Applied Mathematics, 2013, 4, 1313-1319
http://dx.doi.org/10.4236/am.2013.49177 Published Online September 2013 (http://www.scirp.org/journal/am)
A Fast Recognition System for Isolated Printed Characters
Using Center of Gravity and Principal Axis
Ahmed M. Shaffie, Galal A. Elkobrosy
Department of Engineering Mathematics and Physics, University of Alexandria, Alexandria, Egypt
Email: amshaffie.oq@gmail.com
Received January 22, 2013; revised February 22, 2013; accepted Mar ch 2 , 2013
Copyright © 2013 Ahmed M. Shaffie, Galal A. Elkobrosy. This is an open access article distributed under the Creative Commons
Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.
ABSTRACT
The purpose of this paper is to propose a new multi stage algorithm for the recognition of isolated characters. It was
similar work done before using only the center of gravity (This paper is extended version of “A fast recogn ition system
for isolated printed characters using center of gravity”, LAP LAMBERT Academic Publishing 2011, ISBN: 978-3-
8465-0002-6), but here we add using principal axis in order to make the algorithm rotation invariant. In my previous
work which is published in LAP LAMBERT, I face a big problem that when the character is rotated I can’t recognize
the character. So this adds constrain on the document to be well oriented but here I use the principal axis in order to
unify the orientation of the character set and the characters in the scanned document. The algorithm can be applied for
any isolated character such as Latin, Chinese, Japanese, and Arabic characters but it has been applied in this paper for
Arabic characters. The approach uses normalized and isolated characters of the same size and extracts an image signa-
ture based on the center of gravity of the character after making the character principal axis vertical, and then the system
compares these values to a set of signatures for typical characters of the set. The system then provides the closeness of
match to all other characters in the set.
Keywords: OCR; Pattern Recognition; Confusion Matrix; Image Signature; Word Segmentation; Character
Fragmentation
1. Introduction
Character recognition is not a difficult task for human
beings. The purpose of optical character recognition
(OCR) systems is to recognize the characters from im-
ages. OCR systems have been divided into two catego-
ries, namely: on-line and off-line techniques. On-line
techniques are mainly dependant on the motion of hand
while the characters are being written; hence this tech-
nique is mainly used in the recognition of hand written
documents. One of the main problems of that technique
is that it cannot be used for already written documents
and for printed characters and its need for special digitiz-
ers or PDA where sensors pick up the pen-tip movements
as well as pen-up/pen-down switching. There is multiple
papers exist explaining it, for which, a well formed sur-
vey is given by Nouboud et al. [1]. On-line techniques
provide be tter results than off-line techniques as it uses a
significantly larger set of information which is not avail-
able for off-line techniques which are only dependent on
stored images. However, off-line techniques are the only
techniques available for hard copy and printed papers.
Cowell et al. [2] give a procedure for an OCR system.
This paper uses the same schedule but it uses the feature
of center of gravity instead of counting pixels in rows
and columns as Cowell et al. [2]. The reason for using
this different technique is because the pixel counting
method is a very exhaustive technique as it requires pass-
ing through every pixel and provides no methods of im-
proving the recognition and to decrease the confusion
between characters as proposed in this paper by using the
center of gravity. Tabassam Nawaz et al. [3] use the
same schedule of Cowell [2] but change the method of
extracting the features to be the number of consecutive
ones and zeros using a method called “Chain Code”.
However, this method also requires a lot of calculations
and processing. Multiple different techniques exist for
character recognition such as “structural information” as
number of holes and strokes, but these techniques cannot
be used for every character set and have to be revised
completely for each different character set as there is a
lot of style of characters such as English, Arabic and
C
opyright © 2013 SciRes. AM
A. M. SHAFFIE, G. A. ELKOBROSY
1314
Chinese characters, every one of them has its own char-
acteristics. In the Arabic character set, one of the main
features is dots. Up to 3 dots can exist for Arabic charac-
ters, and hence no one criteria can be used to apply for all
of these character sets. Some techniques for these styles
are applied for Arabic characters by Cowell et al. [4]
using “thinning” and “feature extraction”, however, that
technique was slow and cannot be modified to another
character sets easily. One of the main problems of the
techniques used by OCR systems is that the character is
wrongly identified, so th e features can be tested by bu ild-
ing a confusion matrix Cowell [5] to determine whether
this technique is good fo r this character set or it will lead
to a problem in the recognition phase. And it has been
derived a way to resolve this conflict Cowell [6]. The
proposed work here gets its importance from its scale and
rotation invariant and of course because its calculations
are minimum while its accuracy is very good and can be
tuned by doing more calculations to get more accuracy.
