Zernike Moment Feature Extraction for Handwritten
Devanagari (Marathi) Compound Character
Recognition
Karbhari V. Kale, Senior Member, IEEE
∗
, Prapti D. Deshmukh
†
, Shriniwas V. Chavan, Student Member, IEEE
‡
,
Majharoddin M. Kazi, Student Member, IEEE
§
, Yogesh S. Rode, Student Member, IEEE
¶
∗‡§
Department of Computer Science and IT, Dr. B. A. M. University, Aurangabad, Maharashtra, India - 431004
†
Dr. G. Y. Pathrikar College of Computer Science, MGM, Maharashtra, India - 431005
‡
Department of Computer Science, MSS’s ACS College, Ambad, Jalna, Maharashtra, India - 431203
∗
Email: kvkale91@gmail.com,
†
prapti.research@gmail.com,
‡
shripc@gmail.com,
§
mazhar940@gmail.com,
¶
ys.rode@gmail.com
Abstract—Compound character recognition of Devanagari
script is one of the challenging tasks since the characters are com-
plex in structure and can be modified by writing combination of
two or more characters. These compound characters occurs 12 to
15% in the Devanagari Script. The moment based techniques are
being successfully applied to several image processing problems
and represents a fundamental tool to generate feature descriptors
where the Zernike moment technique has a rotation invariance
property which found to be desirable for handwritten character
recognition. This paper discusses extraction of features from
handwritten compound characters using Zernike moment feature
descriptor and proposes SVM and k-NN based classification sys-
tem. The proposed classification system preprocess and normalize
the 27000 handwritten character images into 30x30 pixels images
and divides them into zones. The pre-classification produces three
classes depending on presence or absence of vertical bar. Further
Zernike moment feature extraction is performed on each zone.
The overall recognition rate of proposed system using SVM and
k-NN classifier is upto 98.37%, and 95.82% respectively.
Keywords—Handwritten Character, Devanagari Compound,
Zernike, SVM, k-NN.
I. INTRODUCTION
Handwritten character recognition is gaining popularity
for many years and attracting researchers for the purpose of
potential application development. These potential applications
reduce the cost of human efforts and save the time. Some
of its potential application areas are like bank automation,
postal automation [1]–[3] etc. Similarly the biometric and
criminal identification system uses scanned handwritten script
for forensic and Historic Document Analysis (HDA) and
represents an excellent study area within the research field of
biometrics and forensic science.
The technical challenge in handwritten character recogni-
tion comes from three sources: Symbol: an ideal shape that
occurs in hierarchy and symbol are arranged in complex form
at different level in organization. Deformation: shape variation
in each symbol to undergoes geometric transformation (transla-
tion, rotation, scaling, stretching) and complex representation.
Defect flaw in image owing to print, scan, quantized, binary
etc. Handwritten and Printing character demands diverse ap-
proach, handwritten consist of extended stroke and printed
consist of normal shaped blobs.
Research in handwritten character recognition focuses on
two main approaches i.e. on-line and off-line. In on-line char-
acter recognition system captures data by the sensors during
writing process, which makes the information dynamically
available according to the strokes. While, off-line character
recognition takes place in static form where images are cap-
tured or scanned after completion of the writing process on
paper/sheets. Both the tasks are challenging for automatic
character recognition, specifically in off-line character recog-
nition requires more efforts due to various reasons viz. large
variations in shape of characters due to pen ink, pen width,
and accuracy of devices, stroke size and location, effect of
physical and mental situation of the writer on writing style, in
turns effect the recognition accuracy.
Character recognition problem becomes more challenging
even in on-line and off-line in Indian Language Scripts due
to several reasons [4]. The Indian scripts have character set
with large number of characters. The shape of the characters
in Indian scripts is more complex and may have modifiers.
These modifiers may found at above, below or in-line with
the character. The modifiers are the vowel that changes their
shapes when they get connected with the consonants. The
scripts may have some character pairs that are looks alike and
cause difficulty in classification. Some of Indian languages
like Devanagari, Bangla are having the specific problem in
compound characters where two or more consonants join with
each other to form a special character [5], [6].
