RESEARCH PAPER

Int. J. of Recent Trends in Engineering and Technology, Vol. 1, No. 1, Nov 2009

An Efficient OCR for Printed Malayalam Text

using Novel Segmentation Algorithm and SVM
Classifiers
Bindu Philip1 and R. D. Sudhaker Samuel2
1, 2
Department of Electronics & Communication, S J College of Engineering, Mysore, India
Email: 1binduthomas25@yahoo.co.in and 2sudhakersamuel@yahoo.co m

Abstract—This paper describes an Optical Character levels of accuracy. (Such systems are also available for
Recognition (OCR) System for printed text documents in many European languages as well as some of the Asian
Malayalam, a South Indian language. Indian scripts are rich languages such as Japanese, Chinese etc.) However, there
in patterns while the combinations of such patterns makes are not many reported efforts at developing OCR systems
the problem even more complex and these complex patterns
are exploited to arrive at the solution. The system segments
for Indian languages.
the scanned document image into text lines, words and An automatic character recognition system is one of
further characters and sub-characters. The segmentation the most fascinating and challenging areas of pattern
algorithm proposed is motivated by the structure of the recognition with a wide range of practical applications
script. A novel set of features, computationally simple to like mail sorting, forms processing, preserving historical
extract are proposed. The approaches used here are based documents in editable format, desktop publication,
on the distinctive structural features of machine-printed text backup files of rare books, reading aid for blind, and
lines in these scripts. A lateral cross-sectional analysis is other applications involve language processing, word
performed along each row of the normalized binary image indexing, library automation. The different challenges
matrix resulting in distinct features. The final recognition is
achieved through classifiers based on the Support Vector
that exist in Malayalam script are its large character set of
Machine (SVM) method. The proposed algorithms have roughly more than 900 characters, similarity of character
been tested on a variety of printed Malayalam characters shapes, and complexity of character structure. This paper
and currently achieve recognition rates between 90.22% addresses all these challenges.
and 95.31 %.
Index Terms—Malayalam Script, OCR, Structural II. SOME EXISTING OCR TECHNIQUES FOR INDIAN
approach, Segmentation, Support Vector Machine (SVM) SCRIPTS
Classifier.
Some of the existing techniques used in OCR for
I. INTRODUCTION Indian scripts work is presented here. Pal & Chaudhuri
[2] reported a complete OCR system for printed
Optical character recognition system has received
Devnagari here headline deletion is used to segment the
considerable attention in recent years due the tremendous characters from the word. An OCR for Telugu is reported
need for digitization of printed documents. In this paper by Negi, et. al[4], where instead of segmenting the words
we describe a document image analysis system that can
into characters as usually done, words are split into
handle printed text documents in Malayalam, the official connected components (glyphs). Some contributions that
language of the south Indian state of Kerala. The input to report the use of SVM classifier are, a font and size
the system is the scanned image of a page of printed
independent OCR system for printed Kannada documents
Malayalam text. The output is an editable computer file using support vector machines reported by Ashwin T V
containing the text data in the printed page. The most and P.S Sastry [1]; Seethalakshmi, et. al, [3], reported a
important phases involved are Segmentation and Feature Tamil OCR using Unicode and SVM classifier. Renju
Extraction. The task of separating lines and words in the John, et.al., [6] reported work on isolated Handwritten
document is fairly independent of the script and hence
Malayalam Character Recognition based on 1 D Wavelet
can be achieved with conventional projection profiles Transform. Recognition of Isolated handwritten character
techniques. However, due to the peculiarities of the images based on k-nearest neighbour classifier is reported
Malayalam script, a novel segmentation scheme is
by Lajish, et.al., [5]. A comprehensive study on the
proposed whereby words are first segmented to a sub- success rate of well known feature extraction methods in
character level and the individual pieces are recognized. terms of recognition accuracy and computational
These are then put together for the recognition of the
complexity is yet to be reported. There is hardly any
characters. The proposed system employs a classifier reporting on techniques used for printed Malayalam
based on the concept of Support Vector Machines OCR.
(SVM).
Currently there are many OCR systems available for
handling printed English documents with reasonable