2. An Overview of the Proposed System
The paper’s approach in recognition makes use of five
phases as outlined below:
Read input image.
Line and characters segmentation.
Normalize character to a standard size, 100 × 100
pixel resolutio n has been used in the implem ent at i on.
Extract the character signature.
Compare the character signature with the signature
templates of the character set.
2.1. Text Image and Text Line Segmentation
The segmentation of the image is done at two levels.
First, the text in the image is split into lin es of text using
the horizontal projection technique (i.e. location of hori-
zontal lines of zero density of pixels, given the line of
text from the horizontal projection technique, indicates
the beginning of a segment and the subsequent location
of another zero density line of pixels indicates the end of
a segment, thus an entire segment is located). Second,
each line of text is split into characters using vertical
projection technique (i.e. location of vertical lines of zero
density of pixels, given the line of character from the
vertical projection technique, indicates the beginning of a
segment and the subsequent location of another zero
density line pixels ind icates the end of a segment, thus an
entire segment is located) [7]. Figure 1 illustrates the
text image to text lines segmentation and the text line to
individ ual charact e r segmentati o n.
2.2. Normalization of the Fragmented
Characters
One of the important stages in the process is to insure
that the input character has the same dimensions and the
same orientation as the characters used to create the sig-
nature or the configuration file, so the procedure start by
unified the orientation of the characters by using the
character principle axis, so the normalization started by
getting the principle axis of the character and rotate the
character image to make the principle axis vertical, and
then trimming the white p arts of the character image then
it is scaled to 100 × 100 pixels. Figure 2 shows the case
of Arabic character alef and mim. The figure on the right
shows the scanned character as scanned and the figure on
the left shows the character after it expanded so that it
has the size required and therefore touches each side on
the 100 × 100 square as shown [2 ].
2.3. Get the Center of Gravity
The signature of each character is produced by getting
the center of gravity of the normalized character as if the
character is a uniform body and the center of gravity co-
ordinate XG, YG is calculated using the following for-
mula
XG = thegma xi/n
YG = thegma yi/n
where:
n is the num ber of pixels.
x, y is the coordinate of the black pixels in the image
of the character.
This approach is applied to the Arabic characters as
Figure 1. Image and line text segmentation.
Figure 2. Normalized characters and other original form.
Copyright © 2013 SciRes. AM
A. M. SHAFFIE, G. A. ELKOBROSY
Copyright © 2013 SciRes. AM
1315
the configuration file which contains the number of seg-
ments and each character signature. Figure 4 shows the
structure of the proposed learning process.
the algorithm treats each character as a body and at-
tempts to get the center of gravity for it; a major chal-
lenge with this technique is that more than one character
have the same center of gravity or that two centers of
gravity are approximately the same. A solution for this
problem is provided by dividing the character into 4
bodies and in some cases into 9 bodies and getting the
centers of gravity of the divided bodies and then use
these points to identify the character. It is found that,
when the character is divided, there is no conflict be-
tween characters signature and so this signature can be
used successfully in recognition. Figure 3 shows the cen-
ter of gravity of a character and then show the center of
gravity after the character is divided into 4 and 9 seg-
ments.
Second, the recognition process of the input image; the
program separates the image first into lines, then to ind i-
vidual characters and asks for the configuration file that
the program will use to recognize the characters in the
image and then classify the characters by getting the
nearest character in the configuration file by calculating
squared equilidian distance to all the character set and get
the minimum distance. Figure 5 shows the structure of
the proposed OCR system.
4. The Confusion Matrix
The closeness of match between every character can be
calculated and is put in a matrix called the “confusion
3. The Implementation of the System
The system described above has been fully implemented
using Java in two processes. First, learning process at
which the configuratio n file is created. The configuration
file contains the characterset signature and determine the
number of segments used to get the signature, the input
of this process is the full character set and then the sys-
tem normalizes every character and gets the center of
gravity based on the number of segments, the output is Figure 3. The center of gravity location of ein Arabic char-
acter usnig 1, 4, 9 segmentation model.
