2009 10th International Conference on Document Analysis and Recognition
Enhanced Text Extraction from Arabic Degraded Document Images using
EM Algorithm
Wafa Boussellaa1 , Aymen Bougacha1, Abderrazak Zahour2, Haikal EL Abed3, Adel Alimi 1
1
University of Sfax, REGIM, ENIS, Route Soukra, BPW, 3038, Sfax, Tunisia
2
IUT, Université du Havre, Place Robert Schuman, 76610 Le Havre, France
3
Technical University Braunschweig, Institute for Communication Technology (IfN), Germany
Email:{wafa.boussellaa,adel.alimi,elabed}@ieee.org, Aymen.Bougacha@gmail.com,
abderrazak.zahour@univ-lehavre.fr
In this paper an enhanced text extraction method is
proposed on the basis of the maximum likelihood (ML)
estimation for the segmentation problem. The difficult
task lies in how to estimate the parameters of the
likelihood functions and the number of segments.
Eventually, the expectation maximization algorithm
(EM) algorithm is introduced in order to improve the
parameter estimation. The initial estimates for the EM
algorithm are given by the k-means clustering
algorithm to avoid the problem of random initial
selection. The text and background segmentation is
performed by the conventional ML method. The
segmented image is used to produce a final colored
restored document image
The rest of this paper is organized as follows.
Section 2 gives an overview of previous work of
degraded document image enhancement. Section 3 and
4 describes the proposed algorithm in details.
Experimental results are presented in section 5.
Conclusion and future work are given in last section.
Abstract
This paper presents a new enhanced text extraction
algorithm from degraded document images on the
basis of the probabilistic models. The observed
document image is considered as a mixture of
Gaussian densities which represents the foreground
and background document image components. The EM
algorithm is introduced in order to estimate and
improve the parameters of the mixtures of densities
recursively. The initial parameters of the EM
algorithm are estimated by the k-means clustering
method. After the parameter estimation, the document
image is partitioned into text and background classes
by the means of ML approach. The performance of the
proposed approach is evaluated on a variety of
degraded documents comes from the collections of the
National library of Tunisia.
1. Introduction
2. Related work
The automatic processing of degraded historical
documents is a challenge in document image analysis
field which is confronted with many difficulties due to
the storage condition and the complexity of their
content. For historical degraded and poor quality
documents, enhancement is not an easy task. The main
interest of an enhancement step of historical documents
is to remove information coming from the background.
Background artifacts can derive from many kinds of
degradation, such as scan optical blur and noise, spots,
underwriting, or overwriting. Most previous document
image enhancement algorithms have been designed
primarily for binarization. Binarization aim to extract
text from distorted degraded documents and its related
methods are proposed for processing gray documents
which have not been extended for color documents.
978-0-7695-3725-2/09 $25.00 © 2009 IEEE
DOI 10.1109/ICDAR.2009.220
According to the literature, approaches that deals
with document image enhancement and text extraction
are based on binarization or foreground/background
separation techniques. Most previous document image
enhancement algorithms have been designed primarily
for binarization. Binarization is performed either
globally or locally. Global thresholding methods are
not sufficient since document images usually are
degraded including shadows, non-uniform illumination
and low contrast. Local methods are shown to perform
better according to recent exhaustive survey of image
binarization methods presented in [15]. Two main
approaches are distinguished, based local thresholding
methods and clustering based methods.
Based local thresholding techniques have been
proposed to estimate a different threshold for each
743
pixel according to the grey-scale information of the
neighboring pixels. Some of these popular methods,
namely Otsu's thresholding technique [13] locally
adaptive technique [11,14].
Other adaptive methods specially designed for
historical and distorted documents are based on
adaptive threshold segmentation. Gatos et al. [8]
presents a binarization methodology based on
background estimation used to segment the image,
various pre- and post-processing steps are needs in this
approach. Oh et al. [12] presents an iterative algorithm
based on water flow models and a hierarchical
thresholding. This method deals with low contrasted
documents images.
An iterative approach for segmenting degraded
document images is described by Kavallieratou et al
[10]. It consist in obtaining an initial segmentation
using a global thresholding and applying a local
thresholding on the areas which are incorrectly
segmented and detected.
Other methods for historical document image
enhancement are driven by the goal of improving
human readability of the documents are Based
clustering methods. These methods are dedicated for
foreground/background separation of color document
images using a classification approach. Garain et al. [7]
have proposed an adaptive method for foregroundbackground separation in low quality color document
images. A connected component labeling is initially
implemented to capture the spatially connected similar
color pixels. Next, Dominant background components
are determined to divide the entire image into a number
of grids which representing local uniformity in
illumination background.
