High-Level Shape Representation in Printed Gujarati Characters
Mukesh M. Goswami
1
and Suman K. Mitra
2
1
Faculty of Technology, Dharmsinh Desai University, Nadiad, Gujarat, India
2
Dhirubhai Ambani Inst. of Information and Communication Tech., Gandhinagar, Gujarat, India
Keywords:
Pattern Recognition, Character Shape Representation, Shape Similarity, Character Recognition, Gujarati
Characters.
Abstract:
This paper presents extraction and identification of the high-level stroke (HLS) from printed Gujarati charac-
ters. The HLS feature describes a character as a sequence of predefined high-level strokes. Such a high-level
shape representation enables approximate shape similarity computation between characters and can easily be
extended to word-level. The shape similarity based character and word matching have extensive application
in word-spotting based document image retrieval and character classification. Therefore, the proposed fea-
tures were tested on printed Gujarati character database consisting of 12000 samples from 42 different symbol
classes. The classification is performed using k-nearest neighbor with shape similarity measure. Also, a shape
similarity based printed Gujarati word matching experiment is reported on a small word image database and
the initial result are encouraging.
1 INTRODUCTION
India, being a multilingual country, has more than
22 officially listed languages written in 12 different
scripts. Substantial work in character classification,
OCR, and word-spotting is reported in the literature
for dominating Indian scripts, like Devanagari, Ben-
gali, Tamil, and Telugu. However, many scripts such
as Gujarati still lakes attention of researchers. Despite
years of efforts, the word-level accuracy of the OCR
system for many Indian scripts have remained low as
compared to western text mainly due to the large and
complex character set including base characters, mod-
ifiers, and conjunct symbols (Kompalli et al., 2005).
Therefore, the researchers are motivated to explore
the recognition free approach for document image re-
trieval in many Indian scripts (Srihari et al., 2006;
Hassan et al., 2009; Tarafdar et al., 2010; Jawahar
et al., 2004). As a result, the shape similarity based
character and word recognition for Indian scripts have
gained considerable interest in recent time.
The majority of the recognition free systems de-
pends on the shape-based features for the matching
of characters and words (Doermann, 1998). Such
a system demands features that are compact yet ef-
ficient in describing the high-level shape of charac-
ters. Also, it should be easy to compute the shape
similarity between characters and words (Yang et al.,
2008). Therefore, The current paper investigates a
technique for compact and high-level shape represen-
tation of characters using the sequence of predefined
high-level strokes (HLS). Such a sequential represen-
tation facilitates an efficient shape similarity match-
ing between characters and words using dynamic pro-
gramming based algorithm.
The rest of the paper is organized as follow: we
start with a brief discussion on the related work in
Section 2. Section 3 outlines the representation, ex-
traction, and identification of high-level stroke. Sec-
tion 4 discusses shape similarity computation be-
tween characters. Section 5 describes the experimen-
tal setup for classification of printed Gujarati char-
acters using k-nearest neighbor and shape similarity
measure as well as the shape similarity based word-
matching experiments. Finally, the paper is concluded
in Section 6.
2 RELATED WORK
Much work is found in the literature for the classifi-
cation of characters from both North as well as South
Indian script families. The character classification for
prominent North Indian scripts like Devanagari, Ben-
gali, Gurmukhi, and Oriya is discussed in (Chaud-
huri and Pal, 1998; Chaudhuri et al., 2001; Lehal
418
Goswami, M. and Mitra, S.
High Level Shape Representation in Printed Gujarati Character.
DOI: 10.5220/0006191104180425
In Proceedings of the 6th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2017), pages 418-425
ISBN: 978-989-758-222-6
Copyright
c
 2017 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved

and Singh, 2000), whereas the work on major Dra-
vidian languages such as Tamil, Telugu, and Kannada
is available in (Jawahar et al., 2003; Aparna and Ra-
makrishnan, 2002; Lakshmi and Patvardhan, 2002).
