12th WSEAS International Conference on COMMUNICATIONS, Heraklion, Greece, July 23-25, 2008

A BINARIZATION ALGORITHM FOR HISTORICAL MANUSCRIPTS

VENTZAS, DIMITRIOS 1
Professor
Department of Computer Science Technology & Telecommunications
TEI of Larisa, Greece,
http://www.teilar.gr
1
Correspondence Member

NTOGAS, NIKOLAOS
Post Graduate Student
MSC in Computer Science
Department of Computer Science Technology & Telecommunications &
Staffordshire University Faculty of Computing Engineering &
Technology, TEI of Larisa, Greece,

Abstract: Binarization methods are applied to document images for discriminating the text from the background based
on pure thresholding and filtering combined with image processing algorithms. The proposed binarization procedure
consists of five discrete steps in image processing, for different classes of document images. A refinement technique
enhances further the image quality. Results on Byzantine historical manuscripts are discussed and potential applications
and further research are proposed. The main contribution of this paper is to propose a simple and robust binarization
procedure for pre-filtered historical manuscripts images, and simulation results are also presented.
Keywords: Image processing, document, binarization, denoising, global, local, thresholding.
Introduction: Academic libraries, institutions and
historical museums pile-up or preserve documents in
storage areas. Our work in this paper contributes to
documents safe and efficient preservation in its
original state through out the years and their
unconditional exploitation to researchers, a major issue
for historical documents collections that are poorly
preserved and are prone to degradation processes, see
fig. 1. Documents digitalization, allows access to wider
public, while cultural institutions and heritage
organizations create local or national digital libraries
accessed through the internet. Our work concentrates
on basic techniques used for image enhancement and
restoration, denoising and binarization. The entire
system is implemented in visual environment using
Matlab programming and MathWorks Inc, Image
Processing Toolbox ( MathWorks 2004 ).
Figure 1: Byzantine Manuscript; taking photos of
Byzantine manuscript. Digital photo and Intensity
Histograms
Denoising refers to the removal of noise on the
image ( Sonka et al, 2008 ) and binarization refers to
the conversion of a grayscale image to binary. Both
techniques are basic stages in our image processing of
Byzantine historical manuscripts. Denoising are
filtering methods that eliminate the noise, enhance the
quality of text characters and make the background
texture uniform (Gonzalez et al, (2002), Papamarkos,
(2001)). Binarization (thresholding) converts the
grayscale document image to binary, by changing the
foreground pixels (text characters) to black and
background pixels to white. The paper presents the
need for degraded historical manuscripts images
preservation by binarization implemented by a
procedure based on image preparation, type
classification and refinement of pre-filtered images in
spatial (mean, median and Wiener filter) and frequency
(Butterworth and Gaussian low pass filter) domain.
The work concentrates on text image enhancement and
ISSN: 1790-5117

41

ISBN: 978-960-6766-84-8

12th WSEAS International Conference on COMMUNICATIONS, Heraklion, Greece, July 23-25, 2008

restoration, denoising and binarization using Matlab.

Binarization is obtained by global (Otsu) and local
(Niblack, Sauvola and Bernsen method) thresholding.
The method consists of image acquisition, preparation,
denoising, thresholding and final refinement steps.
Filtering, thresholding and final binarization results on
poor quality text images with various but
systematically classified degradation problems enables

us to compare our methodology to existing binarization

techniques.
Denoising - Filtering:
Denoising methods are
divided in filtering in spatial and frequency domain (
Motwani et al, 2004 ). In our work we implemented
three filters in spatial domain (mean, median and
Wiener filters) with various windows sizes and two
filters in frequency domain (Butterworth and
Gaussian).

Spatial Domain filters

Linear filtering in spatial domain is performed by applying a filter with a weighted sum of neighbouring pixels. The
weight is defined by the filter. Filtering is achieved by convolution and convolution kernel is the correlation kernel
rotated by 180o ( Gonzalez et al, 2002 ).
Mean filter
Median Filter
Wiener filter
The simplest linear filter is the Median filter is a non linear filter. For Wiener filter, known as
mean filter. The intensity of every A a1, a 2, a3,....an
,
and minimum mean square error
pixel in the image is replaced with
filter, is an adaptive linear
the averaged value of intensity of a1 a 2 a3 ..... an R the new value filter, applied to an image
its neighbour pixels. The new of intensity of a pixel (i, j) of an image I is locally, by taking into account
value of intensity of a pixel (i, j) of given by:
the local image variance.
an image I is given by:
When the variance in an
image is large the Wiener
1
a n +1 , if n is odd
I (i , j ) =
I ( x, y )
filter results in light local
2
M ( x , y )N
smoothing, while when the