Drira et al. [6] proposed a recursive method of
unsupervised clustering. It classifies the pixels of
document image in three clusters (background, original
text, and show-through effect) in the degraded
document. Thereafter the show-through effect must be
eliminated and replaced by the color of the
background.
Agam et al. [2] have described a novel approach based
on probabilistic models (EM) for foreground and
background separation. This algorithm deals with low
contrasted document images.
The observed document image is considered as a
mixture model of two Gaussian densities which
represents the foreground (YF) and background (YB)
document image components. The data of this model
are determined by a random vector X of the probability
density function (PDF) which is written as the
following, equation 1:
K
F ( x,θ ) = ∑ π k f k ( x;θ k )
(1)
k =1
Where:
K =2: The number of densities assumed a
priori.
(
θ
=
µ
YB , σ YB ) : The mean and the standard
1
deviation vectors of YB component.
θ 2 = ( µ YF , σ YF ) : The mean and the standard
deviation vectors of YF component.
The PDF f1 ( x;θ1 ) and f 2 ( x;θ 2 ) are used for
maximum likelihood clustering. Then, the estimation
of θ1 and θ 2 parameters are performed using of the
EM algorithm which need an initializing parameter
step. This step is the first one of our segmentation
algorithm detailed in the following section.
4. Proposed Method
The proposed approach is considered as a novel
view and improvement of our proposed methodology
published in [3]. This approach belongs to our system
PRAAD (Pre-processing and Analysis Tool for Arabic
Ancient Document) [4].
Our approach process both color and grayscale
document images. To apply our approach on color
document image, we convert the image to YIQ colors
space and operate on Y luminance channel. This choice
is justified by the fact that the human vision is very
sensitive to the change of luminosity. Moreover, the
variation in light intensity caused by the uneven
background of poor degraded is captured in Y channel.
According to the poor quality of document images and
the pale colors and degradations coming from the
background artifacts which affect the foreground
contrast, we apply a stretching to the intensity values
of image histogram using a proportion value. Then, the
image YC is produced with proportion between 2%
and 8% which gives correct results. After this needed
pre-processing step, the contrasted image is operated
by the segmentation algorithm detailed in the
following section.
3. The EM Algorithm
The EM algorithm mentioned in [5] is an iterative
algorithm for calculating the maximum-likelihood or
maximum-a-posterior estimates when the observations
can be viewed as incomplete data. Each iteration of the
algorithm consists of an expectation step followed by a
maximization step.
744
4.2. The segmentation algorithm
µ̂
Our proposed text and background segmentation
algorithm operates in three steps which are presented
below.
( t +1)
k
∑ zˆ H
=
∑ zˆ
n
(t )
i
i =1 ik
n
(t )
i =1 ik
: Estimates a priori
∧
means µ k of k -th density of the mixture.
4.2.1. Initial estimation
Initial estimates θ1 and θ 2 and their corresponding
mean
and
standard
deviation
vectors
(0)
(0)
(0)
(0)
( µ YB
, σ YB
, µ YF
, σ YF
) for EM algorithm are calculated
using k-means clustering method presented in [3].
σˆ
2
n
i +1
zˆ ik( t ) ( H i − µˆ k( t +1) ) 2
∑i =1 zˆik(t )
n
: Estimates
∧
Until
Q(θ ;θˆ (t ) ) − Q(θ ;θˆ (t −1) ) < ε
.
4.2.3. Maximum likelihood segmentation
The image segmentation is carried out by the
conventional ML method using θ ( EM ) . The ML method
estimates the probability that a pixel belongs to its
corresponding class which is text or background and
assigns it when its probability is maximal. We are
using the probability distributions of Raleigh law.
According to the distribution, the likelihood is
expressed in the following equations (According to the
Raleigh law):
]
1
∑
=
covariance matrix σ k of k -th density of the
mixture.
4.2.2. Iterative Estimation by EM algorithm
The EM algorithm is iteratively carried out with the
initial estimates θ ( 0 ) and the intensity histogram H of
YC document image. The EM algorithm converges
when difference of old estimates and new estimates are
less than some threshold ε and the final estimates
θ ( EM ) is obtained. The EM algorithm contributes to the
segmentation algorithm by way of improving the
parameters of the mixture of densities on the basis of
the ML criterion. The EM algorithm in initialized as
below.
Algorithm EM
Input:
K=2
θ(0) = π1(0) ,π2(0) , µ1(0) ,σ1(0) , µ2(0) ,σ2(0) : estimates vectors
by k-means algorithm.