© 2009 ACEEE 178

DOI: 01.IJRTET. 01.01.287
RESEARCH PAPER
Int. J. of Recent Trends in Engineering and Technology, Vol. 1, No. 1, Nov 2009

III. MALAYALAM SCRIPT The characters and sub characters in printed Malayalam
Malayalam is a Dravidian language with about 35 text have uniform distance of separation and thus
million speakers. It is spoken mainly in the south western segmentation of full characters in a Malayalam word is a
India, particularly in Kerala. The Malayalam script is great challenge. The segmented fragments are now
derived from the Grantha script, a descendant of the normalized to a height of m1 pixels preserving the length
ancient Brahami script. The character set consists of 13 of the characters. It is significant to mention that m1 = 50
vowels, 2 left vowel signs, 7 right vowel signs, some was found to be an optimum value after several trials.
appear on both sides of the Conj/consonant, 30 The problem now reduces to the characterization of mxn
commonly used conjuncts, 36 consonants and vowel image matrices. The value of m however can be fixed at
signs are shown in Figure 1. The dependent vowels do an optimal value to obtain distinct features of the entire
not stand on their own, but are depicted in combination data set economically. The character recognition engine
with a consonant or consonant cluster [7]. The now performs feature extraction by finding a set of
positioning of the dependent vowel may be to the left, to vectors, which effectively represent the information
the right, or both to the left and right of the consonant/ content of a character. A novel set of features,
conjunct, depending on the vowel sign being attached [8] computationally simple to extract are proposed and the
as shown in Figure 1. printed Malayalam characters are classified using
Malayalam has remarkably distinct lateral variations as hierarchical SVM classifiers.
compared to many other Indian languages with a number
of curls and twists in the characters. Another very V. SEGMENTATION

A Malayalam character could consist of several

uniformly spaced unconnected components such as a
vowel (only at the beginning of a word) or a consonant/
conjunct along with vowel signs. Conventional
techniques like horizontal and vertical projection profile
methods fail to segment the complete character correctly
because of the equal space between the characters of a
word and the sub characters of a character within a word.
An appropriate and novel technique is proposed in this
paper to segment Malayalam Characters. We considered
the fact that printed Malayalam characters can have a
maximum of three segments. The first segment could
have either none or possibly one or two left vowel signs
(a unique case). The second segment would be the core
character which could be either a vowel or a consonant or
a conjunct while the third segment could again have
either none or one of the seven right vowel signs as
shown in Figure 1. An example of a character with all
three segments is shown in Figure 2.

Figure 1: Malayalam Vowel Signs

interesting feature of this script is that the number of

columns varies from 53 to a phenomenal 347 columns
over the entire extended character set of the language.
Figure 2: A Typical Malayalam Character
IV. IMPLEMENTATION MODEL

The stages involved in the development of the OCR Let X represent a consonant/ conjunct and Y represent a
engine are image acquisition, preprocessing, vowel. Further let 0 represent vowel signs appearing to
segmentation, normalization, feature extraction and the left of the consonant/ conjunct while 1 represent
classification. A printed document containing vowel signs appearing to the right of the consonant/
Malayalam text is scanned on a flatbed scanner at 300 dpi conjunct. The valid character sequences are of the form
for digitization. This digitized image is preprocessed for Y, X, X1, 0X, 0X1 and 00X, where each of these sub
removal of background noise and the grey scale image is characters Y, 0, X, and 1 are segmented out using the
converted to a binary image after which line classical vertical projection profile method. These
segmentation and word segmentation is performed using segmented sub characters are applied to the logic shown
classical horizontal and vertical projection profiling in the flow chart of Figure 3 and the classification search
technique. Character segmentation is then done using a space as shown in Figure 4 where V represents vowels(13
novel segmentation algorithm as explained in Section V. in all), C represents consonants (36 in all), Conj
© 2009 ACEEE 179
DOI: 01.IJRTET. 01.01.287
RESEARCH PAPER
Int. J. of Recent Trends in Engineering and Technology, Vol. 1, No. 1, Nov 2009

