Tamil Characters Recognition and Retrieval
Abdol Hamid Pilevar
Language Engineering Lab, Computer Engineering Department, Bu Ali Sina University, Hamedan, Iran
Keywords: Feature Analysis, Text Retrieval, Character Recognition, Natural Language Processing.
Abstract: In this paper the shape of the vertical projection curves are considered. The behavior of the edges of vertical
projection curve is selected for creating the feature vectors of the characters. The edges of the vertical
projection curve traced and the direction of the movement in the edges has been mapped by Eleven
Direction Method (EDM) method .The direction codes have been extracted and saved as features vectors of
the characters. The method is tested on the Tamil printed text documents. The testing data are collected
from various legal documents. The test documents contain alphabet, special characters. A technique named
EDM is used to search and retrieve the characters from Tamil text databases. The effectiveness and
performance of the proposed algorithm have been tested with 10 separate sample data of 6 different fonts.
The experiments shows that more than 97% of the Tamil characters are recognized correctly therefore, the
proposed algorithm and the selected features perform satisfactorily.
1 INTRODUCTION
Incorrect segmentation of merged characters is one
of the main causes for errors in the recognition. As
segmentation errors induce recognition errors, the
performance of segmentation is crucial for the whole
OCR process. For segmenting the merged characters
in a work which is called cut-off point or an
inflexion, the point with the smallest interior angle is
detected and the whole stroke is split into two
adjacent curves by this point (Lu et al., 2008).
Character recognition based on pixel distribution
probability of character image (Wang, 2008). A
method for segmentation of touching italic
characters (Li et al., 2004). A segmentation
technique for touching Thai type written (Watch,
2004). A segmentation of machine printed
Gurmukhi text (Davessar, 2003). A segmentation
technique far recognizing the touching Thai type
written (Electronic et al., 2004). A lossy/lossless
compression method for printed typeset bi-level text
images is proposed for archiving purposes (Grailu et
al., 2009). A recognition method of line-touching
characters without line removal (Hotta et al., 2008).
Links between landscape aesthetic theory and visual
indicators (ode et al., 2008). A technique for
identification and segmentation of Bengali printed
characters (Sattar et al., 2007). Vertical and
horizontal text lines are detected without prior
assumption. The touching characters belonging to
different lines are detected (Faure et al., 2009). A
method for segmentation of touching italic offered
(Li et al., 2004). A two pass algorithm for the
segmentation and decomposition of Devanagari
composite characters-symbols into their constituent
symbols (Bansal et al., 2002). A recursive
segmentation algorithm for segmenting touching
characters (Liang et al., 1994).
Furthermore, since segmentation errors often
raise chain effects, the performance of segmentation
is crucial for the whole OCR process. In this paper,
we address this issue by proposing a robust method
for detecting the touching characters in Tamil text
document images. We also present a technique for
recognition of isolated Tamil character images using
simple features.
2 PREPROCESSING
In the present system, character images have been
obtained by optical scanning of the character images
on plain paper. The input data obtained by scanning
of printed text is contaminated with noise and
contains redundant information. Preprocessing
includes noise removal, elimination of redundant
information as far as possible, segmentation, and
scaling. The segmentation started by scanning the
487
Hamid Pilevar A..
Tamil Characters Recognition and Retrieval.
DOI: 10.5220/0004030504870493
In Proceedings of the 7th International Conference on Software Paradigm Trends (ICSOFT-2012), pages 487-493
ISBN: 978-989-8565-19-8
Copyright
c
 2012 SCITEPRESS (Science and Technology Publications, Lda.)