The work found for the printed Gujarati charac-
ter classification is less compared to many other In-
dian scripts. Some of the early contributions include
work by Antani and Agnihotri (Antani and Agnihotri,
1999), who used moment features with minimum
Hamming distance and k-Nearest Neighbor (kNN)
classifier to claim an accuracy of 67% on a small
database of 800 samples. The most noticeable con-
tributions include work by Dholakia et al. (Dholakia
et al., 2007) that uses wavelet features with Neural
Network and kNN classifiers. The overall accuracy
claimed was of 96-97% on a database of 4173 sym-
bols of 119 classes. Goswami et al. (Goswami et al.,
2011) have used Self-Organizing Map(SOM) projec-
tion with the k-NN classifier and reported 84% ac-
curacy on the moderately sized database. Hassan et
al. (Hassan et al., 2014) have used Multiple Ker-
nel Learning based Support Vector Machine (MKL-
SVM) classifier with multiple features, like fringe
distance map (FM), shape descriptor (SD), the his-
togram of gradients (HoG). The accuracy claimed was
97-98% on a database of 16000 symbols including
modifiers, base characters, and conjuncts. Recently,
Goswami and Mitra (Goswami and Mitra, 2015) have
used low-level stroke features with the k-NN classi-
fier for printed Gujarati character classification and
claimed an accuracy of 95-98%.
Some initial work reported for word image re-
trieval on Devanagari, Bengali and Sanskrit script in-
cludes (Chaudhury et al., 2003; Srihari et al., 2006;
Bhardwaj et al., 2008) which uses Geometric Graph,
GSC, and Moment features, respectively. (Kumar
et al., 2007) and (Meshesha and Jawahar, 2008) gives
major contribution in document image retrieval for
Telugu script. The experiment uses multiple fea-
tures like Fourier descriptor, projection profiles, mo-
ments, etc., and the word images are compared using
DTW (Rath and Manmatha, 2003). The experiments
were carried out on a huge word image database ex-
tracted from 1800 pages of 7 machine-printed Telugu
books. Other noticeable contribution includes (Has-
san et al., 2009) that uses shape descriptor features
with hierarchical locality sensitive indexing for word
image retrieval from Devanagari, Bengali, and Malay-
alam scripts. (Tarafdar et al., 2010) have used a se-
quence of shape code with string edit distance for
word image matching from Devanagari, Bengali, and
Gurmukhi script. To the best of our knowledge, no
work is reported in the literature for word image
matching and retrieval for Gujarati script.
Figure 1: Formation of character as a set of high-level
strokes where each high-level stroke is described as some
sequence of shape primitives(i.e. lines, curves, and points).
The majority of the work reported until now uses
transform domain, geometrical, or statistical features
that give only a local view of the character shape and
also generate a large feature vector. The global shape
of the character can well be described using the native
high-level strokes used to form a character. However,
the decomposition of character into native stroke is
not experimented much for the Indian script.
3 HIGH-LEVEL STROKE
FEATURES
In the case of off-line text, the HLS can be described
as a sequence of object pixel between two feature
points in one pixel wide thinned character image(as
shown in Figure 1). 48 major and minor HLS are
identified that are visually non-redundant and suffi-
cient to describe any middle zone symbols in Gujarati
script (see Figure 2). Each HLS, in turn, is described
as a sequence of shape primitives like lines, curves,
junction points, and endpoints. Figure 3 shows the
outline of the proposed method. The process be-
gins by taking Low-Level Stroke (LLS) matrix com-
puted using template matching algorithm proposed
in (Goswami and Mitra, 2015). The HLS is extracted
from the LLS matrix using a junction point based
stroke scanning algorithm (discussed in Section 3.2).
The extracted HLS are identified using linear chain
Conditional Random Field (CRF) (Charles and Mc-
Callum, 2011) (discussed in Section 3.3).
3.1 Representation of High-Level
Stroke
The HLS used to describe a character in Gujarati
script are highly cursive and elongated (see Figure 2)
High Level Shape Representation in Printed Gujarati Character
419

Figure 2: The set of major and minor high-level strokes
present in Gujarati characters.