Figure 4. The structure of the proposed learning process.
Figure 5. The structure of the proposed OCR system.
A. M. SHAFFIE, G. A. ELKOBROSY
1316
matrix” [2,5]. The application has a tab to get the confu-
sion matrix for the selected configuration file to provide
indications about the suitability of using this configura-
tion file in recognition, but in some cases there are more
than one character that has nearly the same signature and
so the number of segments can be changed in order to
solve this conflict.
5. Screen Shoots of the Implemented
Program
The program consists of three tabs. First, recognition tab
in Figure 6, this tab is used for loading the image which
contains the characters which has to be recognized, then
a configuration file which contains the signature of the
character being loaded, and then the characters are rec-
ognized and written in the selected output file. second,
learning tab in Figure 7, at this tab, the “Load Alpha”
button is used to load the character set images, and after
that the ”Scale & Trim” button is used to normalize all
characters, then from the List of Values the “number of
segments” is selected, and the center of gravity is calcu-
lated and saved in a configuration file to use it later in the
recognition tab. Third, confusion matrix tab in Figure 8,
in this tab a configuration file is loaded and the program
calculates the distance between each pair of the charac-
ters in the character set to show them in the grid and to
know if this configuration file can be used efficiently or
not.
6. Performance Evaluation
In this part we will conduct with a blackbox evaluation
for two Arabic OCR products which are our proposed
Figure 6. Recognition tab.
Figure 7. Learning tab.
Figure 8. Confusion matrix tab.
OCR system and Cowell OCR system. Our proposed
system and Cowell system deal with a character image
with resolution 100 × 100 pixels and of course this will
make the comparison process very easy and very expres-
sive.
6.1. Transformations Invariance
Our proposed system and Cowell system dial with a
character image normalizing the character image and
resizing it to resolu tion 100 × 100 pixel so this will make
both system translation and scale invariant , but Cowell
system will face a big problem if the character image is
Copyright © 2013 SciRes. AM
A. M. SHAFFIE, G. A. ELKOBROSY 1317
rotated a certain angle this will show that Cowell system
is not rotation invariant. Our proposed algorithm is rota-
tion invariant as we first rotate any character image to
make its principal axis vertical and this will guarantee
that anycharacters have the same shape will have the
same principal axis and by the way have the same orien-
tation. For example Cowell syste m failed torecognize the
image in Figure 9 but our proposed system can recog-
nize all the characters in this text image.
6.2. Number of Comparisons
Of course the number of comparisons while recognizing
any text image is very important as it affects the execu-
tion time of the system. In Cowell system it depends on
comparing the number of black pixels in each row and in
each column and treat with this numbers as the signature
of the character, So when we try to recognize a single
character we have to compare 200 number (100 number
which denote the number of black pixels in each row and
100 number which denote the number of black pixels in
each column). In our proposed system we get the signa-
ture by getting the center of mass of the character and so
we have one point coordinate it means that we reduce the
number of comparisons to only two comparisons. We
add an option as mentioned before to increase the accu-
racy by dividing the character image into 4 or 9 images
and get the center of mass to every part. Now in case we
divided the character image to 4 parts we will have 8
comparisons and if we divided the character image to 9
parts we will have 18 co mparisons. Figure 10 shows the
relation between the number of comparisons and the
number of characters in the text image when we use
Cowell system and our proposed system with different
number of segments.
6.3. Execution Time
One of the very important factors in any OCR system is
the execution time. Cowell claim that his technique can
identify in the region of 100 characters per second and
we test our proposed system and we get in the region of
250 characters per second. Figures 11 and 12 show the
relation between the number of characters in the page
and the execution time it take in milliseconds when we
use 9 segments characters and 4 segments characters
respectively.
6.4. Resolution of the Image and Font Size
Of course when we increase the resolution of the scanned
image, and the font size we get more details about the
character. So the effect of the font size is also shown.
Our proposed algorithm will be applied on different im-
ages each one has different font size and different resolu-
tion of image. From the results, we concluded that, using
300 dpi resolutions is the best choice giving average ac-
curacy about 98% and no need to use resolution more
Figure 9. Text image contained some characters rotated
different angles.
