An Empirical Intrinsic mode based characterization of
Indian Scripts
Kavita Bhardwaj
∗
Indian Institute of Technology
New Delhi,India
kavitab788@gmail.com
Santanu Chaudhury
Indian Institute of Technology
New Delhi,India
schaudhury@gmail.com
Sumantra Dutta Roy
Indian Institute of Technology
New Delhi,India
sumantra.dutta.roy@gmail.com
ABSTRACT
In this paper, we describe a novel technique for Document script
identification(DSI) from printed documents, using Empirical Mode
Decomposition (EMD). The intrinsic decomposition nature can adap-
tively decompose script images into a series of modes representing
different local features of script images. In this method, Radon
transformed script images are decomposed into finite set of IMFs
(Intrinsic Mode Functions). The energy concentration in a particu-
lar orientation characterises a script texture as it indicates the domi-
nance of individual script in that direction. We demonstrate how the
proposed method use these IMFs as feature vectors to distinguish
various scripts.
Keywords:
Empirical mode decomposition(EMD), Radon transform, Intrinsic
mode function, AdaBoostM1
1. INTRODUCTION
The Identification of the script used in printed documents is use-
ful for the digitization of the conventional paper documents, sort-
ing of document images according to the scripts in which they are
written, for selecting appropriate script-specific OCRs for the re-
trieval of online archives of document images or for indexing of
documents in digital library.
Ghosh et al. [1] proposes the categorisation of script recogni-
tion methods as structure-based and visual appearance-based. He
discussed the methods of both categories at page-level, paragraph-
level, word-level and character-level. A vast survey is presented
for each of the categories. By reffering Wang et al. [9] and Ghosh
et al. [1], it is found that according to the feature extraction, all
the methods lying under any category are grouped into three major
categories- Statistical-information based methods, Structure-based
methods, Texture-based methods. Statistical information-based al-
gorithms use character density distribution, vertical and horizontal
projections, for classifying printed documents. Waked et al. [8]
used bounding box size distribution, character density distribution,
∗
Corresponding author
Permission to make digital or hard copies of part or all of this work for
personal or classroom use is granted without fee provided that copies are not
made or distributed for profit or commercial advantage and that copies bear
this notice and the full citation on the first page. Copyrights for components
of this work owned by others than ACM must be honored. Abstracting with
credit is permitted. To copy otherwise, to republish, to post on servers or to
redistribute to lists, requires prior specific permission and/or a fee.
DAR ’12, December 16, 2012, Mumbai, IN, India
Copyright 2012 ACM 978-1-4503-1797-9/12/12 ...$15.00.
vertical and horizontal projections for the classification of printed
documents. Lam et al. [5] has also used statistical features for
script identification in printed-documents. These methods are more
useful for scripts that differ significantly in style. Structure-based
methods focus on extraction and analysis of connected components
and use the identification results of these â
˘
AIJsignaturesâ
˘
A
˙
I to de-
termine the script(s) used. These methods in general have advan-
tage of discriminating similar scripts. Hochberg et al. [2] exploited
the shape characteristics of "textual symbols" for the identification
of script(s). Pal and Chaudhuri [6], presented the script charac-
teristics and shape based features for script identification. Visual-
appearance and texture analysis-based methods are related, because
according to appearance of any text block, corresponding texture
analysis-based method can be used for extraction of features. Joshi
et al. [4] proposes the Gabor function-based texture analysis to
extract features and used hierarchical classification to distinguish
among the script(s). Tan [7] developed Gabor function-based tex-
ture analysis for machine-printed script identification that discrim-
inates Chinese, Latin, Greek, Russian, Persian, and Malayalam
script documents.
For our problem, we propose an algorithm based on Empirical
Mode Decomposition(EMD) for textural analysis of script classes.
The directionality and periodicity reflect the effective directions for
textural processing of subpatterns. Each script class will always ex-
hibit a specific periodicity at a particular angular orientation. This
is observed in Radon transformed image of the script classes con-
sidered for the problem and they are decomposed in different mode
functions to compute directional energy specified by each IMF.
The scripts involved in this paper are Devnagari, Roman(English),
Malayalam, Bangla and Gurumukhi. The cosine similarity measure
is used as our measure to define the most similar script class for the
descrimination. We use Adaboost binary decision tree to improve
the classification.