Figure 3: Outline of the high-level stroke extraction from
printed Gujarati characters.
hence direct mathematical representation is infeasi-
ble. However, any complex HLS shape can be de-
scribed as a sequence of shape primitives like points,
lines, and curves(Figure 1). A template matching
based algorithm was proposed in (Goswami and Mi-
tra, 2015) to extract such shape primitives called low-
level strokes (LLS). The algorithm takes MxN skew
corrected, binarized, thinned character image as an in-
put and generates a MxN matrix of LLS as output (as
shown in Figure 4).
Figure 4: Output of the Low-Level Stroke extraction algo-
rithm(Goswami and Mitra, 2015).
The LLS features can be used as elementary build-
ing blocks to represent a HLS (as shown in Figure 1).
Apart from the sequence of LLS, it is also needed to
know the direction information of LLS. For example,
as shown in Figure 5(a), the sequence of LLS is same,
but the direction is different resulting into two distinct
HLS. Thus, the HLS can be defined as – a sequence
of Direction Encoded Low-Level Strokes (DLLS) be-
tween two feature points(i.e. junction point or end-
points). Figure 5(b) shows 10 basic LLS combine
with 8 direction code to obtained 18 different DLLS
(Figure 5(c)).
3.2 Extraction of High-Level Stroke
A junction point based stroke extraction algorithm is
proposed to obtain the direction and sequence infor-
mation. The algorithm takes MxN matrix of LLS ob-
tained from the input character image (shown in Fig-
ure 4) and extracts the high-level strokes present in
the character. Each LLS in the sequence is replaced
by corresponding DLLS depending on the direction
information obtained while scanning. The necessary
steps for the junction point based scanning algorithm
are described as follow.
1. STEP 1: If a character has more than one stroke
it has at least one junction point. During the first
step, the middle region of LLS matrix is scanned
in left to right order to obtain the list of junction
points present in the matrix(Figure 6(a)).
2. STEP 2: The 3x3 neighborhood of each junction
point in the list is scanned in clockwise order to
obtain the starting point of each HLS originating
from the given junction point. The clockwise scan
resembles left to right and top to bottom writ-
ing order of Gujarati script. The touching junc-
tion points are handled by recursively invoking the
scanning algorithm for each junction in the 3x3
neighborhood of current junction point.
3. STEP 3: A contour tracing algorithm is used to
extract the HLS starting from the start point (ob-
tained in step 2) till either an end point or a junc-
tion point is not reached. Each LLS in the contour
has exactly two neighbors, already visited previ-
ous point and next unvisited point. The direction
of LLS is obtained by finding the relative position
of next point w.r.t. the current LLS. Thus, LLS
is combined with direction information to obtain
DLLS used to generate the output sequence. The
HLS so extracted are deleted from the LLS matrix
to avoid duplication.
4. STEP 4: Repeat Step 2 and 3 for each junction
point in the list obtained in Step 1
ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods
420

Figure 5: (a) Different HLS with same sequence of LLS but different direction (b) Direction encoding of LLS (c) Set of
Directional Embedded Low-Level Strokes (DLLS).
5. STEP 5: Many characters in the Gujarati script
are formed using single high-level strokes only,
hence does not have any junction point. In such
scenario, the starting point is obtained by find-
ing the first endpoint in the top-left quadrant of
the LLS matrix using zigzag scan. If the end-
point does not exist in the top-left quadrant then
top-right, bottom-left, and bottom-right quadrants
are scanned respectively, to obtained the starting
endpoint. If the character contains neither a junc-
tion point nor an endpoint (for example symbol
”zeros”), then the first LLS obtained in zigzag or-
der from top-left quadrant is selected as starting
point(see Figure 6(b)). Once the starting point is
obtained the contour tracing algorithm discussed
in step 3 is used to extract the HLS.
3.3 Identification of High-Level
Stroke(HLS)
The stroke extraction algorithm discussed in the pre-
vious section represent every character sample hav-
ing k HLS as a set of k order pairs (y
1
, X
1
), (y
2
, X
2
)
. . . , (y
k
, X
k
) where y
i
gives i
th
HLS label and X
i
gives corresponding DLLS vector. The identifica-
tion of HLS refers to the problem of predicting the
sequence of values y
1
, y
2
, . . . , y
k
given the sequence
of DLLS vectors X
1
, X
2
, . . . , X
k
where each X
i
=
h
x
i1
, x
i2
, . . . , x
im
i
. The stroke label y
i
∈
{
1, 2, . . . , 48
}
and elements of DLLS vector x
i j
∈
{
1, 2, . . . , 18
}
.