variance is small, it gives an

where M represents the number of median( A) =
1

local smoothing (
pixels in the neighbourhood N.
a n + a n , if n is even improved
MathWorks,
2004 ).
2

n
+1
2

Frequency Domain Filters

Spatial frequency filtering is implemented by low pass filters which perform Fourier transform. These are smoothing
frequency filters, since they smooth edges and sharp transitions in an icon, such as noise. Low frequencies in the
Fourier transform of an image are responsible for the grey level appearance over the smoothed areas. On the other hand,
high frequencies are responsible for the presence of details, edges ( Ventzas, 1994 ) and noise in the image. There are
two types of low pass filters that are described here: Butterworth and Gaussian.
Butterworth Low Pass Filter
Gaussian Low Pass Filter
Butterworth filter is a low pass filter with transfer Gaussian low pass filter removes effectively the noise but
function:
blurs the image. The mathematical form for twodimensional Gaussian filter is given by

H (u , v ) =

1
1 + [ D (u , v ) / D0 ]2 n

H (u, v) = e D

where D0 is a specific non negative quantity, and D(u,v) it

the distance from point (u,v) to the centre of the frequency
rectangle [ Gonzalez et al, 2002].

( u ,v ) / 2 2

where D(u,v) it the distance from the origin of the Fourier

transform ( Gonzalez et al, 2002 ).

Binarization - Thresholding:
Robust binarization
gives the possibility of a correct extraction of the
sketched line drawing or text from its background. For
the binarization of images many algorithms have been
implemented. Thresholding is a sufficiently accurate
and high processing speed segmentation approach to
monochrome image. This paper describes a modified
logical thresholding method for binarization of
seriously degraded and very poor quality gray-scale
document images. This method can deal with complex
signal-dependent noise and variable background
intensity caused by non uniform illumination, shadow,
smear or smudge and very low contrast images. The
outcome binary image has no obvious loss of useful
information. Firstly, we analyse the clustering and
ISSN: 1790-5117

connection characteristics of the character stroke from

the run-length histogram for selected image regions
and various inhomogeneous gray-scale backgrounds.
Then, we propose a modified logical thresholding
method to extract the binary image adaptively from the
degraded gray-scale document image with complex
and inhomogeneous background. It can adjust the size
of the local area and logical thresholding level
adaptively according to the local run-length histogram
and the local gray-scale inhomogeneity. Our method
can threshold various poor quality gray-scale document
images without the need of any prior knowledge of the
document image and without manual parameter finetuning and without taking into account character
geometric features. It keeps useful information more
42

ISBN: 978-960-6766-84-8

12th WSEAS International Conference on COMMUNICATIONS, Heraklion, Greece, July 23-25, 2008

accurately without over connected and broken strokes

of the characters, and thus, has a wider range of
applications compared with other methods.
A review of the recent research on
binarization is given here. Otsu (1979) suggested a
nonparametric automatic optimal threshold selection
for picture segmentation in order to maximize the
separability of the resultant classes in gray levels.
Kapur et al (1985) and Niblack (1986) introduced a
maximum entropy algorithm that divides the histogram
of the image into two probability distributions, one
representing the objects and one representing the
background. Yanowitz, D.L. and A.M. Bruckstein
(1989) presented a method for finding a threshold
surface, by a gradient map of the image, to point at
well-defined object boundaries for local thresholds.
Solihin, Y. and C.G. Leedham (1999) proposed a new
class of histogram based global thresholding
techniques based on a two stage thresholding approach
of foreground, background, and a fuzzy area. Wu et al
(1999) automatically detected and extracted text in
images from different sources, including video,
newspapers, advertisements, stock certificates,
photographs, using multiscale texture segmentation and
spatial cohesion constraints, by a histogram-based
binarization algorithm.
Yang and Yan (2000)
presented a logical adaptive thresholding method to
binarize seriously degraded and very poor quality grayscale document image with complex signal-dependent
noise. Sauvola et al (2000) presented a new method for
adaptive document image binarization, where the page
is considered as a collection of subcomponents such as
text, background and picture. Zhang et al (2001)
described problems of distorted images of scanned
thick, bound documents, to remove shade and adjust
the warped words, with location, shape and orientation.
Randolph et al (2001) suggested a binary domain
approach that enhances fax documents by directional
filter bank enabling edges and contours in the text
letters to be smoothed appropriately. Leedham, et al
(2003) proposed new thresholding techniques and
compared against existing algorithms. Wu et al (2003)
experimented with a multi-stage global thresholding
approach followed by a local spatial thresholding,
which works well for simple and complex images of
postal envelopes. Fan et al (2003), proposed spatial
correlations of wavelet coefficients by replacing the
thresholding process with a diffusion process for
highly corrupted document images. Bartolo et al (2004)
introduced accurate binarization of a low-level digital
image without user-defined parameters restrictions on
Bernsens algorithm that classifies correctly image of
poor quality, with inhomogeneous paper background,
suitable for text shadow boundaries removal. Gatos et
al (2004) proposed a digital image binarization scheme
for low quality historical documents by five distinct
steps: a pre-processing low-pass Wiener filter, a rough
estimation of foreground regions using Niblack s
approach, a background surface calculation by
interpolating neighbouring background intensities, a
thresholding by combining the calculated background
surface with the original image and finally a postprocessing quality and connectivity step. Sezgin et al
ISSN: 1790-5117