The rest of the paper is organized as follows. In (Sec. 2), we sum-
marize the proposed approach and the framework for the problem is
described in (subsec. 2.1). The results are shown through Table 3in
(Sec. 3). Experimental observation are described in (Sec. 4). Con-
clusions and future work are discussed in (Sec. 5).
2. THE PROPOSED APPROACH
The method described in this paper involves four main steps.
• First, the preprocessing is performed initially to remove noise
that includes binarization of document images.
• Second, Radon transform is computed on document images
of each script at different angles of orientation between 0
◦
to 90
◦
. The unique characteristic of each script is observed
at a particular orientation in radon transformed image. The
Figure 1: Schematic Block Diagram of the System
transformed image is decomposed using empirical mode de-
composition(EMD) in a finite set of IMFs.
• Third, the energy is computed corresponding to each Intrin-
sic Mode Function(IMF) at different orientations. The en-
ergy at a particular orientation characterises a script class tex-
ture as it indicates the dominance of individual script class in
that orientation.
• This angle of orientation and the IMF exhibiting maximum
energy helps to chose a distinguishing feature vector for in-
dividual script.
• Fourth, the feature vectors for test script images are obtained
similarly and then classified using AdaBoost binary decision
tree.
Figure 1 shows a schematic diagram of the system. We will dis-
cuss each of the modules in detail in the upcoming sections of this
paper.
2.1 A Generalised Framework to Script Iden-
tification
We propose a general framework to address the problem of script
identification. This proposed scheme (for document script iden-
tifcation exploits the textural characteristic of each script. The
constructing patterns of each script are formed by oriented lin-
ear/curvilinear subpatterns. The energy distributed at different ori-
entations, characterises each script according to the textural char-
acteristics of each script. For instance, the devnagari script is char-
acterised by the dominance of horizontal lines, whereas Malay-
alam script is characterised by the dominance of vertical lines and
curves. While Roman(English) script is a good mix of linear and
curved subpatterns. On the other hand, energy is distributed more
or less evenly amongst different orientations for scripts which have
curved patterns, like Malayalam and Gurumukhi. Here, in the pro-
posed framework, features are learned for each script from prior
knowledge(from the training data set) of target script classes. These
extracted features contain finer and sufficient discriminating details
of the scripts. A script class is separated from other script class by
exploiting its unique features.
2.2 Feature Extraction
The energy based features are used corresponding to each script
class. We use Empirical mode decomposition to capture the ori-
ented local energy . As discussed earlier, the constructing subpat-
terns of each script are composed of linear/ curvilinear subpatterns.
Therefore any script class can be distinguished from other based on
the structural subpattern behavior.
From the prior observation(having the knowledge about struc-
tural subpatterns) of each script, we have empirically determined
the angular orientation.
We have computed the oriented local energy features for the dis-
crimination between script class. Also in Table 2, we have defined
the average energy and the variance corresponding to each script
class. While testing a document or defining a script class to a doc-
ument, the energy corresponding to the test document is compared
with the average energy corresponding to each class. The variance
for each script class is also computed. The projections are taken us-
ing Radon transform at different orientations for each script class
between angle 0
◦
to 90
◦
. The transformed script images are de-
composed using EMD(Empirical mode decomposition) to analyze
local characteristics and a finite set of IMF(s) are obtained. We
have considered first four IMF(s) in our experiments.Since, all the
Intrinsic mode functions(IMF) obtained does not contain sufficient
energy in that orientaion and can not be used as discriminating fea-
ture. We have empirically determined the IMF in which the en-
ergy distribution is maximum for that script class and moreover the
script class is distinguishable from other script class. The angu-
lar orientation at which Radon transform is computed and the IMF
compositly represents the distinguishing feature vector for each
script class.
Table 1 shows the feature vectors selected for the script classes
considered for our problem.
Features Extracted for each script class
Feat_vector Script class angle of orient. IMF
FV1 Devnagari 90
◦
IMF1
FV2 Malayalam 90
◦
IMF3
FV3 Gurumukhi 0 − 10
◦
IMF1
FV4 Roman 0
◦
IMF4
FV5 Bangla 90
◦
IMF2
Table 1: Feature vectors defined for script classes
The average energy and the variance corresponding to each fea-
ture vector is shown below in Table 2.
