Arabic Character Segmentation Using Projection-Based Approach with Profile's Amplitude Filter

characters vary with respect to their

this is a great challenge in Arabic tex

Because of the different nat

characters may overlap vertical

compounds of characters at certain

word segments such as "ـﺤﻣ , ﺣﻤـ

represented by single atomic grap

Traditional Arabic font for examp

graphemes, and another common

fewer ligatures) like Simplified Ara

graphemes [1, 17, 18].

r position in the word and

xt [1].

ture Arabic text fonts,

lly

to produce certain

n positions of the Arabic

ﺣ , ـﺠﻧ " which can be

phemes called ligatures.

ple contains around 220

less involved font (with

abic contains around 151

Some Arabic characters have si

ب" and another characters have dou

and other characters have triple do

doted characters exhibit a big

processed.

ingle dot such as "ج , ن ,

uble dots such as "ـﻳ , ـﺗ"

ـﺷ ". The

ots such as "ـﺛ ,

problem while being

This paper presents Arabic

algorithm. The proposed algorithm

approach concepts to separate line

using profile's amplitude filter and s

of the paper is organized as foll

different segmentation techniques.

proposed algorithm. Section 4 dem

performance analysis. Section 5 con

character segmentation

uses the projection-based

es, word, and characters

simple edge tool. The rest

lows: Section 2 reviews

Section 3 presents the

monstrates the results and

ncludes this paper.

II. SEGMENTATION

TECHNIQUES

In this part, methods of how to

contains Arabic text into characte

three seg

This

is done using

segmentation, word

segmenta

segmentation.

o convert the image that

er images are discussed.

line

gmentation stages:

character

ation,

and

A. Line segmentation approaches:

Projection-based approach; in w

being summed along the horizontal

this is referred as a horizontal proje

along the vertical axis for each x val

image and this is called vertical pro

18].

which pixels of image are

axis for each y value and

ection [2-5, 10,14-18 ] or

lue on the segmented line

ojection [2, 3, 12, 13, 14-

Smearing approach; in which

along the horizontal direction are

between the white space is calcula

within a predefined threshold, it is

The text lines are bounded with co

pixels [6, 10, 11].

consecutive black pixels

e smeared. The distance

ated. If the distance lies

filled with black pixels.

onnected shapes of black

Grouping approach; in which

constructed by grouping neighborin

text lines are iteratively

ng connected components

2013 International Conference of Information and Communication Technology (ICoICT)

This profile has information about the text lines that are

indicated by the regions with the black intensities as in Fig.

2. On the other hand, the blank lines appear as a drop in the

black intensities. The text lines can be extracted by

comparing the profile with a pre-defined threshold and this

can be achieved by performing the algorithm described in

Fig. 3.

Fig. 3 shows the line segmentation algorithm that accepts

an image written in Arabic and extracts its lines. This is done

using the horizontal axis profile on two stages. The first one

is to locate each connected group of dark regions in the

profile. The other one is to decide which dark region(s) can

be considered as a separate line.

Figure 2. Horizontal axis profile for a sample image with different

fonts and different word size.

Figure 3. Line segmentation algorithm.

based on certain perceptual criteria such as similarity,

continuity and proximity [7, 10].

Hough-based approach in which the Hough transform is

used for locating straight lines in text images [8-10].

B. Word and character segmentation approaches:

There are four main approaches that deal with connected

characters in an Arabic word [1].

a. Assuming that the input is already segmented into

characters (i.e., no character segmentation will be

needed).

b. Segmenting input words into primitives smaller

than a character then collecting each group of

primitives into character while being recognized.

c. Segmenting words into characters. This is the most

difficult approach in cursive nature languages.

d. Recognizing input words, as a whole, with no

segmentation.

C. Histogram-based Algorithms

Several algorithms use histogram-based techniques for

OCR in different languages. None of these algorithms solved

all the problems associated with OCR in Arabic language. In

Telugu script, the text document image may contain

overlapped lines and characters and no cursive nature

between characters [15,21]. Devnagari is used for writing

Hindi, Marathi, Sanskrit and Nepali languages. Characters

may be separate or connected with a horizontal line at the

upper part, known as Shirorekha and no overlapping

challenge considered [16]. Only line and connected parts

segmentation is performed on Arabic language and the main

difficult in Arabic language, which is to separate characters

and this, is not considered in [20].

