EXTRACTION OF OBJECTS AND PAGE SEGMENTATION OF
COMPOSITE DOCUMENTS WITH NON-UNIFORM
BACKGROUND
Yasser Alginahi, Maher Sid-Ahmed, Majid Ahmadi
Univerisyt of Windsor, 401 Sunset Ave, Windsor, Ontario,Canada
Keywords: Statistical features, Multi-Layer Neural Networks, Non-uniform background.
Abstract: In designing page segmentation systems for documents with complex background and poor illumination,
separating the background from the objects (text and images) is very crucial for the success of such system.
The new local based neural binarization technique developed by the authors will be used to extract the
objects from document images with complex backgrounds. This algorithm uses statistical and textural
feature measures to obtain a feature vector for each pixel from a window of size
)12()12(
+
×+ nn
,
where
1≥n
. These features provide a local understanding of pixels from their neighbourhoods making it
easier to classify each pixel into its proper class. A Multi-Layer Perceptron Neural Network (MLP NN) is
then used to classify each pixel in the image. The results of thresholding are then passed to a block
segmentation stage. The block segmentation technique developed is a feature-based method that uses a
Neural Network classifier to automatically segment and classify the image contents into text and halftone
images. The results of page segmentation are then ready to be passed into an OCR system that will convert
the text image into a format the can be stored and modified
.
1 INTRODUCTION
Document image analysis is an important area of
research in image processing, pattern recognition
and computer vision. The goal of our research is to
process grey level document images with complex
backgrounds, bad illumination and poor contrast.
The motivation behind most of the applications of
off-line text recognition is to convert data from the
conventional media into electronic media. In this
paper, a document segmentation system is presented
to transfer grey level composite images with
complex backgrounds and poor illumination into
electronic format that is suitable for efficient storage
retrieval and interpretation. Such applications are
bank cheques, security documents and form
processing.
There are many threshold selection schemes
published in the literature, and selecting an
appropriate one can be a difficult task. Thresholding
of documents can be categorized into two main
classes: global and local thresholding. Global
thresholding techniques use a single threshold value;
on the other hand, local thresholding compute a
separate threshold based on the neighbourhood of
the pixels. (Sahoo et al., 1998) showed that the
Otsu’s (Otsu, 1979) class separability thresholding
method is the best global thresholding method. In
(Trier and Jain, 1995), Trier and Jain showed the
Niblack’s (Niblack, 1986) method to be the best
local thresholding method compared to other
methods. Few methods used NNs in thresholding
grey scale images into two levels. The technique
proposed by Koker and Sari (Koker and Sari, 2003),
uses NNs to select a global threshold value for an
industrial vision system based on the histogram of
the image. The method developed by Papamarkos
(Papamarkos, 2001) uses Self Organizing Feature
Maps (SOFM) to define two bi-level classes. Then,
the contents of these classes are used with the fuzzy
C-mean algorithm to reduce the character blurring
effect. Both methods are not suitable for
thresholding composite images with complex
backgrounds. In this paper, a new threshold selection
algorithm is proposed which handles images with
non-uniform and complex backgrounds. The new
method uses a MLP NN with statistical and textural
feature measures as inputs to the network.
344
Alginahi Y., Sid-Ahmed M. and Ahmadi M. (2005).
EXTRACTION OF OBJECTS AND PAGE SEGMENTATION OF COMPOSITE DOCUMENTS WITH NON-UNIFORM BACKGROUND.
In Proceedings of the Second International Conference on Informatics in Control, Automation and Robotics - Signal Processing, Systems Modeling and
Control, pages 344-347
DOI: 10.5220/0001167903440347
Copyright
c
 SciTePress

2 STATISTICAL FEATURE
MEASURE
The new NN local thresholding method, takes
advantage of the document texture characteristics by
considering the statistical texture descriptors in a
neighbourhood of pixels. The statistical texture
descriptors used were defined over a window of size
)12()12( +×+ nn centered at ),( ji ,
where
2=n . The features extracted are as follows:
2.1 Actual Pixel Value
The first feature extracted is the center pixel
),( jip in the window.
2.2 Mean
The mean,
ij
µ
, of the pixel values, in the defined
window.