Since each HLS y
i
in the character depends on
Figure 6: High-Level Stroke Extraction using (a) Junction
Point (b) Endpoint or (c) First LLS in zigzag order from
top-left corner.
corresponding DLLS vector X
i
as well as other y’s
present in the character, the Conditional Random
Fields (CRF) (Charles and McCallum, 2011), an
undirected probabilistic graphical model, is employed
next to identify the HLS from the DLLS vector. The
selection of CRF is justified in this context because
it not only captures the dependency between HLS y
i
and DLLS vector X
i
but also consider the dependency
between the current HLS y
i
with other HLS present
in the character. In the simplest case, known as linear
High Level Shape Representation in Printed Gujarati Character
421

chain CRF, the current HLS y
i
depends on DLLS vec-
tor X
i
and the previous HLS y
i−1
. The conditional
probability of the set of HLS Y =
{
y
1
, y
2
, . . . , y
k
}
given the set of corresponding DLLS vectors X =
{
X
1
, X
2
, . . . , X
k
}
can be computed using Equation 1.
P(Y |X) =
1
Z(X)
n
∏
i=1
exp
∑
j
λ
j
f
j
(y
i
, y
i−1
, X
i
)
!
(1)
where Z(X) gives normalizing factor, f
j
gives fea-
ture function and λ
j
gives parameters of CRF, which
needs to be learned from labeled training database us-
ing standard gradient optimization algorithm like L-
BFGS. The HLS y
i
in each character sample are la-
beled manually using stroke labeling tool to gener-
ated a labeled stroke database required for training the
CRF model.
4 SHAPE SIMILARITY
COMPUTATION
CRF model discussed in the previous section
takes as an input the sequence of DLLS vec-
tors
{
X
1
, X
2
, . . . , X
k
}
and labels corresponding HLS
{
y
1
, y
2
, . . . , y
k
}
for each character. Thus, a character,
after HLS identification, is represented as a sequence
of HLS C =< y
1
, y
2
, . . . , y
k
>. Therefore, the approx-
imate shape similarity between the characters can be
obtained by finding the similarity between HLS se-
quences for the given characters. Since it is needed to
compare the entire HLS sequence (end to end align-
ment) to match the characters, the global sequence
alignment techniques proposed by (Needleman and
Wunsch, 1970) can be used to find the regions of sim-
ilarity between two HLS sequences.
The algorithm, shown in Figure 7, gives the length
of the maximum matching sub-sequence between two
sequences. However, unlike traditional Longest Com-
mon Subsequence (LCS) algorithm which does not
assign a penalty to mismatch, the NW algorithm as-
signs a penalty of -1 to mismatch as well as the gap.
The similarity score is obtained by dividing the length
of maximum matching subsequence by the minimum
of the length of two sequences. The highest value
of the similarity score is 1 if both the sequences are
same and close to 0 if they are dissimilar. The set of
all high-level strokes can be divided into two groups
based on their importance in describing a character
class, namely major and minor strokes (see Figure 2).
Therefore, the indicator function, I, in the original al-
gorithm is replaced by a customized similarity score
matrix, S
48×48
, which gives more weight to major
strokes than minor strokes while computing the simi-
larity between the HLS sequences.
5 EXPERIMENTS AND RESULTS
Two different experiments are reported in the fol-
lowing Section, namely printed character classifica-
tion and word matching, to show the effectiveness of
shape similarity measure computed using HLS repre-
sentation.
