Recognition Of Handwritten Meitei Mayek Script Based On Texture Feature

International Journal on Natural Language Computing (IJNLC) Vol.7, No.5, October 2018
DOI: 10.5121/ijnlc.2018.7510 99
Recognition Of Handwritten Meitei Mayek Script
Based On Texture Feature
Sanasam Inunganbi and Prakash Choudhary
Department of Computer Science Engineering, NIT, Manipur
Abstract
Recognition of Manipuri Script called Meitei Mayek is still in the infant stage due to its complex structure.
In this paper, an attempt has been made to develop an offline Meitei Mayek handwritten character
recognition model by exploiting the texture feature, Local Binary Pattern (LBP). The system has been
developed and evaluated on a large dataset consisting of 3,780 characters which are collected from
different people of varying age group. The highest recognition rate achieved by the proposed method is
93.33% using Support Vector Machine (SVM). So, the contribution of this paper is bi-fold: firstly, a
collection of a large handwritten corpus of Meitei Mayek Script and secondly developing
character recognition model on the collected dataset.
Keywords
Handwritten character recognition, Meitei Mayek script, Local Binary pattern, Support Vector Machine.
1. Introduction
Optical Character Recognition is one of the interesting and challenging topics in Pattern
Recognition and Computer Vision. It may be defined as the procedure follow to recognized
printed or written text, which can be acquired either off-line or online. Recognition of printed
script is relatively easy because of its standard size and style. Recognition of handwritten script
is the real challenge faced by research community because of massive variation in the writing
style of different individuals. The general procedure followed by an OCR algorithm, irrespective
of nature of script are image acquisition, pre-processing, feature extraction and recognition.
Handwritten character recognition has multiple applications such as digitization of documents,
automation in mailing system and bank data processing, etc. The result of OCR systems can act
as an intermediate step in translation or speech synthesis. Meitei Mayek is already encoded into
Unicode [1]. Therefore, the OCR output of text as Unicode would help to maintain a universal
standard for further processing.
Meitei Mayek is the official language of Manipur, a state in north-eastern India. It belongs to
Tibeto-Burman branch of the Sino-Tibetan language family and is also spoken in Bangladesh
and Myanmar. The script has been reinstated recently, and that is why only a few researchers
have been accomplished. The current script is the construction of the ancient Meitei Mayek
script. The script consists of 27 alphabets (18 original alphabets called Eeyek Eepee and 9
additional letters called Lom Eeyek, derived from the original alphabets), 8 letters with short

100
ending (lonsum Eeyek), 8 vowel signs (Heitap Eeyek), 10 digits (Cheising Eeyek), 4 punctuation
marks (Khudam Eeyek). Interestingly the alphabets of Meitei Mayek Scripts are named after parts
of human body. The 27 alphabets of the Meitei Mayek Scripts are given in Figure 1 with the
meaning of their names.
Figure 1: Meitei Mayek Alphabets
Various Character Recognition System has been developed for different languages in the world.
However, very few research papers existed in literature for recognition based on Meitei Mayek
Script. In 2010, identification of the handwritten character of Meitei Mayek script using Artificial
Neural Network was proposed by [2]. The authors have used 594 samples images, out of which
459 have been trained, and the remaining 135 had been used for testing. Probabilistic and fuzzy
feature and their combination had been used as the feature vector in their work. In 2013, an OCR
system of Meitei Mayek Script on printed documents was proposed using Support Vector
Machine (SVM) in [3]. They had scanned 100 pages at 300 dpi from textbook, magazine, and
newspaper and stored in ``tiff" or ``png" format. Chain code and directional distribution had been
used as local features, aspect ratio and longest vertical run had been used as the global feature for
classification. In 2014, another OCR system for Meitei Mayek Script using Neural Network (NN)
with back propagation had been proposed in [4]. The authors had used 1000 samples of images,
500 images were used for training while the remaining 500 were used for testing. They had used
the binary pattern as the feature for their model. Recognition of Meitei Mayek numerals or digits
had also been carried out using SVM in [5-6] and NN in [7]. Various work performed on Meitei
Mayek script has been summarized in Table 1.