∑∑
+
−=
+
−=
+
=
ni
nix
nj
njy
ij
yxp
n
),(
)12(
1
2
µ (1)
2.3 Standard deviation
The Standard deviation,
ij
σ
, is the estimate of the
mean square deviation of grey pixel value,
),( yxp
,
from its mean value,
ij
µ
.
∑∑
+
−=
+
−=
−
+
=
ni
nix
nj
njy
ij
yxp
n
2
)]),([
)12(
1
µσ (2)
2.4 Skewness
Skewness,
ij
S , characterizes the degree of
asymmetry of a pixel distribution around its mean.
3
2
),(
)12(
1
∑∑
+
−=
+
−=
⎥
⎥
⎦
⎤
⎢
⎢
⎣
⎡
−
+
=
ni
nix
nj
njy
ij
ij
ij
yxp
n
S
σ
µ
(3)
2.5 Kurtosis
Kurtosis,
ij
K , measures the relative peakness or
flatness of a distribution relative to normal
distribution.
3
),(
)12(
1
4
2
−
⎪
⎭
⎪
⎬
⎫
⎪
⎩
⎪
⎨
⎧
⎥
⎥
⎦
⎤
⎢
⎢
⎣
⎡
−
+
=
∑∑
+
−=
+
−=
ni
nix
nj
njy
ij
ij
ij
yxp
n
k
σ
µ
(4)
The –3 term makes the value zero for a normal
distribution.
2.6 Entropy
Entropy,
ij
h , describes the distribution variation in a
region. Entropy of pixel
),( jiP
can be calculated
as
∑
−
=
−=
1
0
log
L
k
kkij
pph (5)
Where
k
p is the probability of the k
th
grey level,
which can be calculated as
2
)12/( +nz
k
,
k
z is the
total number of pixels with the k
th
grey level and L is
the total number of grey levels in the window.
2.7 Relative Smoothness
Relative Smoothness,
ij
R , is a measure of grey-level
contrast.
2
1
1
1
ij
ij
R
σ +
−=
(6)
2.8 Uniformity
Uniformity,
ij
U , is a texture measure based on
histogram and is defined as:
∑
−
=
=
1
0
2
L
k
kij
PU (7)
Before computing any of the descriptive texture
features above, the pixel values of the image were
normalized by dividing each pixel by 255 in order to
achieve computation consistency.
3 NN LOCAL THRESHOLDING
TECHNIQUE USING MLP
The new local NN thresholding technique uses a
MLP neural network (Sid-Ahmed, 1995) to classify
document images into background and foreground.
Random representative sample pixels from several
images with different complex backgrounds and
degraded images were selected as training sets for
the NN.
EXTRACTION OF OBJECTS AND PAGE SEGMENTATION OF COMPOSITE DOCUMENTS WITH
NON-UNIFORM BACKGROUND
345

Figure 1: Example results showing the thresholding of the Otsu’s, Niblack’s and proposed methods
The features related to the neighborhood pixels in
the neighborhood window, were calculated and the
actual output value for the pixel was recorded, 1 for
background and 0 for foreground. These feature
values were then fed into the MLP NN to train the
network. The MLP NN is trained by supervised
learning using the iterative back-propagation
algorithm, which minimizes the mean square error
between the network's output and the desired output
for all input patterns. Once the NN has been trained,
the weights are used in the classification phase.
During classification, image data feature vectors
extracted from each pixel in the image are fed into
the network that performs classification by assigning
a class number, either 1 or 0, for each pixel. Figure
1 shows the results of the Otsu’s, Niblack’s and the
proposed NN local thresholding techniques for an
example image.
4 NEURAL SEGMENTATION
4.1 Introduction
The RLSA (Wong and Casey, 1982) technique is one
of the most widely used top-down segmentation
algorithms. It is used on binary images to classify
images into text and halftone images. By linking
together the neighbouring black pixels that are
within a certain threshold to form blocks of text
and/or images, this method is applied row-by-row
(horizontal smearing) and column-by-column
(vertical smearing), then both results are combined
in a logical AND operation. From the RLSA results,
single blocks of text lines and images are produced.
Wahl et al. in (
Wong and Casey, 1982) used a
statistical classifier to classify these blocks, but in
this work a NN classifier will be used to
automatically segment the contents of the images.
Figure 2 shows the thresholding, and segmentation
process.