III. THE PROPOSED ALGORITHM

text

This paper concerned with segmentation procedure that

accepts an Arabic

image and outputs separated

characters. It has three stages. The first one is responsible for

detecting and separating lines in the text. The segmented

lines are then passed throw the second stage which is

designed to get words from text lines. Finally, the third stage

takes those words and produces the character representation

of each word. The proposed algorithm considers the cursive

and the partially overlapping natures between characters .A

pre-stage is used for detecting and correcting the skew for

the scanned text image. The algorithm in [19] is used to

perform the skew angle correction.

A. Line segmentation:

Line segmentation is done using Image Axis Profile

method that calculates the horizontal axis profile for the

binarized text image [10]. The horizontal axis profile matrix

(cid:1835)(cid:3037) is calculated by summing pixels values P(i,j) along the X-

axis for each y value as shown in (1).

(cid:2919)(cid:2880)(cid:2924)

I(cid:2920) (cid:3404) (cid:3533)(cid:3435)255 (cid:3398) P(cid:4666)i, j(cid:4667)(cid:3439)

(cid:4666)1(cid:4667)

(cid:2919)(cid:2880)(cid:2868)

where i, j are X and Y - axis indices respectively, n = X-

resolution.

978-1-4673-4992-5/13/$26.00 ©2013 IEEE

123

2013 International Conference of Information and Communication Technology (ICoICT)

B. Word segmentation:

Fig. 4 shows an Arabic text line image and its

corresponding vertical profile. The text line is segmented,

beginning from the left side to the right, into connected parts.

These connected parts are clustered to the corresponding

word. The algorithm in explained in Fig. 5. Each word is an

input image to the character segmentation stage.

Figure 4. A test sample for the word segmentation algorithm

Figure 5. Word parts segmentation algorithm

C. Character segmentation:

Character segmentation process is the most important one

in the OCR system, because character will be then entered to

the recognition stage so it should be correctly separated with

no error to be recognized correctly. As shown in Fig. 6, the

text word/sub-word image is entered to the character

segmentation

the base-line

coordinates for each text line image as shown in Fig. 2,

which is in red color. The base-line is the space around the

maximum value in the horizontal profile [12, 15, 17].

stage which

calculates

978-1-4673-4992-5/13/$26.00 ©2013 IEEE

124

Figure 6. The character segmentation algorithm

The proposed algorithm mainly depends on the vertical

profile and its dc components that correspond to connections

between two characters. So, the existence of Hamza of Alif

character such as أ , إ , ﻹ , points of the characters such as ـﻧ ,

ـﺗ , ـﺛ , ـﺷ and all other characters that have points above or

under the character shape may produce errors in our

technique, so another no-points and Hamza image is to be

constructed from the previous word image . The method of

locating points and Hamza is to locate connected group of

dark pixels. The points and Hamza are small connected areas

above or under the baseline. So, they can be erased

successfully as shown in Fig. 7.

By constructing the vertical profile for the no points word

image. The separation between two characters is considered

as constant amplitude in the profile. A constant amplitude

(low variation) passing filter is designed so that only low

variations in the profile will be passed. The filter's output

pulses are a locus of the characters connections. This locus

takes the shape of separated sequential train of pulses as

shown in Fig. 7.

2013 International Conference of Information and Communication Technology (ICoICT)

the green color lines for the original image with points and

Hamza.

Figure 7. A test sample for the character segmentation algorithm

Each pulse in the locus shape will be checked to

determine whether it is a correct characters connection or

not. This is done by using the simple edge tool which finds

step edges along an array of pixel coordinates. This tool

returns the coordinates of all the edges found. If there is only

two edges and the coordinates of these edges lie within the

base-line coordinates, then this pulse is considered as a

character connector as shown in Fig. 7.

To calculate the base-line coordinates, the vertical index

corresponding to the maximum peek in the horizontal axis

profile (cid:1848)-(cid:1861)(cid:1866)(cid:1856)(cid:3040)(cid:3028)(cid:3051) is calculated for each text line generated

from the line segmentation stage as shown in (3).

(cid:1848)-(cid:1861)(cid:1866)(cid:1856)(cid:3040)(cid:3028)(cid:3051) (cid:3404) (cid:1835)(cid:1866)(cid:1856)(cid:1857)(cid:1876)(cid:1841)(cid:1858)(cid:4666)max(cid:4666)(cid:1835)(cid:4667)(cid:4667) (cid:4666)3(cid:4667)

This value is bounded by two lines which are the base line

coordinates as shown in Fig. 2.