(2004) compared and categorized most image

thresholding methods, such as histogram shape,
measurement space clustering, entropy, object
attributes, spatial correlation and local gray-level
surface for NDT and document images, based on the
combined performance measures. Bieniecki et al
(2005) compared between multi-pass algorithms of
global and local threshold by Bernsen method for
proper pixel neighbourhood window size that fits the
size of image objects. Chen et al (2005) compared
global or local thresholding techniques for degraded
historical documents images and introduced a local
feature thresholding decompose algorithm or document
sub regions using quad-tree decomposition. Gatos et al
(2005) investigated closed cavity regions in the
characters and proposed a segmentation-free
recognition procedure for old handwritten manuscript.
J. He at al (2005) compared alternative binarization
algorithms for historical archive documents recognition
performance in a commercial OCR engine. Kitadai et
al (2005) studied text on stained, damaged, and
degraded wood, to extract characters from badly
blurred or missing ink by binarization discriminant
analysis. Kavallieratou (2005) and Kavallieratou et al
(2005) presented a binarization method of document
images and photos. The method uses the fact that the
pixels that compose the text in a document do not
exceed the 10% of its size. Ashley et al (2007) studied
binarisation algorithms of greyscale images in optical
music recognition by pre-processing, that differ
significantly from non-music documents. Badekas et al
(2007) suggested a system for the binarization of
normal and degraded documents for visualisation and
recognition of text characters by a Kohonen adaptive
neural network. Badekas et al (2007) presented a
technique for the binarization of text blocks in colour
document images that contain text and graphics highly
mixed with the background, based on a colour
reduction. Konidaris et al (2007) searched for
keywords in historical printed documents combining
synthetic data and user feedback by synthetic image
words creation and word segmentation.

43

ISBN: 978-960-6766-84-8

12th WSEAS International Conference on COMMUNICATIONS, Heraklion, Greece, July 23-25, 2008

Global thresholding
The simplest implementation of thresholding is to choose an intensity
value as a threshold level and the values below this threshold become 0
(black) and the values above this threshold become 1 (white). If T is the
global threshold of image f(x,y) and the g(x,y) is the thresholding
image,
then:

1,
g ( x, y ) =
0,

(a)

if

f ( x, y ) T
otherwise

Otsus method
Among the global techniques the most
efficient is Otsus technique [7]. Otsus
method applies clustering analysis to the
grayscale data of input image and models
two clusters of Gaussian distribution of
pixels of the image. The optimal threshold
minimizes the class variance of the two
classes of pixels.

(b )
Figure 2: Global threshold (a) grayscale image (b) T=80 (c) T=150

Common problem in document images are changes in

illumination, or local shadows that are difficult to give
a global threshold, i.e. for the whole image, see fig. 2.
Background Surface Thresholding (BST) computes a
surface of background intensities at every point in the
image and performs adaptive thresholding based on
this result. The surface is estimated by identifying
regions of low resolution text and interpolating
neighbouring background intensities into these regions.
The final threshold is a combination of this surface and

Local thresholding
Niblacks method
Niblacks method is based on the
calculation of the local mean and
of local standard deviation (
Niblack, 1986 ). The threshold in
the pixel (x,y) is decided by the
expression:

a global offset. According to our evaluation BST

produces considerably fewer OCR errors than
Niblacks local average method while it is more
runtime efficient. For small windows noisy background
regions and for larger windows inconsistent stroke
width were produced, i.e. the method is related to
neighbouring features and it is less susceptible to
misclassification of large homogeneous regions.