represents conjunct characters(30 in all), VSL represents The process of digitization followed by segmentation
the vowel signs to the left(2 in all) while VSR represents essentially renders the image in the form of an mxn
the vowel signs to the right(7 in all). The end of the matrix. These matrices are then generally normalized
word is identified by a larger valley in the vertical and then converted into a square matrix in order to apply
projection profile. Note that the sub characters are the classical tools of linear algebra for characterization. It
extracted using the smaller valley of projections. All the is easy to see that any form of characterization results in a
sub characters of a word are subjected to this logic in reduction in dimensionality which essentially helps in the
sequence of appearance in the word. The logic used to search process by classification over a large data base.
reduced the search subspace is that the first level search However, there are instances where rich information
subspace has vowels, consonants, left vowel signs and along rows would be lost in the process of reducing the
conjuncts. The first recognized character/ sub character of image matrix to square.
a word falls into one of these four categories only. One good example is the segmented images of the
characters of Malayalam language. It thus makes good
sense to retain the number of columns. A rectangular,
strictly black and white digital image preprocessed to
remove any extraneous noise can be represented by a
matrix A, where
A = ( aij ) ∈ mxn : aij = {0,1} , (1)
usually n > m.
In order to ensure practicality in classification and
identification, for the matrix in equation (1) reduction in
dimensionality is performed to obtain a feature
vector x ∈ , at the same time capturing the distinct
m

information in all the n columns. The methods proposed

in this paper essentially capture useful information along
rows of matrixes and hence we call them lateral analysis.
Selected features along rows are retained rather than
losing them by compression, normalization or by looking
only at the overall characteristic feature of a matrix.
The matrix A as represented in equation (1) has several
lateral features. Three of these are presented here.
Figure 3: Segmentation Logic Flow Chart

The search space further reduces for finding the A. Frequency Capture
This process in principle captures the frequency of
subsequent character or sub character as independent
vowels can appear only in the beginning of a word. transitions along each row. The feature vector, x∈ m
of the matrix A ∈
The logic used behind the choice of search space for the mxn
in this approach is defined by
classifier is based on the sequence of arrangement of the n
segments of the character. This logic facilitates accurate xi = ∑ | ai , j +1 − ai , j | (2)
segmentation. j =1
This captures features of characters with multiple loops
which is a distinct feature of Malayalam characters.

B. Average Gap Analysis

Here, the average gap along each row is computed to
get feature vector x ∈ m of the matrix A ∈ mxn
given by
n
n − ∑ ai , j
j =1
xi = (3)
σi
n +1
σ i = (∑ | bi , j +1 − bi , j |) / 2 with each row padded
Figure 4: Classification Search Space
j =1

VI. FEATURE EXTRACTION with 1s on either ends for the sake for computational
convenience, where
© 2009 ACEEE 180
DOI: 01.IJRTET. 01.01.287
RESEARCH PAPER
Int. J. of Recent Trends in Engineering and Technology, Vol. 1, No. 1, Nov 2009

(bij ) = (aij ) for i =1…m and j = 2…n+1 The optimal solution is obtained when this hyperplane
is located in the middle of the distance between the
(bi ,1 ) = (bi , n + 2 ) = 1∀i convex envelopes of the two classes. This distance is
This captures gaps between the numerous curls in the denoted by dm and is expressed by,
characters, once again a distinct characteristic of 2
Malayalam characters.
dm =
|| w ||
The support vectors are situated on the margins of
the two classes.
C. Absorption If the training vectors membership is defined by
Here, the number of ones along each row is uk = 1 if xk ∈ ω1
computed to get feature vector x ∈ m of the matrix uk = −1 if xk ∈ ω2
Then the support vectors can be written in the form
A ∈ mxn given by
n
{
Ω s = xk | uk ( wT xk + b ) = 1 }
∑a
j =1
i, j
The structure of the SVM classifiers can be modified to
also generate non-linear separating surfaces. The basic
xi = (4) idea is to project the input vectors in higher dimension
n space where the classes become linearly separable. This
transformation is performed by means of a non-linear
This essentially captures the large number of vertical function Φ with modifies the scalar products of the two
strokes, typical of Malayalam characters. input space vectors.
xk → Φ ( xk ) and x j → Φ ( x j )
The features extracted using these methods are grouped
into sub classes and provided to the classification

