Script Identification In Trilingual Indian Documents

R. R. Aparna & R. Radha
International Journal of Image Processing (IJIP), Volume (8) : Issue (4) : 2014 178
Script Identification In Trilingual Indian Documents
R. R. Aparna skandhur@gmail.com
Research Department of computer Science
S.D.N.B.Vaishnav college
AFF. Madras University
Chennai, 600043, India
R. Radha radhasundar1993@gmail.com
Research Department of computer Science
S.D.N.B.Vaishnav college
AFF. Madras University
Chennai, 600043, India
Abstract
This paper presents a research work in identification of script from trilingual Indian documents.
This paper proposes a classification algorithm based on structural and contour features. The
proposed system identifies the script of languages like English, Tamil and Hindi. 300 word images
of the above mentioned three scripts were tested and 98.6% accuracy was obtained.
Performance comparison with various existing methods is discussed.
Keywords: Script, Dilation, Boundary, Centroid, Zone, Blob, Bounding Box.
1. INTRODUCTION
India being a multi-lingual country, most of the documents contains two (bilingual) or three
language (trilingual) scripts. English and Hindi are the most prevalent language being used in
India along with the state language. Important documents like, government documents and
question papers of every state contains two or three languages. Scrip is referred as the graphics
part of the writing system. The writing style and the characters are combined in the script class. A
script can be used by only one language or many languages. Indian and South East Asian
languages are based on Brahmic scripts. North Indian languages are based on Devanagri,
Bengali, Manipuri, Gurumukhi, Gujarati and Oriya scripts. South Indian languages are based on
Tamil, Telugu, Kannada and Malayalam scripts. Automatic identification of scripts is very
essential before feeding into multi-script optical character recognition (OCR) system. Many
methods exist in the literature for script identification. In section 2 existing methods in this field are
discussed. Properties of English, Hindi and Tamil are discussed in section 3. In section 4 the
proposed methodology is explained. Results and performance comparison with various existing
methods are detailed in section 5. Section 6 deals with the conclusion and future work.
2. LITERATURE SURVEY
Various methods exist in the literature for multi-script identification. The script identification is
performed in two ways as local approach and global approach. Identification of script at text line
level or word level or character level is defined as local approach. In global approach the entire
text block is considered [1]. The following methods identify multi-scripts at text line level. [2], [18]
used shape features, statistical features and Water Reservoir based features and classified using
a rule based classifier.[3] have used horizontal projection profile (HPP) and derived a rule based
classifier by calculating the threshold using the successive maxima. Structural features, K-
Nearest Neighbour (KNN) and support vector Machine (SVM) was used by [4] for classification.
Profile based features were used by [5] and classified using KNN. Structural, topological, contour
and Water Reservoir based features were used by [6] and classified using rule based classifier.

Character level script identification technique using Gabor and Gradient features and SVM for
classification [7]. The following methods were performed at word level. Eight features were
obtained by [8] based on top and bottom max row, top horizontal line, tick component, top and
bottom holes, vertical lines and bottom component and classified using KNN. [9] performed
Devanagri script identification using the presence of headlines as [2] and Kannada script
identification using morphological operation. Their method was restricted to font style and size.
[10] have used 400 and 64 dimensional features and SVM was used for classification. [11] has
used Neural networks (NN) for classification and features based on histogram mean distance,
pixel value and vertical zero crossing. [12] used Zone based Gabor features and SVM classifier
was used.
3. PROPERTIES OF ENGLISH, HINDI AND TAMIL SCRIPTS
English language uses Latin or Roman script. English follows the Alphabetic system. Hindi
language uses Devanagri script. The characters of the Hindi words are connected using a
headline called “Shirorekha”. Printed Tamil and English words contain isolated characters. The
structure of the three scripts is divided into three zones (Fig.1). The upper zone (UZ), middle zone
(MZ) and lower Zone (LZ). The character set of English constitutes 52 characters comprising 26
uppercase and 26 lowercase characters. The uppercase English characters stay inside the
middle zone. The lowercase English characters like b, d, f, h, k, l, t containing ascenders
occupies UZ and those containing desecenders like g, j, p, q, y occupies the LZ. But the presence
of pixels in the LZ and UZ by lower case English alphabets is of symmetrical pattern (Fig1.a).
Tamil character contains 12 vowels and 18 consonants. 216 compound characters were formed
by combining vowels and consonants. Almost all the character of Tamil occupies UZ and LZ.
Tamil character that occupies the UZ and LZ has varied pixel density and patterns (Fig1.b). Hindi
character contains 34 consonants and 10 vowels and form combined characters. Since Hindi
characters are connected using headline, the probability of pixel density in the LZ and UZ is very
less (Fig1.c).
FIGURE 1: a) English script b) Tamil script c) Hindi script.
4. PROPOSED METHODOLOGY
The input image is converted to greyscale image and binarized using Ostu method [17], (Fig.2).
The proposed algorithm works by dividing the multi-script words into three zones and the pixel
densities are calculated in the UZ above Headline (HL) and LZ below base line (BL) for
discriminating multi-script (fig.1). The proposed algorithm is based on the contour features and
classifies using a rule based classifier. First it checks whether the script belongs to Hindi by
using the pixel density of UZ and LZ. Classification of English and Tamil script is performed using
the contour features of the boundary pattern such as upper boundary (UB) and lower boundary
(LB) pixel density calculation.