Sauvolas method
Sauvolas method is an adaptive threshold
method ( Sauvola et al, 2000 ). The
computation of local threshold (i.e., for
each pixel separately) is based on
estimation of local mean and local standard
deviation. The threshold value T(x,y) at the
pixel (x,y) is defined by the relation:

T(x,y)=m(x,y)+k*s(x,y)
where m(x,y) and s(x,y) are the
average and the standard deviation
of a local area respectively. The
size of the window must be large
enough to suppress the noise in the
image, but also small enough to
preserve local details of the image.
A window size 15-by-15 works
efficiently. The value of k is used
to adjust the percentage of total
pixels that belong to foreground
object especially in the boundaries
of the object. A value of k = -0.2
ISSN: 1790-5117

(c)

Bernsens Method
Bernsens method calculates the
local threshold value based in the
mean value of the minimum and
maximum intensities of pixels
within a window ( Papamarkos
(2001) ). If the window is centred
at the pixel (x,y) the threshold for
I(x,y) is defined by:

T ( x, y ) =

s ( x, y )
T ( x, y ) = m( x, y ) 1 + k 1

where k and R are constants with usual

values k = 0.1 and R = 128.

Z max + Z min
2

where Zmax and Zmin

are the
maximum and minimum intensity
of the window. This threshold
works properly only when the
contrast is large. The contrast is
defined
as
C ( x, y ) = Z max Z min . If the
contrast is less that a specific value
k the pixels within the window may
be set to background or to
foreground according to the class
that most suitably describes the

44

ISBN: 978-960-6766-84-8

12th WSEAS International Conference on COMMUNICATIONS, Heraklion, Greece, July 23-25, 2008

produces objects separated well

enough from background.

window.
This
algorithm
is
dependent on k value and also on
the size n of window N-by-N.

classify the documents image types and the applied

methods. There is not a single suitable method that can
be applied to all types of images, or an image type to
suppress the output image artifacts. Before image
processing stages the difficult document images have
been grouped to six distinct categories, i.e. image
conditions:

METHOD APPLICATION AND RESULTS:

Historical manuscripts present information
distortions that are visible in the form of poor quality,
shadows, non uniform illumination, low contrast, large
signal dependent noise, smear and strain spurious point
noise and ragged edges. Since documents images vary
in characteristics compared with common images we
IMAGE CONDITION
GOOD
SPOTS and STAINS
SHADOWS or WRINKLES
TRANSPARENT PAGE
THIN STROKES of PEN:
BROKEN CHARACTERS
COLOURED
CHARACTERS

TECHNICAL COMMENTS
Paper acceptable without spots, stains, smears, aging, brightness degradation
Images with spots, stains, smears or smudges, with less or more background noise.
High humidity and illumination variation caused wrinkles effects and shadows
ink seeking from the other side of page and oily page
Images with thin strokes of pen, i.e. stroke width analysis
Broken
Characters with red ink

The below described discrete stages were applied to

pages acquired from historical books and manuscripts
called Codices, from the Holy Monastery of Dousiko
near Meteora, Trikala, see Table 1. The proposed
binarization method was tested on numerous low
quality digitized historical manuscripts and digital text

images. The proposed method is robust to many source

type-related degradations illumination variations and
produces images with very little noise and consistent
stroke width, with minimal prior knowledge of the
document image.

Historical manuscripts digital processing algorithm

The method applied consists of five stages that are described schematically as in fig. 3. These stages are described
analytically as follows:
Stage 1: Image acquisition
The images are acquired by a digital camera with high
resolution ratio, stored in computer and converted to
compressed file formats with low storage requirements.
The digital camera used is a CMOS technology SLR
CANON 1.8II [Canon, 2007 ] with a 50mm lens. The
resolution of the camera was the in the range of 4,368 by
2,904 pixels. The sensors size is 24 mm by 36mm. JPEG
files were used due to lower size and computation time,
especially for old books reproduction, since RAW or TIFF
files are memory and CPU time consuming. MATLAB
IP Toolbox [2004] supports all these format. We worked
with a sufficient number of document images and applied
various filters, extended to many types, sizes, windows,
etc, in order to explore denoising procedures.
Stage 2: Image preparation
The preparation of the image before filtering and
binarization consists of:
a. conversion from RAW to TIFF / JPEG
b. cropping
c. conversion to gray scale.
Stage 3: Denoising
The best denoising filter is derived after comparison of
various types of filters and more specifically the: Mean,
Median and Wiener filters in spatial domain and Gaussian
and Butterworth filters in the frequency domain.
ISSN: 1790-5117