page images then continued by horizontally
detecting the Text lines in each scanned page.
Frequency of black pixels in each row is counted in
order to construct the row histogram.
The position between two consecutive lines,
where the number of pixels in a row is zero denotes
a boundary between the lines. After a line has been
detected, it is scanned vertically. In order to find the
column histogram, the number of black pixels in
each column is counted. If there are n consecutive
vertical scans that find no black pixel, we denote
those columns to be a marker between two words.
The value of n is decided experimentally. To
segment the individual character in a word, the
column histogram is found; number of black pixels
in each column is calculated. If there are more than
one consecutive scans that has no black pixels, then
the region is decided to be a marker between
characters, and again if it is more than n, then
considered as marker between two words. The
isolated characters have been normalized so that size
invariant recognition is possible. Though the
recognition is size invariant, better result is obtained
when the size of the characters is assumed to be
within a specific range. The performance of our
algorithm has been tested, by using the fixed size
characters in creating our characters training sets.
For the developed system, 6pt - 36pt character
sizes have been selected. The characters are scaled
to a standard size (40 x 40) using an efficient scaling
algorithm (Kumar et al., 1997). Scanning commonly
introduces noisy cavities in the character images.
These distortions detrimentally affect the shape of
the characters. Single pixel components of noise are
removed from the character images, before feature
extraction.
3 TAMIL ALPHABET
Tamil has many letters, but most of them are derived
from the 12 vowels and 18 consonants. The other
216 letters are made by combining the sounds of a
vowel and a consonant. Even in writing, they are
viewed as the addition of a vowel and consonant.
From more than 300 Tamil fonts which some of
them are very stylish too, six of the simplest fonts
are selected. For providing the test documents, the
Azhagi editor is used. The samples of the six
selected fonts are shown in Figure 1.
Figure 1: Six different fonts “Chanakya, Ilango, Kamal,
Kannadasan Nambi, and Padma” are used in test data.
4 FEATURE SELECTION AND
EXTRACTION
The features are extracted, and vectors are selected
in the follows stages:
1. Project the pixels Vertically, and then calculate
the number of pixels in each column(VPc).
2. Find the Directional Vectors (DV).
3. Calculate the Angles of Directional Vectors
(ADV) with the X axis.
4. Based on discrete directional table (DDA) values,
convert the directional vector angels (ADV) to
eleven connected chain (EDM) codes.
4.1 Vertical Projection Count (VPc)
The vertical projection of all the Tamil characters
(consonant “, , ,  ...” , borrowed consonants
“, , , , ”, vowels isolated form
“, ,  ...”, vowels compound form “, , 
...”) were found, and the black pixels for each
projected column are count.
4.2 Directional Vector (DV)
The variations of the VPC values from one column
to the next column (in the right side) are found and
the directions of the variation are respectively
registered as directional vectors.
4.3 Angle Calculation of Directional
Vectors (ADV)
The angle of directional vectors are calculated with,
1
360
atan( )
2
i i i
yy 


  
where: i = 1 …n-1
n is the number of features in the feature vector

= 3.1416…
ICSOFT 2012 - 7th International Conference on Software Paradigm Trends
488