Figure 10. Comparing the number of comparisons in john cowel system and our proposed syste m using 1, 4 and 9 character
egments. s
Copyright © 2013 SciRes. AM
A. M. SHAFFIE, G. A. ELKOBROSY
1318
Figure 11. Number of characters execution time using 9
segment characters.
Figure 12. Number of characters execution time using
se it will need more execution time.
7. Conclusions
s a fast recognition system based on
Table 1. Our proposed system accuracy using different font
4
segment characters.
an 300 dpi becauth
And the imagesize will be very large (more than 1.5 MB).
Table 1 shows the result o f using different image resolu-
tion with different font size. The accuracy is calculated
by dividing the number of right recognized characters by
the total number of characters in the input text image.
This paper describe
creating images signatures which can be used for any
character set. The Arabic character set is used here but
this method can be used in any character set. The system
used is very rapid, as it uses the center of gravity of the
character and then calculates the distances to all the
characters to get the nearest one which is selected as the
recognized character. This method starts by normaliza-
tion and done by scaling the character to standard size
and rotating the character image in order to make its
principal axis vertical and this normalization guarantees
that this signature is scale and rotation invariance. The
performance and accuracy of this technique can be
tweaked by changing the number of segments each char-
acter is divided into. The confusion matrix gives an indi-
size and different image resolution.
Image Resolution
Font size150 dpi 200 dpi 300 dpi 400 dpi
11 62% 83% 92% 96%
12 70% 84% 92% 96%
14 72% 85% 94% 98%
16 73% 85% 94% 98%
18 73% 90% 96% 98%
20 76% 90% 96% 99%
22 76% 93% 97% 99%
24 78% 93% 97% 99%
26 78% 93% 99% 99%
28 80% 93% 99% 99%
36 80% 94% 99% 100%
48 80% 96% 100% 100%
72 83% 98% 100% 100%
catorhow me chas are each
Adional wan be in th of sta-
on and fragmen tation of the characters especially in the
“On-Line Recognition of
Handprint Characters,” Pattern Recognition, Vol. 23, No.
9, 1990, pp. 10
to uch thracternear to other.
ditork c donee partegmen
ti
Arabic character as our proposed algorithm dealt with
isolated characters; however, Arabic characters in prac-
tice are not isolated. And there is a lot of features can be
added to the signature such as the number of dots and its
position which can help in decreasing the number of
segments used and so the program will be faster and at
the same time more accurate.
REFERENCES
[1] F. Nouboud and R. Palmondon,
31-1044.
doi.org/10.1016/0031-3203(90)90111-W
[2] J. Cowel and F. Hussain, “A Fast System for Isolated
Arabic Characters,” Proceeding of the 6th International
Conference on information Visualisation IEEE
July 2002, pp. 650-654. , London,
[3] T. Nawaz, S. A. H. S. Naqvi, H. Rehman and A. Faiz,
“Optical Character Recognition System for Urdu (Naskh
Font) Using Pattern Matching Technique,” International
Journal of Image Processing, Vol. 3, No. 3, 2009, pp. 92-
104.
[4] J. Cowel and F. Hussain, “Thinning Arabic Characters for
Feature Extraction,” IV 2001 Proceedings IEEE Confer-
ence on Information Visualization 2001, London, July
2001.
[5] J. Cowell and F. Hussain, “The Confusion Matrix: Iden-
Copyright © 2013 SciRes. AM
A. M. SHAFFIE, G. A. ELKOBROSY 1319
tifying Conflicts in Arabic and Latin Character Recogni-
tion,” CGIM 2000, Las Vegas, 2000.
[6] J. Cowell and F. Hussain, “Resolving Conflicts in Arabic
ive Mutli-Direc-
and Latin Character Recognition,” 19th Eurographics UK
Conference, London, April 2001.
[7] M. Al-A’ali and J. Ahmed, “Optical Character Recogni-
tion System for Arabic Text Using Curs
tional Approach”, Journal of Computer Science, Vol. 3,
No. 7, 2007, pp. 549-555.
doi.org/10.3844/jcssp.2007.549.555.
Copyright © 2013 SciRes. AM