101
Table 1: Various work performed on Meitei Mayek Script
The remaining part of the paper is organized as Section 2 presents the methodology of the
proposed scheme which is subdivided into Image Acquisition, Pre-processing, Feature Extraction,
and Recognition. Section 3 gives the Experimental Results and Section 4 concludes the paper by
providing Conclusions and Future Work.
2. Proposed Methodology For Character Recognition
Four essential stages have been covered to accomplish the recognition problem of Meitei Mayek
Script. The steps that have been considered in recognizing handwritten characters of this script
are Image Acquisition, Pre-processing, Feature Extraction, and Recognition. This paper focuses
on the recognition of the 27 alphabets. A schematic diagram of this recognition system has been
illustrated in Figure 2.

102
Figure 2: Schematic Diagram of the Proposed Method
2.1. Image Acquisition
The first and foremost important step in developing an OCR system is to design and acquire
databases for evaluation and validation of the proposed algorithms or techniques. So, for
development and evaluation of an efficient character recognition system of Meitei Mayek, we
have collected 4 sets of the 27 alphabets in 108 pages. Each alphabet has 35 samples collected
per page so that one set comprises of 945 (27 x 35) isolated characters. So, in total there are
3,780 samples of the alphabets written by 70 different people of various age group. The image
acquisition is associated with demographic information of the writer such as name, address,
occupation, qualification signature, etc. so that this information can be made available for other
application like signature verification as well. The collected image has been appropriately
scanned using a scanner at 300 dpi and stored in ``jpg" file format for further processing.
2.2. Pre-Processing
Pre-processing is vital for further processing and crucial for maximizing recognition accuracy. It
is the procedure performed to enhance image properties and suppress undesirable noise. The use
of pre-processing technique improved document image and prepared it for next stage in Character
recognition system. So, it is essential to have an active pre-processing step to achieve higher
recognition rate. Besides, it makes an OCR system more robust through accurate enhancement,
noise removal, image thresholding, character segmentation, character normalization, additional
space removal around the character and morphological technique. Segmentation is the main pre-
processing perform in our methodology
Figure 3: Pre-processing perform in our character recognition System
The characters have been collected in tabular form where each character occupies one slot, and
every character needs to be isolated for further processing. Firstly, the character table (the section
where writer replicates characters in their writing style) has been separated from the rest of the

103
part by manual cropping as illustrated in Figure 3. Then, every character has been isolated by
diving the table into an appropriate number of rows and columns since the table has been
appropriately aligned, and further processing has been performed to eliminate extra space other
than the alphabets.
Figure 4: Process for obtaining minimum bounding rectangle of a character
From the isolated image, the edge of the alphabet is determined using Sobel operator. To thicken
the border, we have operated the morphological dilation [8] using structural element disk of size
three. Image filling operation is performed on the binary image to fill any hole exists in the
picture. Then, the connected component analysis is implemented on the filled image to obtain
properties like area and bounding box. The bounding box corresponding to the most significant
object (here the alphabet) is taken into account, and the image is cropped accordingly. The whole
procedure can be depicted diagrammatically from Figure 4.
2.3. Feature Extraction
An image can be represented by distinct numerical properties of various feature, which can be
analyzed by a classifier for correct classification. Texture represents an essential property of an
object that does not change with the change of illumination conditions. Texture can outline
regularity, randomness, smoothness or coarseness. Texture gives the spatial arrangement or
structure of color or pixel intensities over a local area or of the whole image. It defines how a
pixel intensity correlates with their neighboring pixels.
The Local Binary Pattern (LBP), introduced by Ojala [9], has been widely used to analyse the
texture of an image. It may be defined as an ordered set of binary numbers formed by
thresholding between central and its neighbour pixels. The order set of binary number thus
generated is then multiplied by the weighted power of two to be converted into decimal. On the
other hand, histogram provides the global appearance of an image. So, the histogram of LBP
image is considered as the feature vector of an image in our recognition system. Mathematically,
LBP can be calculated using equation 1 where gc is the centre pixel of 3 x 3 block and gp are the
neighbouring pixels. For our experiment, we have taken P = 8 and R = 1. After obtaining the LBP
image LBP_I(x,y), the LBP histogram can be obtained using equation 2.