y
i
is y coordination of starting point of
directional vector i.
y
i+1
is y coordination of ending point of
directional vector i.
For example: the graphical depict of a directional
vectors are shown in Figure 2.
Figure 2: The graphical representation of a directional
vector.
4.4 Eleven Connected Chain (ECC)
In Freeman’s coding method characters are usually
encoded either with 8-connected or 4-connected
chain codes. For the work described in this paper,
11-connected chain codes were used (Figure 3). Our
experiments show that eleven directions method has
better results than 4 and 8 or other numbers of
directions. The technique is named EDM coding
system.
Figure 3: Eleven connected chain coding.
4.5 Discrete Directional Angle (DDA)
In our recognition system, templates are presented
by strings of normalized discrete directional angle
values. The angle values are classified as
demonstrated in Table 1.
The number of the classes is selected to be eleven
classes for being compatible with ECC-codes.
The number of the classes is selected to be eleven
classes for being compatible with ECC-codes.
The number of the classes is selected to be eleven
classes for being compatible with ECC-codes.
Table 1: normalized discrete directional angle values
Angle
Angle
Direction
>
<=
1
-99
-81
2
-81
-63
3
-63
-45
4
-45
-27
5
-27
-9
6
-9
9
7
9
27
8
27
45
9
45
63
A
63
81
B
81
99
4.6 EDM-Code
If the calculated discrete directional angles (DDA)
string for a given directional vector in stage (d) are: -
88, -36, 14, 53, and 86 then its EDM-code by
referring respectively to Table1 is 1479B.
The EDM-codes for all of the Tamil characters
(consonant “    ” , borrowed
consonants “    ”, vowels
isolated form “  ”, vowels compound
form “”) are calculated and saved in
EDM-Table.
4.7 Feature Vector
The smallest extracted features vector belongs to
“” and “” with 13 features and the largest
features vector belong to “” with 61 features
in it. Our scaling template size is eleven features and
the feature vectors are classified based on the
number of fitted templates (NFT) in them. For
example NFT of “” and “” are 2, and the NFT of
“” is 6.
The EDM-codes are selected as characters
feature vectors and saved in the feature vectors table
(Table 2).
5 CHARACTERS RECOGNITION
Template-matching algorithms can recognize
touched and broken versions of character templates
without major difficulty. The reason for this
phenomenon is simple: The difference between a
broken image and the ideal one is relatively small.
Tamil Characters Recognition and Retrieval
489

Table 2: characters Feature vectors table
For template- matching, if a gap is introduced in an
image and it does not change the interpretation of
the image, it increases the distance from the image
to all templates so the recognition is not affected.
Statistical classifiers trained to recognize the most
common broken characters, and in this context a
broken character is better seen as the result of a
feature extractor anomaly.
In contrast, most structural approaches to
classification are confused by broken characters,
since a broken piece may drastically change the
representation of the character structure.
Our approach allows recognition of broken
characters by a template matching technique: the
representations of a character with and without a gap
look alike modulo the graph matching. In other
words, the representation does not decide if the gap
corresponds to a missing part or to a real separation
of strokes, but allows both possibilities to coexist
until the matching with models decides which the
best interpretation of the gap is.
This property for gap representation is just an
extension of our paradigm that all singularities on
the image should appear on the input graph, but only
in the context of the models is the interpretation of
the singularities given.
Thus, gaps should be represented as character
parts that may be missing or not from the ideal
image. Gaps are now positive features with
ambiguous interpretations. This brings up two
problems: first, how to identify all gaps, in other
words, how to define a gap, and, second, how to
consider only possible useful ones, since we want to
increase the number of edges in the input graph as
little as possible.
Figure 4 illustrates the segmentation procedure
of the characters “ ”. The optimal Cutting
Points (CP) can be found after several forward and
backward cutting iterations. The segmented patterns
are best matched to the prototypes. For the
characters containing two or more parts, mistake
may happen in the sequences of the recognized
characters. The broken characters will be rejoined in
the merge process using layout context information.
Figure 4: Graphical representation of directional vectors
for the characters ” “.
6 DYNAMIC RECURSIVE
SEGMENTATION
ALGORITHM
The approach presented in this section, the dynamic
recursive segmentation algorithm, executes a
forward segmentation or a backward merge process
dynamically, based on the recognition result of the
current input array and the neighboring character
arrays, until the connected components are accepted
by the classifier as a valid character. Instead of
registering all the cutting points whose
ICSOFT 2012 - 7th International Conference on Software Paradigm Trends
490