45

ISBN: 978-960-6766-84-8

12th WSEAS International Conference on COMMUNICATIONS, Heraklion, Greece, July 23-25, 2008

Stage 4: Thresholding
Thresholding are applied by global (Otsus) and local
(Niblack, Sauvola, Bernsen) thresholding techniques on
previous stage resulting filtered images.
Stage 5: Refinement
A refinement procedure, based on erosion and dilation, is
applied on the binarized image, such that the obtained
image has its characteristics further clarified in the texture
and foreground compared with the background area.
Figure 3: Proposed method stages

Documents with poor

quality paper

Ink wet characters

visible both sides

The brightness of the aging

paper colours

Poor contrast between

foreground and background

Dirty documents with spots,

High humidity cause
stains, smears or smudges.
wrinkles to the paper
Table 1: Documents problems classification

Broken characters, light

handwriting.

Problems due to image

acquisition, illumination,

Denoising results
The filters applied are Mean, Median, Wiener, Gaussian and Butterworth ones. The application of each filter with
variable sizes of window, explored all possible denoising results:
a. Filtering improved the quality of the image, thus preparing it for binarization, see Table 2.
b. Spatial domain filtering using the Mean, Median and especially Wiener filters.
c. Frequency domain filtering using the Butterworth and Gaussian low pass filters.
d. The paper condition is an unexpected factor.
e. The document filtering is a preliminary stage for optical character recognition.

original grayscale image

ISSN: 1790-5117

after mean filter 5-by-5

46

after median filter 7-by-7

ISBN: 978-960-6766-84-8

12th WSEAS International Conference on COMMUNICATIONS, Heraklion, Greece, July 23-25, 2008

after max filter 5-by-5

after min filter 5-by-5

after wiener filter 7-by-7

ideal filter, radius value of 100

Butterworth filter, radius value of 100

Table 2: Filtering on documents

Gaussian filter, radius value of 100

Documents Image Category / Binarization

GOOD CONDITION
SPOTS and STAINS
SHADOWS or WRINKLES
INK SEEKING from other SIDE
THIN STROKES of PEN
RED coloured CHARACTERS

Bernsen
BEST
BAD
BAD
BAD
BAD
BEST

Niblack
BEST
GOOD
BEST
GOOD
BAD
GOOD

Otsu
BEST
BAD
BAD
BAD
GOOD
GOOD

Sauvola
BEST
BEST
BEST
BEST
BAD
GOOD

Table 3: Results from combination of Wiener filter 5-by-5 with binarization methods for each image category
results to almost all of the specified image categories,
see Table 3. Eikvils and Parkers binarization methods
were not included into our comparison, but
thresholding techniques review indicated bad text
detection recall ranking.

Document Image Thresholding:

Binarization is applied to all document image
categories. Image focusing, sharpness and clarification
on the handwritten characters, and texture was
compared with the original ones, see Table 3. The
binarization, based on adaptive global / local
thresholding, is an efficient step in image digitalisation
and works best on high resolution images. The JPEG
produced file formats needed the least computational
effort to be processed. Previous research classifies
threshold methods in two categories, global or local
(adaptive) threshold to separate foreground from
background objects. We have chosen four binarization
methods (Otsus, Niblacks, Sauvolas and Bernsens)
and looked into the results of each one in a range of
document images taken from the Holy Monastery of
Dousiko at Meteora - Trikala, Thessaly, Greece.
Thresholding techniques applied to non-similar types
of documents images revealed the hidden capabilities
of combined filtering and binarization of the
categorized types of images. The conclusions of the
suitable methods for each type of document image are
given in Table 3. The application of Wiener filtering
with a window size of 5-by-5 has produced the best

ISSN: 1790-5117

The most of the Byzantine manuscripts images taken,

belong to the category of images with spots and
stains and with red coloured characters. Niblacks
and Sauvolas methods produce efficient results in
almost all categories except the category of thin strokes
of pen in which global Otsus method has the best
results on the produced binary images. In Byzantine
manuscripts with characters and drawings with red ink
the Bernsens method produced best results. The
proposed algorithms were tested with images including
different types of document components and
degradations, against well-known thresholding
techniques. The results show that the method performs
well in each case qualitatively and quantitatively and
has superior performance compared to other techniques
tested.