104
where n is the number of different labels produced by the LBP operator, and I{A} is 1 if A is true
and 0 otherwise. In this modality, we have obtained the value of n= 256.
LBP has many advantages. First, its computation is simple with a low dimensionality of feature
space. Second, it provides highly discriminant features. It has proven to be highly tolerant to
illumination changes. It is required to extract features for a large number of images for
developing this system and hence simple computation yet efficient that takes less time is
preferred.
The collected dataset of 3,780 characters are processed and 256 LBP features are extracted from
each character which will be used for Recognition. The feature vector space is independent of the
image size.
2.4. Recognition
Recognition is an important step to evaluate the usefulness of database of the proposed method. It
requires some decision to be taken to give labels to the test images of which categories they
belong. This decision is made by a classifier based on the features being selected. Support Vector
Machine (SVM) is an efficient and discriminant classifier. It was first introduced by Boser et al.
[10]. The SVM is a supervised learning algorithm for classification and regression. The SVM has
been considered for our experiment with linear kernel because it is robust to noise and it suffers
less from over fitting. The feature size in our experiment is not very small, and hence there is no
need to transform feature points to a higher dimensional space. That means nonlinear mapping
does not improve the performance. Using linear kernel is good enough. Moreover, a thumb rule
of SVM asserts to use linear kernel when the number of features is larger than the number of
observations.
A guide to the proper understanding of SVM was given by A. Ben-Hur et al. in [11]. Training and
testing phases are present in SVM like other classifiers. During the training phase, SVM
constructs a large margin hyperplane by setting Lagrange multiplier (α) and biased (b).
3. Experimental Results

105
For the technical validation and contribution of the dataset as a standard benchmark platform for
linguistic research on Meitei Mayek Script, we have applied character recognition technique and
presented the results of the experiments. The experiment has been carried out on the collected
samples of characters. The database of 27 alphabets has been collected in sets of four. One set
consists of 945 samples, having 35 samples for each alphabet. Out of 945 isolated characters, 540
(20 from each alphabet) has been used for training and 405 (15 each alphabet) for testing. The
experiment has been carried out separately on each of the four sets using SVM classifier. Each set
has been written by a group of different individuals and does have variation. In the collected
dataset, no two individuals have written the same alphabet twice, and the overall variation found
in the database is around 80% and above. So, the proposed method is robust to various writing
styles. The variance in writing has been depicted in Figure 5. The recognition rate achieved in
each of the sets are 90.62%, 92.59%, 93.33% and 91.36% respectively. The highest accuracy
achieved is 93.33% by the Set 3. The accuracies of each alphabet in Set 3 have been highlighted
in Table 2. It can be examined from the table that majority of the alphabets attain 100% of the
recognition rate and only two classes obtained lesser accuracy of 53.33% and 20%.
Comparison of some of the existing work in literature for recognition of Meitei Mayek alphabet
has been performed with the experimental results of our proposed work, and the results have been
summarized in Table 3. It can be observed from the table that our proposed model for recognition
has achieved higher recognition rate as compared to the one existing in the literature. However,
the accuracy in [3] is higher as they had considered printed documents for their experiment and
ours had been developed and evaluated on the handwritten script.
Table 2: Recognition Rate (in %) for each alphabet in Set 3

106
Table 3: Comparison of Recognition Rate (RR) with the existing methods in literature.
Figure 5: Sample of variation found among the same alphabets over the four sets (a,b,c,d) and (e,f,g,h)
4. Conclusions And Future Work