discrimination function values exceed the specific
threshold, only the cutting points determined by the
recursive segmentation algorithm are recorded for
additional processing with context information and
spelling tools. To prevent from misclassifying the
characters whose shape is similar to other characters,
the minimum distance classifier utilizes multiple
rejection thresholds to control the recursive
segmentation process. R
1
, R
2
and R
3
are respectively
used for the initial input patterns (IP), the Residue
input Pattern (RP) and the forward and backward
input patterns (FP- BP). By properly adjusting R
1
,
R
2
and R
3
(R
1
< R
2
<R
3
), the algorithm achieves the
optimal results.
7 MATCHING THE FEATURE
VECTORS
Given the feature vector for a word, our method
finds vectors and sub-vectors that are homeomorphic
to some prototype. A distance function between
vectors measures the amount of distortion between
vector, sub-vector and prototype. This distance
function represents the minimum transformations
that a vector will undergo so that the matching is
possible. Thus, it is used as a measure of the quality
of the matching. The cost of matching two vectors is
the distance between the vectors.
The cost function and the algorithm that is used
to find sub-vectors are basically the same used for
recognition of segmented words. In order to find
sub-vectors of the entire word vector that match
prototypes, the recognition process is initiated on all
features of the vector. The prototypes will guide the
best recognition that contains the initial feature. The
cost of the sub-vector matching does not include
features that are not matched in the candidate, since
they may belong to a different character.
8 SEARCH ALGORITHM FOR
FEATURE VECTOR
A special tree search technique named EDM-Codes
method is applied, the feature vectors are looked for
using a tree based searching technique as displayed
in Figure 4. The search process is as follows:
1- Extract the image of a word
2- Calculate the feature vector of the image
3- If the number of features is larger than 66, then
it can be touching characters problem(follow
the related routine: executes a forward
segmentation or a backward merge process
dynamically as discussed in sections 4, 5, and 6
)
4- Look in the search tree for the most similar
feature vector
5- If N features are matched then Match = N*
1.515
6- If Match is equal to 100 then it is exactly
matched
else
if Match is less than 100 then
it can be a broken character (follow the
related routine: executes a forward
segmentation or a backward merge process
dynamically as discussed in sections 4, 5, and
6 )
Let us for example: a query image is given and
its feature vector is extracted by the system, then the
searching process will be followed through the
nodes of the search tree as marked in Figure 5.
Figure 5: The search tree: each state represents a character
candidate and the remaining pattern. A path from the root
to a node constitutes a segmentation attempt. Search
pattern for the features vector 39…AB is marked.
9 DISTANCE METRIC
The characters are classified based on the above
mentioned features vectors and for similarity
measurement the weight method is used (Pilevar
2005). The similarity degree S
i
, between the i-th
elements of the feature vectors of character images
f(x) and g(x) is defined as:
Tamil Characters Recognition and Retrieval
491

if (
ij
fg
), then S
i = 1
else
 
 
min ,
max ,
ij
i
ij
fg
S
fg

Where:
f
i
is the feature vector of the i-th character
in the mean set of the training sets
g
j
is the feature vector of the j-th character
in the document.
The similarity degree between character images
f(x) and g(x) is the sum of the similarity degrees
between the corresponding n elements of the feature
vectors derived from the two images, and defined as:
1
n
ii
i
SS 