4.2 Block Labelling
Labels have to be assigned to different blocks to
identify each block separately to be used in the
feature extraction step. All connected pixels must
have the same label. A Local Neighbourhood
Algorithm (Rosenfeld and Kak, 1976) is used which
scans the image horizontally until it hits the first
pixel then a fire is set at this point that propagates to
all 8-neigborhood of the current pixels giving them
the same label.
4.3 Feature Extraction
After block labeling, the coordinates of each block
are known. The next step is to extract features.
Geometrical and statistical features were extracted
from each block of text and images. The following
features were extracted: the height (
i
H ), mean pixel
value (
i
µ
), standard deviation of pixels (
i
σ
) and
black pixel count (
i
BC ) calculated from the binary
image where
i is the label corresponding to each
block.
4.4 Block Classification
The final step in page segmentation is the
classification of each block into its proper class. In
Whal (
Wong and Casey, 1982), a statistical classifier
was used with parameters that were not optimized
and are application dependent. In this paper, a NN
classifier using MLP is used with the features
extracted fed into the NN to produce the proper
classification between text and images. The NN
contained three layers, with 4 nodes at the input
layer, 7 nodes in the hidden layer and one node at
the output layer. The results of block classification
are shown in the example of Figure 2.
ICINCO 2005 - SIGNAL PROCESSING, SYSTEMS MODELING AND CONTROL
346

Figure 2: Example showing the process of thresholding and page segmentation.
5 RESULTS
The new local NN thresholding method was tested
on over hundred images and the results obtained
outperformed other thresholding selection schemes
as seen from Figure 1. This new neural based
technique calculates an optimal threshold value for
each pixel when passed through the NN making it a
suitable method to be used in character recognition
applications. The proposed neural based page
segmentation method which uses the RLSA to
produce blocks of text and images provided good
separation between different contents of a document
image as seen in Figure 2. These results in most
cases do not need any post processing since the
thresholding produced binary images that are free of
noise. The results are then ready to be fed into an
OCR system to extract the text from the image.
Testing this system on hundred images provided a
98.6% rate in separating different contents of an
image into text and halftone images.
6 CONCLUSION
The proposed thresholding technique outperformed
other techniques and the results produced are free of
noise making it a good choice to use in extracting
objects from image documents with non-uniform
background. The Neural based page segmentation
provided a good separation between text/halftone
image contents of an image document
and can be extended to include other contents in
image documents. These results requires no or
minimal post processing to be used in further stages
of document analysis to guarantee the success of the
character recognition system in providing higher
recognition rate.
REFERENCES
PK. Sahoo, S. Soltani, AKC Wong, 1988. “A Survey of
thresholding techniques”, Computer Vision, Graphics
and Image Procs, Vol. 41, pp. 233-260.
N. Otsu, 1979. “A Threshold Selection Method from
Grey-Level Histograms”, IEEE Trans. On System Man
and Cybernetics, vol. 9, no. 1, pp. 62-69.
O. D. Trier and A. K. Jain, 1995. “Goal-Directed
Evaluation of Binarization Methods”, IEEE Trans
Pattern Analysis and Machine Intelligence, Vol. 17,
No. 12, pp. 1191-1201.
W. Niblack, 1986. “An Introduction to Digital Image
Processing” Englewood Cliffs, Prentice Hall, NJ. pp
115-116.
R. Koker and Y. Sari, 2003. “Neural Network Based
Automatic Threshold Selection For an Industrial
Vision System,” Proc. Int. Conf. On Signal
Processing. pp. 523-525.
N. Papamarkos, 2001. “A Technique for Fuzzy Document
Binarization,” Procs. Of the ACM Symposium on
Document Engineering, pp.152-156.
M. A. Sid-Ahmed, 1995. “Image Processing: theory,
algorithms, and architectures”, McGraw-Hill, pp.
313-375.
Wahl, K. Wong and R. Casey, 1982. “Block Segmentation
and Text Extraction in Mixed Text/Image
Documents,” Computer Vision, Graphics and Image
Processing, Vol. 20, pp. 375-390.
A. Rosenfeld and A. C. Kak, 1976. “Digital Picture
Processing,” Academic Press, NY, pp. 347-348.
EXTRACTION OF OBJECTS AND PAGE SEGMENTATION OF COMPOSITE DOCUMENTS WITH
NON-UNIFORM BACKGROUND
347