked (7)

Similar to Optical character recognition performance analysis of sif and ldf based ocr (20)

Recently uploaded (20)

Optical character recognition performance analysis of sif and ldf based ocr