( )
module.
⇒ xTj xk → Φ ( xk ) Φ x j
T
Let the data base consist of k images. These k images
would correspond to k image matrices as represented by the function Φ is replaced by a symmetric and separable
equation (1). Using any of the three approaches, k feature
function called kernel Δ .
vectors ( x k ) ∈ m are computed. It was seen that for
The Kernel Function is defined as
|||≥ ξ∀k ensuring Δ ( xk , x j ) = exp ( −α || xk − x j ||2 )
k −1
consecutive vectors ||| x || − || x
k

that these features are sufficiently distinct to be useful for

Δ : n X n → , Δ ( xk , x j ) = Φ ( xk ) Φ ( x j )
T
classification by using SVM classifier. The method was
successfully applied to the OCR of the complete
The performance of SVM depends on the kernel. We use
Malayalam character set. Malayalam has exceedingly
RBF (Gaussian) kernel, which out performed the other
rich patterns along rows. The transverse stroke caused by
commonly used kernels in the preliminary experiments.
the curly Malayalam characters attracts it to the proposed
Gaussian RBF kernel is given as
Δ ( xk , x j ) = exp ( −α || xk − x j ||2 )
methods. Further for a given font size the Malayalam
character range from 53 columns to a phenomenal 346.
A. Implementation of SVM Classifier
VII. CLASSIFICATION The classification stage is the decision step and
The Support Vector Machine (SVM) classifier in its basic associates a label with an input pattern. The segmentation
form implements two-class classifications. The objective of the characters into segments the word into vowels,
is to further improve the recognition rate by using support consonants, conjuncts, left vowel signs, or right vowel
vector machine (SVM) at the segment classification level. signs and each segment is assigned with a label. The
The advantage of SVM, is that it takes into account both maximum number of segments in a given Malayalam
experimental data and structural behavior for better character is not more than three segments and getting
generalization capability based on the principle of back the original character is based on the logic shown in
structural risk minimization (SRM)[9]. Its formulation the flow chart. SVM classifiers are used to label each of
approximates SRM principle by maximizing the margin the segments V, 0, X and 1. The SVM (binary classifier)
of class separation, the reason for it to be known also as is applied to this multiclass character recognition problem
large margin classifier. by using one-versus-rest type method. The problem now
The fundamental idea of SVM classifiers is to find a is a n-class problem with n equal to the number of
separating hyperplane between two classes segments in total, but during training it was found that
some of the confusion characters have similar appearance
( h : w x + b = 0 ), so that minimal distance with respect
T
though this problem of similarity is well tackled by using
to the training vectors, called margins, is maximum. the feature extraction approaches explained in the Section
© 2009 ACEEE 181
DOI: 01.IJRTET. 01.01.287
RESEARCH PAPER
Int. J. of Recent Trends in Engineering and Technology, Vol. 1, No. 1, Nov 2009