Average energy and Variance for each Feat_vect
Feat_vector Average energy Variance
FV1 0.0338 6.8798 exp −04
FV2 0.0243 2.5480e-04
FV3 0.0633 0.0057
FV4 0.0437 0.0028
FV5 0.0479 8.1585 exp −04
Table 2: Average energy and Variance
2.3 Feature Selection
There are many potential benefits to feature selection like facili-
tating data visualization, data understanding, reducing training and
utilization times and improving accuracy. Feature selection is se-
lecting the most relevant variables, is usually suboptimal for build-
ing a predictor, particularly if the variables are redundant. Here,
we have used the feature selection for selecting appropriate and
relevant features for the problem at hand. The angle of orientation
chosen for computing radon transform is totally dependent on data
visualization and understanding of data. We have empirically com-
puted the angular orientation during training by having knowledge
and understanding of a script class.
Next, as aforementioned, EMD method decomposes a signal into
a set of components called IMF. But all the IMF(s) are not relevant
as they all are not informative. So we have to choose only the IMF
that contain useful information(maximum energy) and discarding
those that share similar amounts of energy. We have not used any
standard method for selection of these features, as they are not help-
ful for our problem. Rather we have empirically found the angular
orientation and IMF which can be used a distinguishing feature for
individual script class.
2.4 Classifier Design
For our problem, we use to define the most similar script class
to the training images as well as to testing document. We use Ad-
aboost binary decision tree as a classifier to improve the results.
This involves training of the classifier and testing the new docu-
ment for different script classes.
• Train the classifier: The system extracts the textural feature
of different script classes and they depend on the charac-
teristics of specific scripts. we have performed the training
based on the distance of a training sample to each class. Dur-
ing training, we extract the above discussed features for our
dataset and use cosine similarity measure as our measure of
similarity for the discrimination. This similarity measure can
be computed amongst arbitrary vectors.
• Classifying a new document: Testing a new document, we
compute the features in the same way as described above.
Cosine similarity measure is used for the comparison with
other script classes in the dataset and the most similar script
class is assigned to the test document. To improve the per-
formance of classification, we use Adaboost binary decision
tree.
2.5 Empirical Mode Decomposition
In this paper, we utilize the aspect of decomposing a signal into
IMFs for analyzing nonstationary and nonlinear time series data
developed by Huang et al. [3]. The intrinsic decomposition nature
can adaptively decompose images into a series of modes represent-
ing different local features of images. The decomposition is based
on the local characteristic of time scale of the data.
• Pre-requisites
– the number of external and the number of zero crossing
must either equal or differ at most by one;
– should be symmetric with respect to local zero mean.
With these two prerequisites, an IMF can be represented with
a meaningful instantaneous frequency.
• Sifting Procedure
From a given signal x(t), we extract IMFs using sifting pro-
cess satisfying the above defined conditions.
1. Identify all extremas
2. interpolate between minima with respective maxima end-
ing up with L
min
(t) respective L
max
(t)
3. compute the mean m(t) = (L
min
(t) + L
max
(t))/2
4. extract the details d(t) = x(t) − m(t)
5. iterate the same process on the residual m(t)
Once this is achieved, the detail is referred to as an Intrin-
sic Mode Function (IMF), the corresponding residual m(t) is
computed and step 5 applies.
• Stopping Criterian
If we go for sifting beyond a limit, we will get IMFs as fre-
quency modulated signal, but constant in amplitude. So stan-
dard deviation is computed as stopping equation to stop sift-
ing. The equation is:
x(t) =
m
X
i=1
IMF
i
+ r
m
(t) (1)
where m is the number of IMFs obtained for a given signal
and r
m
(t) is final residual.
3. RESULTS AND DISCUSSION
We have used 300 samples of each script class and a total of 1500
images database have been used for out experiments. These experi-
ments have been done vice a versa for each pair of script class clas-
sification. We have performed the experiments on training and test-
ing dataset vice a versa. So the accuracy of the experiments shown
below is the average accuracy. As far as the feature extracted for
script classes under consideration is concerned, it depends on the
ratio of linear,curvical and curvilinear textural behavior of the sub-
patterns. For example the ratio of curvical and linear subpatterns
are more in Devnagari and Roman than Gurumukhi and Bangla.