Where:
10 
i
S
1
1
n
i
i




n is the number of features , n<=66
10 EXPERIMENTAL RESULTS
AND DISCUSSION
The effectiveness and performance of our algorithm
have been tested on samples collected from various
images of legal documents belonging to one city.
For testing our method, around 200 printed Tamil
text documents are scanned at 300 dpi and binarized
using the two-stage method described in (Dhanya et
al., 2001). For providing the text documents, the
Azhagi editor is used. Six different fonts (Figure 1)
are implemented for creating the documents.
The textual lines and words segments are
determined from valley points in the horizontal and
vertical projection profiles. A one-pixel margin is
kept while detecting zone boundaries of the
characters. However, it is assumed that all the
characters of a text line are of the same font size.
The extracted characters are normalized to the size
of 40×40 pixels. After the character boxes are
extracted, before starting the character recognition
and recognizing the touching characters process, the
documents are checked for their skew (Pilevar and
Ramakrishnan, 2006). In the ultimate experiment,
ten sets of separate documents given to the EDM
software system, the documents are segmented into
about 100000 characters, and more than 97% of
characters are recognized correctly. However we
couldn’t find any similar work to compare ours with,
but we believe that the outcome of this research is
satisfactory and can be used as a base in Tamil
character recognition systems in practical works.
REFERENCES
Bansal V., R. Sinha, “Segmentation of touching and fused
Devanagari characters”, Pattern Recognition 35, 875–
893, 2002.
Davessar N. M., S. Madan, and H. Singh, “A Hybrid
Approach to Character Segmentation of Gurmukhi
Script Characters,” Pattern Recognition, pp. 4-8, 2003.
Dhanya, D.: “Bilingual OCR for Tamil and Roman scripts.
Master’s thesis, Department of Electrical
Engineering”, Indian Institute of Science, 2001.
Electronics N., C. T. Center, and K. Luang, “Using
Projection and Loop for Segmentation of Touching
Thai Typewritten,” Analysis, vol. 2004, pp. 504-508,
2004.
Faure, C., Vincent, N., “Simultaneous detection of vertical
and horizontal text lines based on perceptual organi-
zation”, Proceedings of SPIE - The International
Society for Optical Engineering, Volume 7247, 2009.
Grailu, H., Lotfizad, M., Sadoghi-Yazdi, H, “A
lossy/lossless compression method for printed typeset
bi-level text images based on improved pattern
matching”, International Journal on Document
Analysis and Recognition, pp. 1-24, 2009.
Hotta, Y., Fujimoto, K., “Line-touching character
recognition based on dynamic reference feature
synthesis”, Proceedings of SPIE - The International
Society for Optical Engineering Volume 6815, 2008.
Kumar S. and Muhammad Mashroor Ali, “An Efficient
Object Scaling Algorithm for raster device”, Graphics
and Image Processing, NCCIS, 1997.
Li Y., S. Naoi, and M. Cheriet, “A Segmentation Method
for Touching Italic Characters,” Pattern Recognition,
pp. 2-5, 2004.
Li Y., S., M. Cheriet, Ching Y, Suen, “A Segmentation
Method for Touching Italic Characters”, Proceedings
of the 17th International Conference on Pattern
Recognition (ICPR’04), 1051-4651/ 2004.
Liang S., M. Shridhar and M. Ahmadi, “Segmentation of
touching characters in printed document recognition”,
Pattern Recognition, Vol. 27, No. 6, pp. 825 840,
1994.
Lu X., X. Liu, G. Xiao, E. Song, P. Li, and Q. Luo, “A
Segment Extraction Algorithm Based on Polygonal
Approximation for On-Line Chinese Character
Recognition,” Japan-China Joint Workshop on
Frontier of Computer Science and Technology, pp.
204-207, 2008.
Ode, Ã., Tveit, M., Fry, G., “Capturing landscape visual
character using indicators: Touching base with
landscape aesthetic theory”, Landscape Research,
volume 33, Issue 1, pp. 89-117, February 2008.
Pilevar A. H, A. G. Ramakrishnan, “Inversion detection in
text document images”, 9th Joint Conference on
Information Science, Taiwan, 2006
Pilevar A. H., “Retrieval of signal from Biomedical
Databases some new approaches”, Ph D thesis,
University of Mysore, 2005.
Sattar, Md. A., Mahmud, K., Arafat, H., Noor Uz Zaman,
A. F. M., “Segmenting Bangla text for optical
ICSOFT 2012 - 7th International Conference on Software Paradigm Trends
492

recognition, 10th International Conference on
Computer and Information Technology”, ICCIT, 2007.
Wang W., “Printed Chinese Character Recognition Based
on Pixel Distribution Probability of Character Image,”
2008 International Conference on Intelligent
Information Hiding and Multimedia Signal
Processing, pp. 1403-1407, 2008.
Watcharabutsarakham S., “Segmentation for touching thai
typewrittens,” Science, pp. 199-202, 2004.
Tamil Characters Recognition and Retrieval
493