The research work on character recognition of Devanagari
script was started in 1970, where Sinha and Mahabala [7] were
presented a syntactic pattern analysis system for the recogni-
tion of Devanagari characters (DC). First research report on
handwritten Devanagari Characters (HDC) was published in
1977 by Sethi and Chatterjee [8], very few work were reported
on OCR in the literature and later on in the next decade
S. Kumar and et. al. contributed more in this domain [9].
An extensive research work on printed Devanagari Characters
and Handwritten Characters was carried out by Bansal [10]–
[12] and Reena et.al, [13], [14] respectively. Recognition
of characters in different languages using Zernike Moments
was reported in [9], [15]–[22]. Researchers have proposed
Chain Code Histogram and directional information gradient
based feature extraction in [22]–[24]. A significant contribution
(IJARAI) International Journal of Advanced Research in Artificial Intelligence,
Vol. 3, No.1, 2014
68 |
P a g e
www.ijarai.thesai.org
by Arora and et. al., proposed feature extraction techniques
namely, intersection, shadow feature, chain code histogram
and straight line fitting features in [25]–[28]. Deshpande and
et. al. [29] has proposed fine classification and recognition
of Devanagari characters. S. Kumar in [30] also extracted
various features and performed comparison using SVM and
MLP. Pal and et al. proposed SVM and MQDF based scheme
for recognition of Devanagari Characters [31]. U. Pal and
T. Wakabayashi [32] given a comparative study of differ-
ent Devanagari Character recognizers which extracts features
based on curvature and gradient information. Sushama Shelke
and et. al. [33] presented a novel approach for recognition
of unconstrained handwritten Marathi characters. Baheti M.J.
and et al. [34] proposed a method based on Affine Invariant
Moment (AIM) for Gujarati numerals using k-NN and PCA
classifiers. Elastic matching (EM) technique based on an Eigen
Deformation (ED) for recognition of handwritten Devanagari
characters is proposed by V. Mane and et.al, [35]. Recognition
of handwritten Bangla compound characters was attempted by
U. Pal and et al. [36] using gradient features. S. Shelke and S.
Apte have reported work on handwritten Marathi compound
characters using multi-stage multi-feature classifier [5], [6].
The literature evidence shows that moment can be con-
sidered as potential features for recognition of characters and
numerals, which motivate us to enrich the several orthogonal
and discrete moment features and test the efficacy of the
system for compound characters. While significant advances
have been achieved in recognizing Roman-based scripts like
English, ideographic characters Chinese, Japanese, Korean,
and Arabic, only few works on some of the major Indian
scripts like Devanagari, Bangla, Gurumukhi, Tamil, Telugu,
are available in the literature [37]–[41].
This paper proposes a novel Zernike moment based fea-
ture descriptor followed by SVM and k-NN neural network
approach for recognition of Marathi Script Basic and Com-
pound Characters derived from Devanagari. The organization
of the paper is as follows: Section 2 deals with properties
of Devanagari derived Marathi script. Database designing and
Proposed System has been discussed in Section 3. Section 4
deals with Zernike Moment based feature extraction technique.
Details about the SVM and k-NN approach used for character
recognition system are elaborated in Section 5. The experi-
mental results are discussed in Section 6. Conclusion of the
paper is given in Section 7.
II. PROPERTIES OF DEVANAGARI BASIC AND COMPOUND
CHARACTER
The basic set of symbols of Devanagari script consists
of 12 vowels (or swar), 36 consonants (or vyanjan). The
alphabet of modern Devanagari script consists of 14 vowels
and 33 consonants also called as basic characters. Writing
style of the Devanagari script is from left to right and the
concept of upper and lower case is absent in the script. In
this script vowel following by a consonant takes a modified
shape, these modified shapes are called modified characters. A
consonant or vowel following a consonant sometime takes a
compound orthographic shape, which we called as a compound
character. Compound characters can be combination of two
consonants as well as a consonant and a vowel. The compound
characters are joined in various ways, by removing vertical
line of the character and then to the other characters from the
left side like My, in another way it is joined side by side
or one above the other like Ó. The example of compound
characters is shown in Fig (1). The split character is half of the
basic character which gets connected to other characters. The
example of split component of compound character is shown
in Fig (2). Compounding of three or four characters also exists
in the script. There are about 280 compound characters in the
Devanagari script [4], [31].