Where:
− ( µ1( 0) , σ 1( 0 ) , µ 2( 0) , σ 2( 0) ) : Means and standard
deviation vectors
(
0
)
(
− (π , π 0) ) = ( 1 , 1 )
[
( t +1)
k
1
f k =1, 2 (YC ) =
µˆ k
2
π
exp( −
YC 2
)
2 2
ˆ
2( µ k
)
π
The pixel (i,j) in YC is labeled Vk (i, j ) according to
the following equation:
2 2
H a histogram vector defined previously
ε the threshold for the algorithm convergence
Output:
⎧Vk (i, j) =1
⎨
⎩Vk (i, j) = 0
θˆ (0) : Local Maximum of likelihood law
Experimental works presented in our last work
presented in [3], proved that for the case of degraded
manuscripts, the Raleigh distribution gives better
results for text-background segmentation. Figure 1 and
Figure 2 show segmentation results of YC image both
in grayscale and color space.
t ←0 ;
θˆ ( 0 ) = θ ( 0 ) : Model initialization
Repeat
(Expectation step)
(t )
Compute the posterior probabilities zˆ i ,k .
zˆi(,tk) =
f k (YC(i, j)) = max(f1 (YC(i, j), f 2 (YC(i, j)))
si non
5. Experimental Results
πˆ k( t ) f ( H i ;θˆk(t ) )
∑
si
K
πˆ l(t ) f ( H i ;θˆl( t ) )
(Maximization Step)
As mentioned in the related work study, the
document images enhancement methods are driven by
the goal of improving human readability of document
text. To evaluate the performance of our proposed
method in enhanced text extraction, we achieve our
tests on 100 scanned images of old degraded
handwritten documents given by the National library of
Tunisia [1].
l =1
π k , µ k ,σ k
Estimates of
πˆ k(t +1) =
∑
n
i +1
zˆ
(t )
i,k
maximizing Q (θ ;θˆ
(t )
)
: Estimates a priori
n
∧
probability π k of k -th density of the mixture.
745
To assess the performance of our method especially for
difficult cases, three degradation types were selected
for evaluation and the images are visually inspected.
The proposed method is evaluated using two sets of
metrics. The first sets are based on two selected and
correlated binarization criteria. The criteria are
misclassification error (ME) and the relative
foreground area error (RAE) proposed by Sezgin and
Sankur [15]. In order to be calculated, these criteria
need the ground-truth binary image which is manually
obtained. The expected values of the above criteria
vary between [0, 1]. In all cases, the measure that is
closer to zero corresponds to the best binarization
result. The analytical score values of the two criteria
obtained for three types of degradations document
image are shown in Table 1.
The second sets used the precision and recall
criteria presented in [16]. The two criteria are defined
below and results are shown in Table 2.
Precision = No. of correctly pixel's foreground
extracted / Nbr of all pixel's foreground extracted by
the proposed method
Recall = No. of correctly pixel's foreground extracted /
total Nbr of pixel's foreground present in the
document.
Where the number of pixel’s foreground present in
the document is calculated by using the ground-truth
image. Precision reflects the performance of removing
the degradation and recall reflects the performance of
extracting the foreground.
These evaluations include comparative results
between the proposed method and some known
binarization algorithms [8, 11, 12, 13, 14]. Moreover
an average performance score is computed for global
decision. The result in Table 1 and Table 2 shows
respectively that our method achieves the best average
value in binarization criteria and in precision and recall
criteria.
Figure 1. ML Text background segmentation in Y
channel: (a) Original degraded image; (b) Background,
(c) Text, (d) bitonal image.
Figure 2. ML Text background segmentation in RGB
color space: (a) Original degraded image, (b)
Background, (c) Text, (d) Reconstructed image.
Table 1. Evaluation’ scores for two binarization criteria obtained by document distortion type
Show-through effects
ME
RAE
0,114
0,119
0,035
0,038
0,005
0,005
0,047
0,035
0,008
0,009
0,003
0,003
Localized spots
ME
RAE
0,089
0,086
0,061
0,066
0,024
0,023
0,032
0,034
0,029
0,027
0,023
0,000
Uneven background.
ME
RAE
0,160
0,168
0,060
0,060
0,024
0,027
0,042
0,041
0,022
0,019
0,014
0,009
AVE
Niblack
0,122
Sauvola
0,053
Otsu
0,018
Gatos
0,038
Oh et al.
0,019
Proposed Method
0,009
Table 2. Evaluation’ accuracy for precision and recall obtained by document distortion type
746
Show-through effects
Niblack
Sauvola
Otsu
Gatos et al.
Oh et al.