47

ISBN: 978-960-6766-84-8

12th WSEAS International Conference on COMMUNICATIONS, Heraklion, Greece, July 23-25, 2008

original document image with spot

original document image with

shadow caused by bad illumination

original image with ink seeking from

other side

binary image with Sauvolas method

after Wiener filter 5-by-5

binary image with Sauvolas method

after Wiener filter 5-by-5

binary image with Sauvolas method

after Wiener filter 5-by-5

original Document image with thin

strokes of pen

original Document image with

characters with red ink

detail of image with black dots before /

after refinement step

binary image with Sauvolas method

after Wiener filter 5-by-5

binary image with

Otsus method
after Wiener filter 5-by-5

detail of image with holes on

characters before/after refinement step

Table 4:

Document image before and after binarization

Document Image Final refinement

The post-binarization refinement improves the appearance of the binary images and text readability, especially in
documents with red ink characters and line gaps or holes. Refinement consists of the successive erosion followed by
dilation operation, and opening on the negative image to remove the remaining black pixels that not belong to the text
characters/ Results of post-binarization refinement are shown in Tables 4, 5 and 6. The refinement significantly
improves the image quality for most of the image types since it clarifies the background area, by clearing up the
appearance of the text, especially when it is applied to images after Niblacks and Sauvolas binarization methods.

ISSN: 1790-5117

48

ISBN: 978-960-6766-84-8

12th WSEAS International Conference on COMMUNICATIONS, Heraklion, Greece, July 23-25, 2008

a. binary image

b. binary image after erosion

(a)

(b)

Table 5: Final Refinement (a) binary image (b) Removing 50 connected pixel
FUTURE WORK:
Potential application fields
include the automation of the combined binarizationfiltering procedure by a neural network and the
extension of the method to a wider area of documental
or non-documental images. Parallel computational
machines and perceptual optical processing techniques
should further increase the methods efficiency. The
application of filtering as a preliminary stage for the
binarization of the document image promises a great
improvement on the quality of the final images. Other
filter schemes in the preliminary stage of digital preprocessing can be investigated. By converting
historical documents and old newspapers (which have
been degraded or partly damaged) to digital formats we
preserve them, in the form of the original document,
for future reproduction. By digitalization and storing of
copies of old books and historical manuscripts, we can
store electronically entire libraries to preserve
historical manuscripts. Such a text images storage
environment and data base is proposed for further
research.

image preparation, image type classification in six

categories according to image condition. The estimated
results for each class of images and each method are
further enhanced by an innovative image refinement
technique and a formulation of a class proper method.
The potential of the applications of preliminary
processing to document images, by adjusting the
binarization method according to the category of the
image, becomes reasonably estimated taking into
account the improvement in the quality of the image as
a whole and the increased readability of the texture.
The results have shown improved image quality for the
six categories of document images which were
described by their separate characteristics. It has turned
out that our methodology performs better compared to
current state-of- the-art adaptive thresholding
techniques and it is robust for document images
comparing with other thresholding methods based on
connectivity and background analysis and might
succeed in a wider range of applications.
Acknowledgements:
We would like to thank the
Computer Science & Telecommunications Technology
Dpt, TEI Larissa, Greece, the Monks of Holy
Monastery of Dousiko, near Meteora, Greece, and
Professor N. Papamarkos.

CONCLUSION:
No algorithm works well for
all types of images but some work better than others
for particular types of images suggesting that improved
performance can be obtained by selection or
combination of appropriate algorithm for the type of
document image under investigation. We have
described algorithms that utilize spatial structure,
global and local features or both. Many algorithms
require extensive preprocessing steps in order to obtain
useful data to work with because document image and
data mining classification techniques is still in infancy.
The purpose of our work on text image binarization
was to introduce an innovative procedure for digital
image acquisition of historical documents based on
ISSN: 1790-5117

c. binary image after dilation

Table 6: Steps of refinement
stage

References
1.
Ashley John, Burgoyne, Laurent, Pugin, Greg
Eustace, Ichiro Fujinaga, A comparative survey of
image binarisation algorithms for optical recognition
on degraded musical sources, ISMIR2007, p509
2.
Badekas Efthimios, Nikos Nikolaou, Nikos
Papamarkos, Text Binarization in Color Documents,
Electrical and Computer Engineering Dpt, Image

49

ISBN: 978-960-6766-84-8

12th WSEAS International Conference on COMMUNICATIONS, Heraklion, Greece, July 23-25, 2008