Figure 1. Sample images of Compound Character
Figure 2. Sample images of Split Component of Compound Character
Marathi script is one of the derived script from Devanagari,
and it is an official language of Maharashtra. Marathi script
consists of 16 vowels and 36 consonants making 52 alphabets.
Marathi script is written from left to right, which does not
have upper and lower case characters. Similar to Devanagari it
has nearly the similar type of compound characters property.
However, the occurrence of compound characters in Marathi
is found to be about 11 to 12%, whereas in other scripts of
Devanagari, it is about 5 to 7% [42].
III. DATABASE DESIGNING AND PR OPOSED SYSTEM
A. Database
At present no dataset of handwritten compound charac-
ters is available for Marathi script derived from Devanagari
and hence we have created handwritten compound characters
dataset for this work and it has been tested with our proposed
system for its recognition, this adds a new contribution in the
literature. Details of this database are provided in Table I.
The database of Handwritten Characters of Marathi from
Devanagari script is created for the purpose of this work, which
contains basic, compound and split components of compound
characters. These data characters were recorded in written
form on special paper sheet from 250 different volunteers of
different age group. (in between 20-40 year old). The recorded
character is then scanned with Flatbed Scanner at 300 dpi. The
size of the image of each character is considered 90x90 pixels
and it is stored in TIFF image format.
Table I. DATASET OF HANDWRITTEN DEVANAGARI BASIC AND
COMPOUND CHARACTE RS
Property Descriptions
Number of subjects 250
Number of basic character 48
Number of compound charac-
ters
45
Number of split compound
characters
15
Number of images given by
each subject
48+45+15=108
Gray/Color Color
Resolution 90*90 pixels
DPI 300
Format TIFF
Total Number of images 27000
(IJARAI) International Journal of Advanced Research in Artificial Intelligence,
Vol. 3, No.1, 2014
69 | P a g e
www.ijarai.thesai.org
B. Proposed System
In the proposed system, we aim at recognizing handwritten
Marathi Devanagari compound characters. This is done by
employing Zernike moment feature extraction using SVM and
k-NN neural network approach. Fig. (3) shows the basic block
diagram of the proposed recognition system, which consists
of different phases begin with input character images, prepro-
cessing, pre-classification of the characters, Zernike moment
based feature extraction and character recognition. The brief
phase wise explanation of the recognition system as follows:
Fig. (3) shows the basic block diagram of the recognition
system. It shows that the handwritten Devanagari character
are scanned and a digitized document is obtained. From it a
particular character is selected, the image character is cropped
and resized into fix row and columns. Each block of the
recognition system is elaborated in following sections.
Read Character
Preprocessing
Structural
Classification
VMBVEB NB
ZM ZM ZM
SVM KNN SVM KNN SVM KNN
Recognition
of Character
Figure 3. Block Diagram of Proposed Character Recognition System
C. Preprocessing
Pre-processing step is performed on the character image to
remove the noise from it and also to minimize the variations
in character styles. Occasionally, the document while scanning
was not clean and so it has produced small dots in the images.
Noise generated by the shaded areas and dots must be filtered
during preprocessing step. Moreover, the characters in scanned
images may found to be skewed, slant and varied in sizes due
to cropping. This has been processed in this step. The typical
flow of preprocessing step is shown in following Fig (4).
1) RGB to Gray Image: The database contains color
character images. In preprocessing the character images are
converted to binary images using rgb2gray utility in MATLAB.
2) Thresholding: This preprocessing step also termed as
binarization process and converts the pixels that are above the
threshold to white and those which are below the threshold to
black. We have set the threshold value Th= 190 to produce
good quality binarized images.
