![](bg2.png)
classified the Arabic baseline extraction methods
into four different groups based on the techniques
used. The simplest one is based on horizontal
projection. Elgammal and Ismail (2001) detected the
baseline by finding peak value of horizontal
projection profile in a printed text-line. This method
has the defect to be very sensitive to the skew
(Pechwitz and Märgner, 2002). A modified
projection technique based on rotating word image
through different angular inclinations is presented by
Al-Rashaideh (2006). The baseline is identified by
finding the maximum value and corresponding angle
among all the peak values is obtained. Pechwitz and
Märgner (2002) proposed the only one work to
detect Arabic handwriting baseline according to the
word skeleton. The main idea of this approach is to
calculate robust features from the skeleton and use
these features for classifying the connected
components into baseline relevant and baseline
irrelevant areas. In a subsequent step, a regression
analysis of points of the relevant objects is done to
estimate the final baseline position. Some
researchers extract baselines after correcting the
slant of the word by a linear regression of the critic
points of the contour having nearly the same
horizontal positions (Farooq et al., 2005). Burrow
(2004) presented a method based on angle detection
by principle components analysis.
Other methods such as the minimization of the
entropy and Hough transform based methods, which
are used for Latin script, are developed and applied
on Arabic script (Côté et al., 1996; Likforman-
Sulem et al., 1995). These methods have the defect
to be expensive in term of calculation time. Lemaitre
et al. (2009) proposed a script independent method
for baseline detection. This method is based on the
principle of the perceptive vision, which combines
several points of view of the same word (from low
to high resolution). Boubaker et al. (2009) described
a baseline detection method which considers
geometric and topologic features. It is tested on
online and offline short Arabic handwritten writing.
Recently, a two-stage Persian/Arabic baseline
detection and correction algorithm is presented by
Ziaratban and Faez (2008). The first stage estimates
the writing path of a text-line by a fitted curve based
on candidate baseline pixels, which are detected
using template matching algorithm. Then the slant
and position of the components in the line is
adjusted. In the second stage, the baseline for each
subword is corrected. Other method of tracing the
baseline in handwritten Persian/Arabic text-line is
proposed by Nagabhushan and Alaei (2010). This
method is based on preparing patches of black and
white blocks all along the text-line, identifying some
candidate points and regressing a curve through
these candidate points to trace the baseline.
The majority of methods presented in literature
failed in estimating the correct baseline for
handwritten text having greater number of ascenders
and descenders. Menasri et al. (2008) described a
baseline extraction method of words overcoming
some difficulties in Arabic script such as the
presence of loops and various shapes for a group of
two or three dots. Inspired by this work, we
developed a baseline estimation method adapted to
Arabic handwritten text-lines.
3 BASELINE ESTIMATION
After scanning a document, some basic
preprocessing tasks like image binarization and
noise reduction have to be performed to increase the
readability of the input by the baseline detection
system. Using the binary image, we perform a noise
reduction filtering. Small holes, produced by writing
and binarization process, are closed and the
unwanted information is deleted by using the
opening and closing morphology operation
respectively (Figure 1(b)).
The developed baseline detection process
consists of three stages: the first one is a basic stage
leading to the detection and removing of diacritical
marks. The second stage extracts the upper baseline
and the lower baseline based on the horizontal
projection histogram. In the final stage, we estimate
more precisely the baseline using support points.
3.1 Diacritical Marks Elimination
More than half of Arabic letters include in their
shape dots which can be one, two or three dots. The
presence of these dots, called diacritical marks, in
their positions allows us to differentiate between
letters that belong to the same family shape. These
diacritical marks lie in either above or under the
baseline depending on the character. In order to
circumvent the bad influences in the process of
baseline detection when using horizontal projection,
we start by removing the diacritical marks based on
the size of the connected components as described in
(Menasri et al., 2008).
A sample of a text-line image
after removing the diacritical marks is shown in
Figure 1(c).
BaselineEstimationinArabicHandwrittenText-Line-EvaluationonAHTID/MWDatabase
431