Processing and Multimedia, Lab, Wiley Periodicals,

Inc, Vol. 16, 262274 (2007)
3.
Badekas E., N. Papamarkos, Optimal
combination of document binarization techniques using
a self-organizing map neural network, Engineering
Applications of Artificial Intelligence 20 (2007) 1124,
Xanthi, Greece
4.
Badekas E, N. Papamarkos, Document
binarization using Kohonen, IET Image Process., 2007,
1, pp. 6785
5.
Bartolo A, K. Camilleri1, J.Borg &P. Farrugia,
Adaptation of Brensen Thresholding Algorithm for
Sketched Line Drawings, Eurographics Workshop on
Sketch-Based Interfaces & Modeling, Hughes & Jorge
Ed, pp. 81-90, 2004.
6.
http://www.canoneurope.com/For_Home/Product_Finder/Cameras/Digit
al_SLR/EOS_5D/index.asp, Canon Europe Ltd, Digital
SRL Camera 5D, 2007
7.
Chen Y. and G. Leedham, Decompose
algorithm for thresholding degraded historical
document images, IEE Proc.-Vis. Image Signal
Process., Vol. 152, No. 6, December 2005
8.
Gatos B, I. Pratikakis . S. J. Perantonis ,
Locating Text in Historical Collection Manuscripts,
Lecture Notes in Computer Science Springer, Vol 3025
Methods and Applications of AI, 2004, ISBN 978-3540-21937-8, p 476-485
9.
Gatos B , I. Pratikakis, S. Perantonis, An
Adaptive Binarization Technique for Low Quality
Historical Documents, Springer, ISSN 0302-9743,
Vol3163, Document Analysis Systems VI 2004
ISBN978-3-540-23060-1p. 102
10.
B. Gatos et al, An efficient segmentation-free
approach to assist old Greek handwritten manuscript
OCR, Journal on Pattern Analysis & Applications,
Springer London ISSN1433-7541, Issue Vol 8, N. 4 /
Feb, 2006, p. 305
11.
Gonzalez, C.R. and E.R. Woods, "Digital
Image Processing", 2nd edition, 2002, Prentice-Hall
Inc., pp 75-278.
12.
He J, Q. D. M. Do, A. C. Downton and J. H.
Kim, A Comparison of Binarization Methods for
Historical Archive Documents, Proc of the 8th
International Conf on Document Analysis and
Recognition (ICDAR05) 2005 IEEE
13.
Kapur, J., P.K. Sahoo, and A.K.C. Wong, "A
new method for gray-level picture. Thresholding using
the Entropy of the Histogram", Computer Vision
Graphics and Image Processing, 1985, vol. 29, pp. 273285.
14.
Kavallieratou E. and H. Antonopoulou,
Cleaning and Enhancing Historical Document Images,
ACIVS 2005, LNCS 3708, pp. 681 688, 2005.
Springer-Verlag 2005
15.
Kavallieratou
Ergina,
A
Binarization
Algorithm specialized on Document Images and
Photos, Document Analysis and Recognition, 2005.
Proceedings. 8th Int Conf ICDAR, 29 Aug.-1 Sept.
2005. p. 463, Vol. 1
16.
Kitadai Akihito, Kei Saito et al, Design and
Prototype of a Support System for Archeologists to
Decode Scripts on Mokkan, Proc. 13th Conference of
ISSN: 1790-5117