5.1 Printed Gujarati Character
Classification
The experiment for printed Gujarati character clas-
sification is performed using k-NN classifier with
shape similarity measure as a distance function to
demonstrate the applicability of shape similarity mea-
sure discussed above. The optimum value of k in
the k-NN depends on the distribution of samples in
database (Murphy, 2012). However, for any given
database, not all the classes have same distribution
(i.e. the samples of some class are denser than oth-
ers). Therefore, a single value of k may not give an
optimum result for all the classes. The current exper-
iment uses a simple heuristics to handle this issue. It
will first find all the neighbors within a tight radius
of the unknown sample (i.e. 80% similarity region in
this case) and predict the class label using majority
voting. However, if no data sample is found within
80% similarity region (i.e. the class has a sparse dis-
tribution of samples), then the 1
st
nearest neighbor is
used to predict the class label. The three-fold cross
validation technique is used to make results more au-
thentic. The average test accuracy over all three runs
is used as the primary performance measure.
The database used in the experiment consist of
12000 samples of 42 middle zone character sym-
bols from Gujarati script. The samples are col-
lected from three different sources namely, machine
printed books (BOOKDB), newspapers(NEWSDB),
and laser printed documents (LASERDB) to ensure
the varieties in terms of font type, style, size, ink
thickness, etc.
The results of the experiment are shown in Ta-
ble 1. The average test accuracy obtained on the com-
bined database is 94.97%. Table 2 shows the com-
parison of the results obtained with existing work. It
is evident that the results obtained are 2-3% lower
than the best-reported work (Hassan et al., 2014) in
the literature. The drop in the accuracy is mainly
due to the compactness of features since the size of
the feature vector in (Hassan et al., 2014) is almost
ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods
422

Figure 7: Compute shape similarity between characters using approximate string matching algorithm.
Table 1: Average test accuracy on BOOKDB, LASERDB,
and NEWSDB using k-NN classifier with shape similarity
measure.
Database 1
st
Run 2
nd
Run 3
rd
Run Average
BOOKDB 98.43% 98.77% 98.50% 98.57%
LASERDB 94.73% 94.67% 90.64% 93.35%
NEWSDB 93.51% 94.10% 89.08% 92.23%
ALL 97.72% 95.07% 92.12% 94.97%
COMBINED
Table 2: Comparison of methods for classification of
printed Gujarati characters.
Author Features/Classifier Accuracy
Hassan et al. HoG, Fringe Map 97-98%
(2014) and Shape Descriptor/
MKL-SVM
Dholakia et al. Daubechies D4 97.59%
(2009) Wavelet Feature/
GRNN
Goswami Histogram of 95.35%
and Mitra Low-Level Stroke/
(2015) K-NN
Proposed High-Level Strokes/ 94.97%
Method K-NN with Shape
Similarity measure
Goswami et al. SOM projection 84%
(2011) k-NN
Antani and Moment features 67%
Agnihotri K-NN with Minimum
(1999) Hamming Distance
100 times the size in the proposed approach. Un-
like, the features used in other character classifica-
tion methods, the HLS provides a high-level view
of the character shape and enables only approxi-
mate matching between character rather than exact.
Therefore, similar looking characters, as shown in
Figure 8, are very often misclassified. Since the
HLS features are extracted from thinned character
image, the structural noise added by thinning algo-
rithm also affects the performance. Therefore, the
accuracy reported on BOOKDB is higher(98.43%)
than LASERDB(94.73%) and NEWSDB(93.51%)
because the newspaper and the laser printed sym-
bols tend to have higher structural noise introduced
by thinning as compared to BOOKDB (Suthar et al.,
2014).
Figure 8: Characters with similar shape that are most com-
monly misclassified.
In summary, the HLS feature provides a compact
representation of the high-level shape of the char-
acter. The feature compactness allows an efficient
comparison between character shape. However, they
provide only inexact matching hence not suitable for
character classification and OCR application. Such,
approximate shape matching is desirable for word-
spotting based experiments where the objective is to
match a query word image with all morphological
variants. Moreover, also the sequential representa-
tion enables fast shape similarity computation using
dynamic programming based sequence matching al-
gorithms. Thus, the features could be useful in shape
similarity based word-matching application.
5.2 Printed Gujarati Word Matching
The idea of shape similarity computation using HLS
can easily be extended at the word-level. As shown in
Figure 9, the skew-corrected, binarized, and thinned
word-image is first segmented into character sym-
bols using connected component analysis. The HLS
are extracted from each symbols using stroke extrac-
tion algorithm (Section 3.2) and identified using CRF
High Level Shape Representation in Printed Gujarati Character
423

Figure 9: High-level stroke representation of Word Image.