The script classes Gurumukhi and Bangla contains more cuvilinear
subpatterns than linear. While if Malyalam is considered, it con-
tains more cuvical subpatterns than Gurumukhi and Bangla even.
So the angle of orientation chosen strictly depends on ratio of the
structural subpattern behavior. While selecteing the feature vector
for any script class the structural behavior of must be analyzed.
Table 3 shows the average classification accuracy evaluated of
all the script classes.
Classification accuracy achieved
Script Script classes
Dev Mal Guru Roman Bangla
Devnagari 97% 3% - - -
Malyalam 4% 96% - - -
Devnagari 97% - - 3% -
Roman 5% - - 95% -
Malayalam - 94% - 6% -
Roman - 3% - 97% -
Devnagari 92% - - 8% -
Gurumukhi 10% - 90% - -
Bangla - - 4% - 96%
Gurumukhi - - 92% - 8%
4. EXPERIMENTAL OBSERVATION
We have applied the proposed script identification(classification)
method on OCR database document images. We have used for our
experimentation 300 documents of each script, so dataset is com-
posed of total 1500 document images. 50% of the images are used
for training and rest 50% for testing of each script. We have used
512 by 512 document images for extracting features. The average
number of characters in 512 by 512 document image is approx.
1500 to 2000. The proposed method gives better performance on
document image with minimum 256 by 256 size , because it can
capture sufficient distribution of pixel intensity according to strokes
of each script class. We have done the experimentation on inter-
changing the traing and testing dataset. So the performance shown
above in (Sec. 3) is the average performance of both experiments
done on different document images of the dataset.
5. CONCLUSION AND FUTURE WORK
This method is developed for identification(classification) of mul-
tiplt script classes individually. The strong potential of the pre-
sented work here is its application and accuracy. The EMD(Empirical
Based Decomposition) based approach has been applied in vari-
ous applications but here is the first of its kind for script identifica-
tion. The proposed method is computationally less time consum-
ing, hence for script specific identification applications, it will be
highly effective as compared to other expensive feature extraction
and classification methods. Also the performance of the method
is reasonably high. In future, We can extend this work for more
number of scripts and even for different script classes. We can also
apply it for text/image separation. As the energy distribution of im-
ages are evenly scattered at diiferent mode functions but in text it
will always be high and oriented than images. Also the proposed
method can be extended to multiple scales for script identification.
Acknowledgment
The authors are grateful to Prof. S.D Joshi, Dept. of Electrical
Engg., IIT, Delhi for their helpful discussion and encouragement
during this work.
6. REFERENCES
[1] D. Ghosh, T. Dube, and S. A.P. Script Recognition - A
Review. IEEE Trans. Pattern Analysis and Machine
Intelligence, 32:2142 – 2161, 2010.
[2] L. Hochberg, L. Kerns, P. Kelly, and T. Thomas. Automatic
Script Identification from images using Cluster-based
Templates. IEEE Trans. Pattern Analysis and Machine
Intelligence, 19:176–181, 1997.
[3] H. Huang and J. Pan. Speech pitch determination based on
hilbert-huang transform. Signal Processing.
[4] G. Joshi, S. Garg, and J. Sivaswamy. A generalized framework
for script identification. Int. J. on Document Analysis and
Recognition, page 55â
˘
A¸S68, 2007.
[5] L. LAM, J. DING, and C. SUEN. Differentiating between
Oriental and European Scripts by Statistical Features.
International Journal of Pattern Recognition and Artificial
Intelligence, pages 63–79, 1998.
[6] U. Pal and B. Chaudhuri. Script Line Separation from Indian
Multi-Script Documents. In Int. Conf. Document Analysis and
Recognition, pages 406–409, 1999.
[7] T. Tan. Rotation invariant texture features and their use in
automatic script identification. IEEE Trans. Pattern Analysis
and Machine Intelligence, pages 751–756, 1998.
[8] B. Waked, S. Bergler, C. Suen, S. Khoury, and C. Y. S. S.
Khoury. Skew Detection, page segmentation, and script
classification of printed document images. pages 4470–4475,
1998.
[9] N. Wang, L. Lam, and C. Y. Suen. Noise tolerant script
identification of printed oriental and english documents using
a downgraded pixel density feature. Pattern Recognition,
International Conference on, pages 2037–2040, 2010.