Scanned
Image
RGB
to Gray
Threshold or
Binarization
Filtering
Boundary
Tracing
NormalizationSkeletonization
Figure 4. Steps in Preprocessing of Image
3) Filtering: To remove the noise present in the binarized
image filtering has been done. We have used Median filter to
remove small black spot in the image and the black shade
appearing at the edges. Further, the documents were cropped
from the edges. Thresholding and filtering steps often resulted
in some broken characters. To rejoin the broken characters,
image dilation operation on the filtered images has performed.
4) Boundary Tracing: Tracing of the boundary identifies
the connected components of the characters in the filtered im-
ages and stores it in array. To find the connected components,
the algorithm starts by traversing the rows of filtered image.
It searches for a foreground pixel, and then it marks that pixel
and picks it. Similarly, marking of all the neighbors of found
pixel in all search directions completed till all the pixels of the
possible character have been traversed and marked. Otherwise,
it will continue the search in the next row. If the size of
any picked connected component is too small than the actual
required size, then the algorithm treats that component as noise
and neglects that component.
5) Normalization: During normalization step, slant in char-
acters is removed and resized to a window. Slant is the average
divergence of the vertical strokes of the character from the right
side of the character. To remove the slant, we used imrotate
with angle θ. At each angle the sum of vertical projection of the
transformed characteris calculated. The angle with maximum
sum of vertical projection is used to finally perform shear
transformation on the character and estimated the slant angle.
6) Skeletonization: In skeletonization, the thickness of the
character is reduced to one-pixel character bound. We have
applied the thinning operation on the character and taken the
precaution, do not to break the character. These operations
were used not only to find the vertical bar and position of ver-
tical bar in the character, but also to extract endpoints, junction
in the character. This features helps in the pre-classification of
the characters. A sample output of the preprocessed character
к is shown in Fig. (5).
D. Pre-classification
Character Images after preprocessing stage consists of
some global and local features. The global feature consists
of presence of vertical line, position of vertical bar in the
(IJARAI) International Journal of Advanced Research in Artificial Intelligence,
Vol. 3, No.1, 2014
70 |
P a g e
www.ijarai.thesai.org
Figure 5. Character Images after Preprocessing
character and enclosed region in the character. The local
features consists of the end points and junction position in
the character. On the basis of global feature, the character is
classified into three major categories based on the presence of
vertical bar i.e. a) character with vertical bar at right (VEB:
Vertical End Bar), b) character with vertical bar at middle
(VMB: Vertical Mid Bar), and c) the character with absence
of vertical bar (NB: No Bar). Vertical bar at right are further
classified into two categories based on whether the vertical
bar and rest of the character are connected or not to the
bar. These pre-classification of characters are shown in the
following Table II and III.
Table II. CLASSIFICATION OF DEVANAGA RI BASIC CHARACTER
Sr. No. Pre Classification Character
1 Character connected vertical
bar at right side
х, G, c, я, J, , t, T, d,
D, n, p, b, B, m, y, r, l,
v, s, q, ", j
2 Character not connected with
vertical bar at right side
g, Z, ш
3 Character with absence of ver-
tical bar
R, C, V, W, X, Y, d, r, h,
Table III. CLASSIFICATION OF DEVANAGA RI COMPOUND CHARACTER
Sr. No. Pre Classification Character
1 Character connected vertical
bar at right side
=y, >y, y, (y, @y, <y,
Sy, &y, ]y, My, Ny, y, Qy,
-y, -m, m, >m, (m, -p,
Sp, ?y, ?l, %y, `y, y,
[v, >v, -v, <v, Mn, ND, =t,
Nя, Hх, (t, n
2 Character not connected with
vertical bar at right side
-к,(к,-P
3 Character with absence of ver-
tical bar
@d, QC, Mh, h, ¤, &d, Nd
E. Local Structural Classification
The local features are detected on the basis of the end
points of the character. We have firstly partitioned the character
into 3x3 image i.e. 9 quadrants and extracted the end points
and junctions in each individual block as shown in Fig. (6).