the International Graphonomics Society (IGS), Salerno,

Italy, pp.54-58, (2005.6)
17.
Konidaris T et al Keyword-guided word
spotting in historical printed documents using synthetic
data and user feedback, Int Journal on Document
Analysis and Recognition, Springer 2007 Vol. 9, N 2-4,
ISSN1433-2833, p 167-177
18.
Leedham G et al "Comparison of
Thresholding Algorithms for Text/Background
Segmentation in Difficult Document Images", Proc of
7th International Conference on Document Analysis
and Recognition, ICDAR 2003, IEEE
19.
Lixin Fan, Liying Fan and ChewLim Tan,
Wavelet Diffusion for Document Image Denoising,
Proceedings of the Seventh International Conference
on Document Analysis and Recognition (ICDAR), 07695-1960-1/03, 2003 IEEE
20.
The MathWorks Inc., "Image Processing
Toolbox - User Guide", User Manual, Matlab, 2004.
21.
Motwani, M. C, M. C. Gadiya, Rakhi C.
Motwani, Frederick C. Harris, Jr, "Survey of Image
Denoising Techniques", Proceedings of GSPx, 2004,
Santa Clara Convention Center, Santa Clara, CA, 2004
22.
Niblack, W., "An Indroduction to Digital
Image Processing" 1986, Prentice Hall, pp.115-116.
23.
Otsu, N., "A threshold selection method from
gray-level histograms", IEEE Trans. Systems, Man,
and Cybernetics, 1979, vol. 9, no. 1, pp 62-66.
24.
Papamarkos, N., "Digital Processing and
Image Analysis", 2001, Athens, Giourdas
25.
Randolph, T. Smith, M, Enhancement of fax
documents using a binary angular representation,
Proceedings of, Int Symp on Intelligent Multimedia,
Video and Speech Processing, 2001. Hong Kong,
China, ISBN: 962-85766-2-3
26.
Sauvola, J. and M. Pietikainen, "Adaptive
document image binarization", Pattern Recognition 33,
2000, pp. 225-236.
27.
Sezgin Mehmet, Bulent Sankur, Survey over
image thresholding techniques and quantitative
performance evaluation, Journal of Electronic Imaging
13(1), 146 165 (January 2004).
28.
Solihin, Y. and C.G. Leedham, "Integral
Ratio: A New Class of Global Thresholding
Techniques for Handwriting images", IEEE Trans. on
PAMI, 1999, vol 21, (no 8), pp. 761-768.
29.
Sonka, M., V. Hlavac, and R. Boyle, "Image
Processing, Analysis and Machine Vision", Toronto :
Thompson Learning, 2008.
30.
Ventzas, D., "Edge Detection Techniques in
the Industry", Advances in Modelling & Analysis,
Series B, vol. 29, No. 2, pp. 57-64, Winter 1993-1994,
AMSE, 1994.
31.
W. Bieniecki and S. Grabowski, Multi-pass
approach to adaptive thresholding based image
segmentation, Computer Engineering Dpt., TU Lodz,
Poland, CADSM2005, February 23, 2005, LvivSlavske, Ukraine
32.
Wu V, R Manmatha, EM Riseman,
TextFinder: An Automatic System to Detect and
Recognize Text In Images IEEE Transactions on
Pattern Analysis and Machine Intelligence, November
1999 (Vol. 21, No. 11) pp. 1224-1229
50

ISBN: 978-960-6766-84-8

12th WSEAS International Conference on COMMUNICATIONS, Heraklion, Greece, July 23-25, 2008

33.
Wu Sue, Adnan Amin,
Automatic
Thresholding of Gray-level Using Multi-stage
Approach, Proceedings of the 7th International
Conference on Document Analysis and Recognition
(ICDAR 2003), 0-7695-1960-1/03, 2003 IEEE
34.
Yanowitz, D.L. and A.M. Bruckstein, "A new
Method for image segmentation", Computer Vision
Graphics and Image Processing, 1989, vol.46, (no 1),
pp.82-95.
35.
Yibing Yang and Hong Yan, An adaptive
logical method for binarization of degraded document
images Pattern Recognition, Volume 33, Issue 5, May
2000, Pages 787-807, Pattern Recognition Society,
Elsevier Science
36.
Zhang, Z. and C. Tan, "Restoration of images
scanned from thick bound documents", in proceedings
of International Conference on Image Processing 2001,
Volume 1, 2001, pp. 1074-1077

N. Ntogas received the diploma degree in Applied Informatics at the Economic University
of Athens, Greece, in 1989. He finished the School of Pedagogical and Technological
Education of Thessalonica, Greece, in 1998. He attained an MSc in Computer Science at
the Staffordshire University, United Kingdom in cooperation with TEI of Larissa, Greece
in 2007. He worked as Computer Programmer and Analyst and as responsible for the
deployment of the computerization of Municipality of Trikala, Greece. Since 1998 he
works as a teacher in the department of Computer Laboratories of 1st Technical Institution
of Trikala. His research interests include, Image Processing, historical documents
acquisition, etc. This work is part of his MSc Dissertation by Research.
Email:
[email protected].
Dr. D. E. Ventzas (1956) SMISA is Electronic Engineer and Professor of Technological
Institute of Larissa, Greece. He owns an MSc in Control Eng and a PhD in Microprocessor
based Instrumentation from Bradford University, Yorkshire, UK. He was Instrument and
Systems Engineer in Hellenic Aspropyrgos Refinery SA, Athens. His research interests
lies in Signal and Image Processing, Process Control and Instrumentation, Biomedical
Engineering and Computer Tools for Instrumentation. He is the author of many research
and review papers in English and books in Greek. He supervised many undergraduate
diploma thesis and postgraduate dissertations.
Email:
[email protected], [email protected]

ISSN: 1790-5117

51

ISBN: 978-960-6766-84-8