FIGURE 2: Binarized images: a) Tamil b) English c) Hindi.
4.1 Classification Phase - I
The given word image is classified in two phases. In phase - I the image is classified to check the
presence of Hindi word. In phase - II, it is classified as English or Tamil.
Step 1: The heights ( ih ) and widths ( iw ) of all the blobs are calculated for the image (Fig2).
Where i=1 to N, N is the number of blobs.
Step 2: Calculate the average width (w )iavg and (h )iavg average heights of all blobs.
Step 3: Divide every ih by (h )iavg and store in an array.
1
( )
i
i
h
v
avg h
= (1)
Step 4: Choose the 1min( )v as it is closer to the average (h )iavg .
Step 5: Extract the minx , miny and ht for the chosen ih (blob).
Step 6: Calculate the height of the MZ using the following formula. (Fig.3)
( )ymh round min ht= + (2)
FIGURE 3: Middle Zone (MZ) Height.
Step 7: Count the pixels row wise for all rows of the entire image and store it in an array (s1).
Step 8: Count the pixels above HL and below BL using
min max
1
2 1,s3 1
y y
i i mh
s s s
= =
 
= = 
 
∑ ∑ (3)
Step 9: Count the no of blobs
The word is classified as Hindi if pixels in BL and HL is zero and the total number of blobs is
equal to one as the Hindi word is connected by a headline otherwise it is classified as English or
Tamil in the next phase.

4.2 Classification Phase – II
Step 1: Extract the boundary features using Morphological operations
Step 2: Dilate the image.(Fig.4a)
Step 3: Calculate the centroid value xc and yc .
Step 4: Extract the boundary co-ordinates and store in the array, ubx , uby and lbx , lby .
(Fig. 4 & 5)
FIGURE 4: a) Obtained Boundary (English) FIGURE 5: a) Obtained Boundary (Tamil) b) UB c) LB.
b) UB c) LB.
Step 5: Extract the upper boundary pixel co-ordinates ( yub ) and lower boundary ( ylb ) pixel co-
ordinates using the following formula,
max( )
1
,y ( )
uby
y ub ub y
i
ub y i c
=
= <∑ (4)
(5)
Step 6: Calculate the total count of pixels in the upper boundary and lower boundary. If the total
pixels is greater than the threshold then it is classified as Tamil else English. The basic idea is
that, the upper boundary of Tamil words (Fig.6a) contains unsymmetrical patterns (peaks and
valleys) by accumulating more pixels. But in English word (Fig.6b) the boundary shape pattern is
more symmetrical without jerks. Hence the pixels between ymin and maxy will be zero or less
for English. Always the characters of Tamil font will be little bigger in size compared to the English
character of same font size.
FIGURE 6: Scatter plot for UP of Tamil and b) Scatter plot for UP of English.
max( )
1
,y ( )
lby
y lb lb y
i
lb y i c
=
= >∑