Figure 6. Presence of End Points in partition block of Character
Thus pre-classification of character is done and put the
character in proper class like VEB, VMB and NB and then
Zernike moment features are extracted for its final classifica-
tion under SVM and k-NN.
IV. ZERNIKE MOMENT BASED FEATURE EXTRACTION
Zernike moments are complex number by which an image
is mapped on to a set of two-dimensional complex Zernike
polynomials. The magnitude of Zernike moments is used as
a rotation invariant feature to represent a character image
patterns [43]. Zernike moments are a class of orthogonal
moments and have been shown effective in terms of image
representation. The orthogonal property of Zernike polynomi-
als enables the contribution of each moment to be unique and
independent of information in an image. A Zernike moment
does the mapping of an image onto a set of complex Zernike
polynomials. These Zernike polynomials are orthogonal to
each other and have characteristics to represent data with no
redundancy and able to handle overlapping of information be-
tween the moments [26]. Due to these characteristics, Zernike
moments have been utilized as feature sets in applications such
as pattern recognition [27] and content-based image retrieval
[28]. These specific aspects and properties of Zernike moment
are supposed to found to extract the features of compound
handwritten characters. Teague [16] has introduced the use of
Zernike moments to overcome the shortcomings of information
redundancy due to geometric moments.
The Zernike moment were first proposed in 1934 by
Zernike [44]. Their moment formulation appears to be one of
the most popular, outperforming the alternatives [45] (in terms
of noise resilience, information redundancy and reconstruction
capability). Complex Zernike moments [46] are constructed
using a set of complex polynomials which form a complete
orthogonal basis set defined on the unit disc (x
2
+y
2
) ≤ 1.They
are expressed as A
pq
. Two dimensional Zernike moments:
A
mn
=
m + 1
π
Z
x
Z
y
f(x, y)[V
mn
(x, y)]
∗
dx dy
where x
2
+ y
2
≤ 1 ▽
y
(x
i
+ h, y
j
+ k)
(1)
where m = 0, 1, 2, ..., ∞ and defines the order, f (x, y)
is the function being described and * denotes the complex
conjugate. While n is an integer (that can be positive or
negative) depicting the angular dependence, or rotation, subject
to the conditions:
m − |n| = even, |n| ≤ m (2)
and A
∗
mn
= A
m,−n
is true. The Zernike polynomials [20]
V
mn
(x, y)V
mn
(x, y) Zernike polynomial expressed in polar
coordinates are:
V
mn
(r, θ) = R
mn
(r)exp(jnθ) (3)
where (r, θ) are defined over the unit disc, j =
√
−1 and
R
mn
(r) and is the orthogonal radial polynomial, defined as
R
mn
(r) Orthogonal radial polynomial:
R
mn
(r) =
m−|n|
2
X
s=0
(−1)
s
F (m, n, s, r) (4)
(IJARAI) International Journal of Advanced Research in Artificial Intelligence,
Vol. 3, No.1, 2014
71 |
P a g e
www.ijarai.thesai.org
where:
F (m, n, s, r) =
(m − s)!
s!
m+|n|
2
− s
!
m−|n|
2
− s
!
r
m−2s
(5)
where R
mn
(r) = R
m,−n
(r) and it must be noted that if
the conditions in Eq. 2 are not met, then R
mn
(r) = 0. The
first six orthogonal radial polynomials are:
R
00
(r) = 1 R
11
(r) = r
R
20
(r) = 2r
2
− 1 R
22
(r) = r
2
R
31
(r) = 3r
3
− 2r R
33
(r) = r
3
(6)
So for a discrete image, if P
xy
is the current pixel then Eq.
(1) becomes:
A
mn
=
(m + 1)
π
X
x
X
y
P
xy
[V
mn
(x, y)]
∗
where x
2
+ y
2
≤ 1 (7)
To calculate the Zernike moments, the image (or region of
interest) is first mapped to the unit disc using polar coordinates,
where the centre of the image is the origin of the unit disc.