(Section 3.3). The HLS sequences of all symbols are
concatenated to generate a single HLS sequence rep-
resenting the word-image. Finally, the shape simi-
larity between word-images is computed by finding
the global alignment score between the HLS sequence
representing the word-images (Section 4).
Figure 10: Shape Similarity matrix computed using HLS
representation of 280 Gujarati word images of 48 different
word groups.
The effectiveness of shape similarity measure
for word-image was verified on a small word-group
database consist of 280 word-images of 48 different
word-groups. Each word-image in the database was
represented as HLS sequence and shape similarity be-
tween each pair of word-images was computed using
shape similarity measure discussed in Section 4. The
results were represented using a similarity matrix of
size 280× 280 (shown visually in Figure 10). It is ev-
ident from Figure 10 that the HLS based shape sim-
ilarity score between the pair of word-images in the
same group is higher the one in different groups.
Figure 11: Similarity Threshold vs Precision and Recall
graph for Gujarati word-group database.
In the word image retrieval experiment, a ran-
dom query image is selected from each word group,
and all matching word images are retrieved from the
database based on the shape similarity threshold. The
value of precision and recall were computed based
on the number of relevant and retrieved images for
each query and averaged over all queries. Figure 11
shows the value of precision and recall versus simi-
larity threshold graph. The optimum values of pre-
cision and recall were 77.61% and 80.91%, respec-
tively with similarity threshold equal to 0.62.
6 CONCLUSION AND FUTURE
WORK
The paper discusses representation, extraction, and
identification of high-level strokes from printed Gu-
jarati characters. The salient characteristics of HLS
features are compactness, high-level shape descrip-
tion, easy to compute shape similarity, and extend-
ability at word-level. The features are tested on the
moderately sized symbol level database of printed
Gujarati characters with the font, size, style, and ink
thickness variations. The experiments were also per-
formed for shape-similarity based word-matching on
a small Gujarati word group database, and the ini-
tial results are encouraging. In future, more extensive
experiments can be carried out on large word image
ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods
424

database. Also, the HLS features can be combined
with other features based on the detailed analysis of
error and misclassification to improve the retrieval re-
sults.
REFERENCES
Antani, S. and Agnihotri, L. (1999). Gujarati character
recognition. In Proc. of the 5th Int. Conf. on Doc-
ument Analysisand Recognition (ICDAR’99), pages
418–421.
Aparna, K. and Ramakrishnan, A. (2002). A complete tamil
optical character recognition system. In Lopresti, D.,
Hu, J., and Kashi, R., editors, Document Analysis Sys-
tems V, pages 53–57. Springer Berlin / Heidelberg.
Bhardwaj, A., Damien, J., and Govindaraju, V. (2008).
Script independent word spotting in multilingual doc-
uments. In Proc. of 2nd Int. Workshop on Cross Lin-
gual Information Access, pages 48–54.
Charles, S. and McCallum, A. (2011). Introduction to con-
ditional random fields. Foundation and Trends in Ma-
chine Learning, 4(4):267–373.
Chaudhuri, B. and Pal, U. (1998). A complete printed
bangla ocr system. Pattern Recognition, 31(5):531–
549.
Chaudhuri, B., Pal, U., and Mitra, M. (2001). Automatic
recognition of printed oriya script. In Proc. of the 6th
Int. Conf. on Document Analysis and Recognition (IC-
DAR’01), pages 795–799. IEE.
Chaudhury, S., Sethi, G., Vyas, A., and Harit, G. (2003).
Devising interactive access techniques for indian lan-
guage document images. In Proc. of the Int. Conf. on
Document Analysis and Recognition (ICDAR), pages
885–889.
Dholakia, J., Yajnik, A., and Negi, A. (2007). Wavelet fea-
ture based confusion character sets for gujarati script.
In Proc. of the Int. Conf. on Computational Intelli-
gence and Multimedia Applications, pages 366–370.
Doermann, D. (1998). The indexing and retrieval of docu-
ment images: A survey. Computer Vision and Image
Understanding, 70(3):287–298.