V. Hierarchal SVM classifiers are used with the viability of the proposed features using the three different
reduction in the search space as shown in Figure 4. structural approaches for classification. Identical
characters were found to be far apart in feature space in
VIII. RESULTS AND ANALYSIS all the three approaches. The methods were tested on
Malayalam was found most appropriate to evaluate the approximately 6000 Malayalam characters and average
three methods of characterization to extract the recognition accuracy of 88.99% for Absorption, 91.88%
extraordinarily distinct and dominant characteristic for average gap and 93.69% for Frequency capture has
features. Feature vectors of all the 620 Malayalam been achieved.
characters including conjuncts extracted based on the The Figure 5 shows some examples of the words used for
methods outlined form the training data set. Around 1000 experimentation having the combination of left vowel
different samples of Malayalam words from different signs, consonants, right vowel signs, along with
textbooks and magazines were scanned to obtain the data consonants which was successfully segmented and
set for training. The data for testing the approach was accurately classified using the approach explained in the
Section V and Section VI for segmentation and feature
TABLE I. extraction respectively.
RECOGNITION RATE OF DIFFERENT TYPES OF MALAYALAM
CHARACTERS BASED ON THE PROPOSED METHODS
ACKNOWLEDGEMENT
Malayalam Absorption Average Frequency This work was supported in part by research grants from
characters Gap Capture
UGC for Major Research Project in Science and
General
93.47 % 94.16 % 95.31 % Technology, F.No. 32-113/2006
characters
Similar
84.23% 90.30 % 92.27 % REFERENCES
characters [1] Ashwin T V, P S Sastry “A font and size independent OCR
Long system for printed Kannada documents using support vector
86.21 % 92.21 % 94.54 %
characters machines”, Saadhanaa, Vol. 27, Part 1, February 2002, pp. 35–58.
Short [2] Pal U. , B. B. Chaudhuri 1997 Printed Devnagari Script OCR
90.22 % 91.34 % 93.64 % System. Vivek, vol.10, pp.12-24
characters [3] Seethalakshmi R., Sreeranjani T.R., Balachandar T., Abnikant
Conjunct Singh, Markandey Singh, Ritwaj Ratan, Sarvesh Kumar 2005
90.62 % 91.39 % 92.69 %
characters “Optical Character Recognition for printed Tamil text using
Unicode” Journal of Zhejiang University SCI 6A(11) pp.1297-
obtained by scanning over 100 different documents. The 1305
[4] Negi Atul, Chakravarthy Bhagvati and.Krishna B 2001 An OCR
approach has been tested successfully and a good system for Telugu. Proc. Of 6th Int. Conf. on Document Analysis
recognition rate is achieved in all the trials. The SVM and Recognition IEEE Comp. Soc. Press, USA,. pp. 1110-1114.
classifier, is used to recognize the characters. A notable [5] Lajish V. L., Suneesh T.K.K. and Narayanan N.K., “Recognition
outcome of our approach is that it is inherently invariant
to font face and proper normalization ensures invariance
to font size. The paper has addressed the problem of
character recognition through three structural approaches
as features. Recognition of Malayalam characters based
on distinct features has been demonstrated with good
accuracy.
Confusion characters usually misclassify in classical
approaches. It is thus appropriate to compare the
proposed approaches on their performances in classifying
Figure5. Examples of Words used for Experimentation
such similar characters. It is fascinating to see that the
frequency capture approach yields better results with of Isolated Handwritten Character Images using Kolmogrov-
good discrimination between similar characters. Smirnov Statistical Classifier and k-nearest Neighbour classifier”,
The performance of the proposed methods in terms of the Proc. Of the International Conference on Cognition and
Recognition ICCR-05, Mandya, Karnataka, December, 2005
recognition rate are compared and is tabulated for five [6] Renju John, G.Raju and D. S. Guru, “1D Wavelet Transform of
different categories of Malayalam characters on the basis Projection Profiles for Isolated Handwritten Malayalam Character
of there column length and structure. It is clear from the Recognition”, Proc. of International Conference on Computational
Table I that the proposed methods perform well and Intelligence and Multimedia Applications 2007, Sivakashi, IEEE
computer society Press, 2007, pp 481-485,.
appear promising for all types of printed Malayalam [7] P. S. Janardhanan. Issues in the development of OCR systems for
characters. Dravidian languages - proceedings of Akshara 94., BPB
IX. CONCLUSION Publications, New Delhi, India 1994.
The system was developed using C++ on Windows XP [8] Malayalam standardization report May 2001
Platform. The proposed approaches have been tested [9] Burges C J C 1998 A tutorial on support vector machines for
pattern recognition. Data Mining Knowledge Discovery 2: 955-
successfully on the extended Malayalam character set. 974.
Several analyses have been performed to illustrate the

© 2009 ACEEE 182

DOI: 01.IJRTET. 01.01.287