An image cutter tool is used on the original text image to

extract sub-images which correspond to each pulse in the

filter response. Each sub-image is extracted vertically

starting from the first index to the last index of each pulse. A

horizontal axis profile is calculated for each sub-image. If the

output of the horizontal axis profile takes the shape of one

pulse with constant amplitude and the vertical index

(cid:1848)-(cid:1861)(cid:1866)(cid:1856)(cid:3040)(cid:3028)(cid:3051) lies inside this pulse, then the pulse coordinates

are considered as the base line coordinates.

Fig. 8 shows a word, its filter output, and the horizontal

axis profile for each sub-image corresponds to each pulse in

the filter output. The vertical index for the second line in Fig.

2 is (cid:1848)-(cid:1861)(cid:1866)(cid:1856)(cid:3040)(cid:3028)(cid:3051) (cid:3404) 29. Hence, the first vertical pulse in Fig. 8

is refused because it gives a horizontal profile from 36 to 41,

which is away from (cid:1848)-(cid:1861)(cid:1866)(cid:1856)(cid:3040)(cid:3028)(cid:3051). The second vertical pulse is

accepted because the vertical index (cid:1848)-(cid:1861)(cid:1866)(cid:1856)(cid:3040)(cid:3028)(cid:3051) lies inside its

horizontal profile (i.e. from 28 to 32). The third vertical

pulse is refused as it gives more than one pulse in its

horizontal axis profile.

Fig. 7 shows the original test images, the no point

images, and output of the low variation passing filter whose

input is the no-point vertical profile for the word. For the

pervious algorithm, only pulses pass are those which marked

with a green line. Character separation operation is done at

978-1-4673-4992-5/13/$26.00 ©2013 IEEE

125

Figure 8. A word with its filter output and the horizontal axis profile for

each sub-image corresponds to each pulse

The character (ر) may overlap in many fonts with the

remaining part of the word. This represents a challenge when

separating the word into connected parts. The proposed

algorithm locate these connected parts as well as points and

Hamza overcoming the problem of separating partially

overlapped characters in [15] as shown in Fig. 9.

Figure 9. Samples with different font and size

IV. RESULTS AND PERFORMANCE EVALUATION

Test operation is done on a document (264 pages, 6,286

lines, 50,931 words, 224,781 characters with no spaces,

275,136 characters with spaces). The line segmentation

algorithm achieved a very large correction ratio that reaches

99.9%. One error is found in the test when dealing with a

group of text lines bounded by a text box. All these lines and

the text box are treated as a single line and this happened

because of the continuity of the text box.

2013 International Conference of Information and Communication Technology (ICoICT)

The word segmentation algorithm achieved a correction

ratio of approximately 99.9%. One error appeared when

dealing with marks such as exclamation, question, comma,

and semicolon. When writing those marks directly after the

word with no separation space (e.g. ؟اذﺎﻣ , ﻰﺘﻣ

), the algorithm

gives an indication that those marks belong to the same

pervious word cluster number as the space between them is

in sufficient to decide that the mark is another word.

!

The character segmentation algorithm achieved also an

acceptable correction ratio that reaches 98%. This algorithm

deals with the compounded characters as a separate character

(e.g. ـﺤﻣ , ـﺠﻧ). These compounded characters will be dealt

with in the recognition stage. This makes the alphabetic

Arabic language equal to 28 basic characters plus 220

compounded characters.

V. CONCLUSION

This paper presents Arabic character segmentation

algorithm. The proposed algorithm uses the projection-based

approach concepts to separate lines, words, and characters.

The lines produced from the line segmentation stage are

entered to the next one which segments them to connected

parts. These connected parts may be separate characters or a

number of connected characters. Those connected characters

are entered then to the character segmentation stage that is

responsible for separating these connected characters. The

proposed algorithm achieved a promising success rate ratio

for line, word and character segmentation.

REFERENCES

[1] B. Al-Badr and S. A. Mahmoud, “Survey and Bibliography of Arabic

Optical Text Recognition,” Elsevier Signal Processing, vol. 41, pp.

49-77, 1995.

[2] K. Bouhilali, M. Kamrouni, and N. Ellouze, “Method

of

Segmentation of Arabic Text Image into Characters,” Kuwaiti

Computer Conferance, pp. 442-446, 1989.

[3] M. Attia and M. El-Mahallawy, “Histogram-Based Lines and Words

Decomposition for Arabic Omni Font-Written OCR Systems;

International Conference on

Enhancements and Evaluation,”

Computer Analysis of Images and Patterns , pp. 522-530, 2007.