5 RESULTS AND DISCUSSION
Samples were collected from the trilingual documents. 300 word samples of Hindi, English and
Tamil were tested using the proposed algorithm and 98.6% accuracy was obtained. In cases
where the Tamil words were very small and having symmetrical pattern was misclassified as
English. The error rate is very low as always the number of pixels will be greater for Tamil
compared to English.
Script Hindi English Tamil
Hindi 100 0 0
English 0 98 2
Tamil 0 2 98
TABLE 1: Accuracy of Script Identification.
There is lack of any comparative analysis of the results with most of the reported works in script
recognition. Experimental results of every proposed method have not been compared with other
benchmark works in the field [1]. As there was no standard evaluation measures available in the
literature. The results were evaluated as using the following statistical measures (Table. 2 & 3).
The proposed algorithm is evaluated using the statistical measures Recall, Precision, Accuracy
and Fscore based on the following metrics. (H-Hindi, E-English and T-Tamil)
True Positive (TP): Correctly identified: (H, E, T identified as H, E, T).
False Positive (FP): Incorrectly identified: (E, T incorrectly identified as T, E).
True Negative (TN): Correctly rejected.
False Negative (FN): Incorrectly rejected.
5.1 Statistical Measure
Recall: This gives the positive cases obtained. It is same as the sensitivity to identify positive
results.
Precision: This measure gives the percentage of positive predictions.
Precision
TP
FP TP
=
+
Accuracy: This measure gives the predictions that were correct.
Fscore: This measure is used as a single measure of performance test by combining all the
above results.
Pr * Re
Fscore 2
Pr Re
ecision call
ecision call
 
=  
+ 
Re
TP
call
TP FN
=
+
Accuracy
TN TP
TP FN
+
=
+

TABLE 2: Statistical Results. TABLE 3: Accuracy of Classification.
5.2 Performance Comparison with the Existing Methods
The proposed algorithm is compared with many of the existing methods in the literature for script
identification of Hindi (H), English (E), Kannada (K), Devanagri (D), Oriya (O),Telugu (Te),
Gurumukhi (G), and Tamil (T) using local approach.
Author & year Script Local
Approach
Technique
Used
Accuracy Remarks
[14]-2002 E, H & K Word-wise NN 98% Fixed font size
and
computationally
complex.
[12]-2003 12 scripts Text line Local features
and water
reservoir
based feature
98% Many features
were extracted
from individual
characters
[13]-2010 K, H & E Text line Top and
bottom profile,
KNN classifier
99.75% Dependent on
font size, time
consuming
[16]-2002 T & E Word-wise Gabor
features, SVM
classifier
96% Computationally
complex
[5]-2005 D, E & U Word-wise Profile based
features,
Water
reservoir
method
97.51% Character level
features were
extracted.
[9]-2009 E, D & B Word-wise 64 & 400
dimensional
features and
SVM as
classifier
98.51% Error rate
increases when
number of
character is
less than three
[3]-2013 H, B, Te &
K
Text line Projection
profile and rule
based
classifier
97.83% Fixed font size
was used
[6]-2013 G & E Character
level
Gabor,
Gradient
features and
SVM classifier
99.45% Complex and
Time
consuming
[10]-2013 D Word-wise Mean distance,
pixel value and
zero crossing,
NN
90% Computationally
complex and
execution time
is very high
[7]-2011 H &E Word-wise Many features, 95%
Predicted
-ve
Predicted
+ve
-ve
case
TN: 0 FP: 4
+ve
case
FN: 0 TP: 296
Measure Accuracy
Accuracy 100%
Recall 100%
Precision 98.6%
Fscore 99%