107
In this paper, we have proposed a character technique on a self-collected database of size 3,780
sample of Meitei Mayek alphabets. Till date, this is one of the largest available datasets for Meitei
Mayek Script. The database has been collected in four sets, and four set of experiments have been
conducted achieving varying accuracy. The variance in accuracy is because different individuals
have written the database. Some individual has written with a pen while some, with a pencil.
Moreover, each has their unique characteristic of writing. Hence, we have achieved varying
accuracy. This issue can be resolved, and efficiency can be improved further in revision with new
character recognition technique.
This paper only deals with the recognition of isolated Meitei Mayek alphabets. In future,
identification of dataset having complete sentences of Meitei Mayek Script can be performed. In
future, the focus can be made on the full-length dataset of both printed and handwritten, encoding
the corpus, benchmarking of the dataset on different standards.
References
[1] P. Daniels, “The unicode consortium: The unicode,” Language: journal of the Linguistic Society, vol.
69(1), p. 225–225, 1993.
[2] T. Thokchom, P. Bansal, R. Vig and S. Bawa, “Recognition of handwritten character of Manipuri
Script,” JCP, vol. 5(10), pp. 1570-1574, 2010.
[3] S. Ghosh, U. Barman, P. Bora, T. . H. Singh and B. Chaudhuri, “An OCR system for the Meetei
Mayek script,” in Fourth National Conference on Computer Vision, Pattern Recognition, Image
Processing and Graphics (NCVPRIPG), 2013.
[4] R. Laishram, P. . B. Singh, S. S. D. Thokchom , S. Anilkumar and A. U. Singh, “A neural network
based handwritten Meitei Mayek alphabet optical character recognition system,” in IEEE
International Conference on Computational Intelligence and Computing Research (ICCIC), 2014.
[5] C. J. Kumar and S. K. Kalita, “Recognition of handwritten numerals of manipuri script,” International
Journal of Computer Applications, vol. 84(17), 2013.
[6] K. A. Maring and R. Dhir, “Recognition of cheising iyek/eeyek-manipuri digits using support vector
machines,” Ijcsit, vol. 1(2), 2014.
[7] R. Laishram, A. U. Singh, N. C. Singh, A. Singh and H. James, “Simulation and Modeling of
Handwritten Meitei Mayek Digits using Neural Network Approach,” in Proc. of the Intl. Conf. on
Advances in Electronics, Electrical and Computer Science Engineering-EEC, 2012.
[8] B. R. Masters, R. C. Gonzalez and R. Woods, “Digital image processing,” Journal of biomedical
optics, vol. 14(2), no. 029901, p. 029901, 2009.
[9] T. Ojala, M. P. ainen and D. Harwood, “A comparative study of texture measures with classification
based on featured distributions,” Pattern, vol. 29 (1), pp. 51-59, 1996.

108
[10] B. E. Boser, I. M. Guyon and V. N. Vapnik, “A training algorithm for optimal margin classifiers,” in
Proceedings of the fifth annual workshop on Computational learning theory, 1992.
[11] A. Ben-Hur and J. Weston, “A user’s guide to support vector machines,” Data mining techniques for
the life sciences, pp. 223--239, 2010.
Authors
Sanasam Inunganbi, is currently pursuing PhD in NIT Manipur. She completed her
M.Tech from PDPM IIITDM Jabalpur in Computer Science and Engineering. Her
area of interest includes Pattern Recognition, Image processing.
PrakashChoudhary, obtained his Ph.D. in Computer Science and Engineering from
MNIT Jaipur. He received his M.Tech. in Computer Science and Engineering from
VNIT Nagpur and B.E. from Rajasthan University in Information Technology. He
is currently working as Assistant Professor in NIT Manipur since 2013. His area of
research interests are Biomedical Image analysis, Pattern Recognition, Annotation
and retrieval of Images.

Recognition Of Handwritten Meitei Mayek Script Based On Texture Feature

More Related Content

What's hot (19)

Similar to Recognition Of Handwritten Meitei Mayek Script Based On Texture Feature (20)

Recently uploaded (20)

Recognition Of Handwritten Meitei Mayek Script Based On Texture Feature