Goswami, M. and Mitra, S. K. (2015). Classification of
printed gujarati characters using low-level stroke fea-
tures. ACM Trans. Asian Low-Resour. Lang. Inf. Pro-
cess., 15(4):25:1–26.
Goswami, M., Prajapati, H., and Dabhi, V. (2011). Classi-
fication of printed gujarati characters using som based
k-nearest neighbor classifier. In Proc. of the Int. Conf.
on Image Information Processing, pages 1–5. IEEE.
Hassan, E., Chaudhury, S., and Gopal, M. (2009). Shape de-
scriptor based document image indexing and symbol
recognition. In Proc. of the 10th Int. Conf. on Doc-
ument Analysis and Recognition (ICDAR’09), pages
206–210.
Hassan, E., Chaudhury, S., and Gopal, M. (2014). Feature
combination for binary pattern classification. Interna-
tional Journal of Document Analysis and Recognition
(IJDAR), 17(4):375–392.
Jawahar, C., Kumar, P., and Kiran, S. (2003). A bilingual
ocr for hindi-telugu documents and its applications. In
Proc. of the 7th Int. Conf. on Document Analysis and
Recognition (ICDAR’03), pages 408–412.
Jawahar, C. V., Balasubramanian, A., and M., M. (2004).
Word-level access to document image datasets. In
Proceedings of the workshop on computer vision,
graphics and image processing.
Kompalli, S., Setlur, S., and Govindaraju, V. (2005). Chal-
lenges in ocr of devanagari documents. In Proc. of the
8th Int. Conf. on Document Analysis and Recognition
(ICDAR’05), pages 1–5. IEEE.
Kumar, A., Jawahar, C., and Manmatha, R. (2007). Efficient
search in document image collections. In Yagi, Y.,
editor, ACCV:LNCS, volume 1 of 4843, pages 586–
595. Springer-Verlag Berlin / Heidelberg.
Lakshmi, C. and Patvardhan, C. (2002). A multi-font ocr
system for printed telugu text. In Proc. of the Lan-
gauge Engineering Conference, pages 7–17.
Lehal, G. and Singh, C. (2000). A gurmukhi script recogni-
tion system. In Proc. of the 15th Int. Conf. on Pattern
Recognition (ICPR’00), pages 557–560.
Meshesha, M. and Jawahar, C. (2008). Matching of word
image for content-based retrieval from printed doc-
ument images. International Journal of Document
Analysis and Recognition (IJDAR), 11(1):29–38.
Murphy, K. (2012). Machine Learning: A Probabilistic Per-
spective. The MIT Press, Cambridge, Massachusetts
London, England.
Needleman, S. B. and Wunsch, C. D. (1970). A gen-
eral method applicable to the search for similarities
in the amino acid sequence of two proteins. Journal
of Molecular Biology, 48(3):443–453.
Rath, T. and Manmatha, R. (2003). Word image match-
ing using dynamic time wrapping. In Proc. of the
Int. Conf. on Computer Vision and Pattern Recogni-
tion (ICVRP), volume 2, pages 521–527.
Srihari, S., Srinivasan, H., Huang, C., and Shetty, S. (2006).
Spotting words in latin, devanagari and arabic scripts.
Vivek, 16(3):2–9.
Suthar, S., Goswami, M., and Thakkar, A. (2014). Empir-
ical study of thinning algorithms on printed gujarati
characters and handwritten numerals. In Meenakshi,
N., editor, Proc. of the 2nd Int. Conf. on Emerging Re-
search in Computing, Information, Communication,
and Applications (ERCICA’14), volume 2, pages 104–
110. ELSEVIER.
Tarafdar, A., Mondal, R., Pal, S., Pal, U., and Kimura, F.
(2010). Shape code based word-image matching for
retrieval of indian multi-lingual documents. In Proc.
of the Int. Conf. on Pattern Recognition (ICPR), pages
1989–1992.
Yang, M., Kpalma, K., and Ronsin, J. (2008). A survey of
shape feature extraction techniques. In Yin, P., editor,
Pattern Recognition, pages 43–90. IN-TECH.
High Level Shape Representation in Printed Gujarati Character
425