KNN
[8]-2011 K, E & H Word-wise Features
based on
Morphology
and KNN
95.54% Suitable only for
fixed font size
and font style
[15]-2002 D, T & O Word-wise Gabor filters 97.33 Computationally
complex
Proposed
method-2014
H, E &T Word-wise Zonal and
boundary
features
using rule
based
classifier
98.6% Simple, fast,
efficient and
accurate
TABLE 4: Performance Comparison.
6. CONCLUSION AND FUTUREWORK
The existing methods are computationally complex and have used more features based on
structural and pixel density feature. But the proposed method is very simple, efficient, fast and
accurate for discriminating the scripts in bilingual or trilingual documents. Majority of the Indian
bilingual or trilingual documents contains 2 or 3 scripts. The documents contain Hindi, English
and the state language. Hence the proposed algorithm can be used for script identification of
trilingual documents containing Tamil, Hindi and English or Bilingual documents containing any
two combinations of these three scripts. The future work will be concentrated towards script
identification of Kannada, Telugu, and Malayalam in trilingual or bilingual documents with the
combination of Hindi and English as the other two languages.
7. REFERENCES
[1] D.Ghosh, T.Dube and A.P.Shivaprasad. “Script Recognition: A review,” IEEE
Transactions on pattern analysis and machine Intelligence, vol. XX, 2009.
[2] U.Pal and B.B.chaudhuri. “Identification of different script lines from multi-script documents,”
Image and Vision Computing, vol. 20, pp. 945-954, 2002.
[3] O.Prakash, V.Shrivastava and A.Kumar. “An efficient approach for script identification,”
International journal of computer Trends and Technology (IJCTT), vol. 4, 2013.
[4] Rajesh Gopakumar, N.V.SubbaReddy, Krishnamoorthi Makkithaya and U.Dinesh Acharya.
“Script Identification from multilingual Indian documents using structural features,” Journal of
computing, vol. 2, 2010.
[5] S.Chanda and U.Pal. “English , Devnagari and Urudu text Identification,” in Proc.
International conference on Document Analysis and Recognition, pp. 538-545, 2005.
[6] R.Rani, R.Dhir and G. Lehal. “Script Identification of pre-segmented multifont characters and
digits,” in Proc. ICDAR, 2013, pp.1150-1154.
[7] M.Swamy Das, D.Sndhya Rani, C.R.K.Reddy and A.Govardhan. “Script identification from
multilingual Telugu, Hindi and English Text documents,” International Journal of Wisdom
based computing, vol.1, 2011.
[8] B.V.Dhandra and M.Hangarge. “Morphological reconstruction for word level script
identification,” International journal of computer science and security, vol.1, 2011.

[9] S.Chanda, S.Pal, K.Franke, U.Pal. “Two-stage approach for word wise script Identification,”
in Proc. ICDAR, 2009.
[10] D.Yadav, S.Sanchez-cuadrado and J.Morato. “Optical acharacter recognition for Hindi
language using a NN approach,” Journal of information processing systems, vol. 9, 2013.
[11] R.Rani, R.Dhir and G.Lehal. “Modified Gabor feature extraction method for word levelscript
identification –Expermentation with Gurumukhi and English scripts,” International journal of
signal processing, image processing and pattern recognition, vol. 6, pp. 25-38, 2013.
[12] U.pal and B.B.Chaudhri. “Indian script Character recognition: A survey,” Pattern recognition,
vol. 37, pp.1887-1889, 2004.
[13] M.C.Padma and P.A Vijaya. “Script identification from trilingual documents using profile
based feature,” International journal of computer science and applications, vol.7, pp.16-33,
2010.
[14] S.B. Patil and N.V.Subbareddy. “Neural network based system for script identification in
Indian documents,” Sadhana, vol. 27, pp. 83-97, 2002.
[15] P.B.Pati, S.S.Raju, N.Pati and A.G.Ramakrishnan. “Gabor filters for Document analysis in
Indian Bilingual Documents,” in Proc. IEEE ICISIP, 2004.
[16] D.Dhanya, A.G.Ramakrishnan and P.B.Pati, “Script identification in printed bilingual
documents,” Sadhana, vol. 27, pp. 73-82, 2002.
[17] R.C.Gonzalez, R.E.Woods and S.L.Eddins. “Digital Image Processing using Matlab,”
NewDelhi, Tata Mcgraw-Hill, 2011.
[18] U.Pal and B.B.Chaudhri. “Script line separation from Indian Multi script documents,” in Proc.
ICDAR, pp. 406-409,1999.

Script Identification In Trilingual Indian Documents

More Related Content

What's hot (12)

Similar to Script Identification In Trilingual Indian Documents (20)

Recently uploaded (20)

Script Identification In Trilingual Indian Documents