Those pixels falling outside the unit disc are not used in the
calculation. The coordinates are then described by the length
of the vector from the origin to the coordinate point, r, and
the angle from the x axis to the vector r.r Polar co-ordinate
radius, θ.θ Polar co-ordinate angle, by convention measured
from the positive x axis in a counter clockwise direction. The
mapping from Cartesian to polar coordinates is:
x = r cos θ y = r sin θ (8)
where,
r =
p
x
2
+ y
2
θ = tan
−1
y
x
(9)
However, tan
−1
in practice is often defined over the interval
x
2
≤ θ ≤
x
2
, so care must be taken as to which quadrant
the Cartesian coordinates appear in. Translation and scale
invariance can be achieved by normalising the image using the
Cartesian moments prior to calculation of the Zernike moments
[47]. Translation invariance is achieved by moving the origin
to the image’s COM, causing m
01
= m
10
= 0. Following this,
scale invariance is produced by altering each object so that its
area (or pixel count for a binary image) is m
00
= β, where
β is a predetermined value. Both invariance properties (for a
binary image) can be achieved using :
h(x, y) = f
x
a
+ ¯x,
y
a
+ ¯y
where a =
r
β
m
00
(10)
and h(x, y) is the new translated and scaled function. The
error involved in the discrete implementation can be reduced
by interpolation. If the coordinate calculated by Equation 58
does not coincide with an actual grid location, the pixel value
associated with it is interpolated from the four surrounding
pixels. As a result of the normalization, the Zernike moments
|A
00
| and |A
11
| are set to known values. |A
11
| is set to
zero, due to the translation of the shape to the center of the
coordinate system. This however will be affected by a discrete
implementation where the error in the mapping will decrease as
the shape (being mapped) size (or pixel-resolution) increases.
|A
00
| is dependent on m
00
, and thus on β
|A
00
| =
β
π
(11)
Further, the absolute value of a Zernike moment is rotation
invariant as reflected in the mapping of the image to the unit
disc. The rotation of the shape around the unit disc is expressed
as a phase change, if φ is the angle of rotation, A
R
mn
is the
Zernike moment of the rotated image and A
mn
is the Zernike
moment of the original image then:
A
R
mn
= A
mn
exp (−jnφ) (12)
Moment based features are extracted from the each zone of the
scaled character bitmapped image. The image is partitioned
into zone and features are extracted from each zone. In this
paper Zernike moments based feature extraction is proposed
for off-line Devnagari Handwritten Basic and Compound Char-
acter. To get the feature set, at first, the image is segmented to
30 x 30 blocks, and partitioned as feature set as follows and
the List of the first 8 order Zernike moments is given in Table
IV.
Feature set 1: Fig. 7 (a) is considered as a whole character
image.
Feature set 2: Fig. 7 (b) shows the image divided into four
equal zones.
Feature set 3: Fig. 7 (c) shows the image divided into
three vertical equal zones.
Feature set 4: Figure 7 (d) shows the image divided into
three horizontal equal zones.
(a) (b) (c) (d)
Figure 7. Partition of Devanagari Character into feature set
Table IV. THE FIRST 8 ORDER ZERNIKE MOMENTS
Order Dimensionality Zernike Moments
0 1 A
0,0
1
2 A
1,1
2
4 A
2,0
, A
2,2
3
6 A
3,1
, A
3,3
4
9 A
4,0
, A
4,2
, A
4,4
5
12 A
5,1
, A
5,3
, A
5,5
6
16 A
6,0
, A
6,2
, A
6,4
, A
6,6
7
20 A
7,1
, A
7,3
, A
7,5
, A
7,7
V. CLASSIFICATION AND RECOGNITION
The classification stage is the decision making part of a
recognition system and it uses the features extracted in the
previous stage. We have used Support Vector Machine (SVM)
and k-NN for the purpose of Classification and recognition.
(IJARAI) International Journal of Advanced Research in Artificial Intelligence,
Vol. 3, No.1, 2014
72 |
P a g e
www.ijarai.thesai.org