[4] Z. Razak, K. Zulkiflee, M. Y. I. Idris, E. M. Tamil, M. Noorzaily, M.

Noor, R. Salleh, M. Yaakob, Z. M. Yusof, and M. Yaacob, “Off-line

Handwriting Text Line Segmentation : A Review,” International

Journal of Computer Science and Network Security, vol. 8, no. 7,

2008.

[5] G. Tímár, K. Karacs, and Cs. Rekeczky, “Analogic Preprocessing and

Segmentation Algorithms For Offline Handwriting Recognition,” in

Proc. of the 7th IEEE International Workshop on Cellular Neural

Networks and their Applications, pp.407-414, 2002.

[6] Y. Li, Y. Zheng, D. Doermann, and S. Jaeger, “A new Algorithm for

Detecting Text Line in Handwritten Documents,” International

Workshop on Frontiers in Handwriting Recognition, pp. 35-40, 2006.

[7] M. Feldbach and K. D. Tönnies, “Line Detection and Segmentation in

Historical Church Registers,” International Conference on Document

Analysis and Recognition, pp. 743-747, 2001.

[8] L. Likforman-Sulem, A. Hanimyan, and C. Faure, “A Hough Based

Algorithm for Extracting Text Lines in Handwritten Documents,”

International Conference on Document Analysis and Recognition,

vol. 2, pp. 774-777, 1995.

[9] G. Louloudis, B. Gatos, I. Pratikakis, and K. Halatsis, “A Block-

Based Hough Transform Mapping for Text Line Detection in

Handwritten Documents,” International Workshop on Frontiers in

Handwriting Recognition, 2006.

[10] L. Likforman-Sulem , A. Zahour, and Bruno Taconet, “Text Line

Segmentation of Historical Documents: a Survey,” International

Journal on Document Analysis and Recognition, vol. 9, pp. 123 - 138,

2007.

[11] Z. Shi and V. Govindaraju, “Line Separation for Complex Document

Images Using Fuzzy Runlength,” International Workshop on

Document Image Analysis for Libraries, 2004.

[12] L. Zheng, A. H. Hassin, and X. Tang , “A new Algorithm for

Machine Printed Arabic Character Segmentation,” Pattern

Recognition Letters, vol. 25, pp. 1723–1729, 2004.

[13] B. M. F. Bushofa and M. Spann , “ Segmentation and Recognition

of Arabic Characters by Structural Classification,” Elsever Image

and Vision Computing, vol. 15 ,pp.167-179, 1997.

[14] V. Kumar, P. K. Sengar, “Segmentation of Printed Text in

Devanagari Script and Gurmukhi Script,” International Journal of

Computer Applications, vol. 3, no. 8, pp. 24-29, 2010.

[15] M. S. Das, C. Reddy, A. Govardhan and G. Saikrishna,

“Segmentation of Overlapping Text Lines, Characters in Printed

International Journal of

Images,”

Telugu Text Document

Engineering Science and Technology, vol. 2, pp. 6606-6610, 2010.

[16] V. J. Dongre and V. H. Mankar, “Devnagari Document Segmentation

Using Histogram Approach,” International Journal of Computer

Science, Engineering and Information Technology, vol.1, no.3, pp.46-

53, 2011.

[17] S. Elaiwat, M. A. Abu-zanona, “A Three Stages Segmentation Model

for a Higher Accurate off-line Arabic Handwriting Recognition,”

World of Computer Science and Information Technology Journal,

vol. 2, no. 3, pp. 98-104, 2012.

[18] M. A. Abdullah, L. M. Al-Harigy, and H. H. Al-Fraidi, “Off-Line

Arabic Handwriting Character Recognition Using Word

Segmentation,” Journal of Computing, vol.4, pp.40-44, 2012.

[19] A. M. Al-Shatnawi and K. Omar, “Skew Detection and Correction

Technique for Arabic Document Images Based on Centre of Gravity,”

Journal of Computer Science, pp. 363-368, 2009.

[20] S. Brook and Z. Al Aghbari, “Classification of Personal Arabic

Handwritten Documents,” Transactions on Information Science &

Applications, vol. 5,pp. 1021-1030, 2008.

[21] N. Sridevi and P. Subashini, “Segmentation of Text Lines and

Characters in Ancient Tamil Script Documents using Computational

Intelligence Techniques,”

International Journal of Computer

Applications, vol. 52, no.14, 2012.

978-1-4673-4992-5/13/$26.00 ©2013 IEEE

126