Proposed Method
Precision
53%
99%
96%
75%
100%
100%
Recall
96%
72%
100%
93%
93%
97%
Localized spots
Precision
66%
99%
89%
98%
97%
93%
Recall
94%
63%
99%
81%
84%
93%
Uneven background
Precision
41%
96%
82%
96%
97%
96%
Recall
95%
49%
100%
65%
83%
90%
AVE
Precision
53%
98%
89%
89%
98%
96%
Recall
95%
61%
99%
79%
86%
93%
EM Algorithm”, J. Royal Statistical Soc., Series B
(Methodological), vol. 1, no. 39, pp. 1-38, 1977.
[6] F. Drira, F. Le Bourgeois, H. Emptoz, “Restoring
Ink Bleed-Through Degraded Document Images
Using a Recursive Unsupervised Classification
Technique”, Document Analysis Systems VII ,
Springer Berlin/Heidelberg, 2006, pp.38-49.
[7] U. Garain, T. Paquet, L. Heutte, “On ForegroundBackground Separation in low Quality Document
Images”, International Journal of Document
Analysis 8(1): pp. 47–63, (2006).
[8] B. Gatos, I. Pratikakis, S. J. Perantonis, “Adaptive
degraded document image binarization”, Pattern
Recognition 39 (3), pp. 317–327, 2006.
[9] M. Junker, R. Hoch, and A. Dengel, “On the
Evaluation of Document Analysis Components by
Recall, Precision, and Accuracy”, Proc. Fifth Int’l
Conf. Document Analysis and Recognition, pp.
713-716,1999.
[10] E. Kavallieratou, E. Stamatatos, “Improving the
Quality of Degraded Document Images”, Second
International Workshop on Document Image
Analysis for Libraries, 2006.
[11] W. Niblack, “An introduction to digital image
processing”, Prentice-Hall, Englewood Cliffs, NJ,
pp. 115–116, 1986.
[12] H.-H. Oh, K.-T. Lim, and S.-I. Chien, “An
improved binarization algorithm based on a water
flow model for document image with
inhomogeneous
backgrounds”,
Pattern
Recognition, 38(12): pp.2612–2625, 2005.
[13] N. Otsu, “A threshold selection method from gray
level histogram”, IEEE Transactions in Systems,
Man, and Cybernetics, 1979, vol. 9, pp. 62-66.
[14] J. Sauvola, M. Pietikainen, "Adaptive document
image binarization", Pattern Recognition 33(2), pp.
225–236, 2000.
[15] M. Sezgin, B. Sankur, “Survey over image
thresholding
techniques
and
quantitative
performance evaluation”, J. Electron. Imaging,
13(1), pp.146–165, 2004.
[16] C. L. Tan, R. Cao, P. Shen, “Restoration of
Archival Documents Using a Wavelet Technique”,
IEEE Trans. Pattern Anal. Mach. Intell., 24(10),
pp. 1399-1404, 2002.
6. Conclusion and future work
The developed enhanced text extraction algorithm
operates firstly with a pre-processing step based on
contrast adjustment of the document image. Then, this
image is segmented into text and background
components with the EM based segmentation
algorithm. This algorithm is composed of three steps:
(1) EM initializing, (2) EM estimation, (3) ML
segmentation. According to the evaluated results, our
method performs the best average values compared by
the others methods. These values are about 0.009 mean
errors in enhanced text extraction and accuracy about
96% rate for precision and 93% rate for recall.
7. Acknowledgement
This research is carried out within the framework of
the DAAD project ‘‘In the way of information society’’
and the research cooperation projects between Tunisia
and German. The Authors thanks the National library
of Tunisia and the National Archives of Tunisia [1] for
the access to its large document images database of
Arabic historical documents.
References
[1] National library of Tunisia
http://www.bibliotheque.nat.tn.
[2] G. Agam, G. Bal, G. Frieder, O. Frieder, “Degraded
document image enhancement”, Proc. SPIE 6500, pp.
C1–11, 2007.
[3] W. Boussellaa, A. Zahour, and A. Alimi, “A
methodology
for
the
separation
of
foreground/background in Arabic historical manuscripts
using hybrid methods”, Journal of Universal Computer
Science, 14(2):284–298, 2008.
[4] W. Boussellaa, A. Zhour, B. Taconet, A. Alimi, and A.
Benabdelhafid, “PRAAD: Preprocessing and analysis
tool for Arabic ancient documents”, In 9th International
Conference on Document Analysis and Recognition, vol.
2, pp. 1058–1062, 2007.
[5] A. P. Dempster, N. Laird, and D.B. Rubin,
“Maximum Likelihood from Incomplete Data via the
747
Aymen Bougacha