Annals of Fuzzy Mathematics and Informatics

Volume x, No. x, (Month 201y), pp. 1–xx @FMI

ISSN: 2093–9310 (print version) c Kyung Moon Sa Co.
ISSN: 2287–6235 (electronic version) http://www.kyungmoon.com
http://www.afmi.or.kr

Text extraction from low-contrast non-uniform

illuminated document images using mathematical
morphology and fuzzy operator

(Mrs)Sharmistha Bhattacharya (Halder), Md. Atikul Islam

Received 29 May 2013; Accepted 11 August 2013

Abstract. Document image binarization has been studied for decades
and many practical binarization techniques have been proposed for dif-
ferent kinds of degraded document images. Document images captured
with electronic devices in low-contrast and non-uniform lightening condi-
tions often cause serious problems in document layout analysis and optical
character recognition systems (OCR). In this paper a new efficient method
for extraction of text from low-contrast non-uniform illuminated document
images is proposed. Morphological operator Bottom-Hat Transform and
fuzzy intensification operator are combined to extract text from degraded
document images. Experiments are performed using gray-scale and colour
document images. Results of proposed method and other methods are
shown and discussed.

2010 AMS Classification: 94DXX, 68U10.

Keywords: Bottom-Hat Transformation; Contrast Removal Fuzzy Operator;
Non-uniform illuminated, document image, Binarization.
Corresponding Author: Md.Atikul Islam([email protected] )

1. Introduction
D ocument image binarization to extract text stroke pixels from the background
of document image is an active research area and has been studied for decades. Many
practical binarization techniques have been proposed. Sezgin and Sankar[41]classified
the different techniques into six categories: histogram based methods, clustering
based methods, entropy based methods, forground attribute based methods, spatial
binarization method. Wen, Li and Sun [52] again divide these methods into three
categories: clustering based methods [7, 8, 17, 31], threshold-based methods and
hybrid methods. The threshold methods categories into two groups: one is global
threshold methods [1, 11, 18, 19, 25, 42] which assign a single threshold for the entire
 Md.Atikul Islam et al./Ann. Fuzzy Math. Inform. x (201y), No. x, xx–xx

image. However global thresholding methods are usually not suitable for degraded
document images, because they do not have a clear bimodal pattern that separates
foreground text and background. So the other method local thresholding are better
approaches for degraded document images with non-uniform background and fore-
ground distribution. The local threshold method based on the information contained
in the neighbourhood of each pixel, or in the region of the image. One drawback
of local thresholding approaches is that the thresholding performance depends on
the window size and hence the character stroke width. In Bernsen’s method [4], the
threshold is a function of the highest and lowest gray values. Whereas Niblack [27]
and Sauvola [43] threshold value depends on mean and standard deviation of the gray
scale. Yanowitz and Bruckstein (YB) [53] were classically locally adaptive methods.
Su, Lu and Tan [44] based on local maximum and minimum filters. By shading
estimation and compensation Lu and Tan [21]proposed a binarization method for
badly illuminated document images. Based on edge information [9, 22, 36, 37], based
on gradient information [6, 32, 49] are locally adaptive methods. Kim et al. [20]
proposed an effective algorithm based on a water flow model for binarization of de-
graded document images. They consider document image as a three dimensional
terrain. Binarization of images was obtained by thresholding the amount of water
filled in valley’s (as text). An improved version of Kim’s method was proposed by
Oh et al [30]. Valizadeh and Kabir [50] proposed an adaptive water flow model
for binarization of degraded document images. They consider image surface as a
three-dimensional terrain and pour water on it. The water finds the valleys and fills
them. There algorithm controls the rainfall process, pouring the water, in such a
way that the water fills up to half of the valley’s depth. After stopping the rainfall,
each wet region represents one character or a noisy component. To segment each
character labeled the wet regions and regarded them as blobs. They used Multilayer
Perceptron to label each blob as either text or non-text. Some widely used hybrid
models were Gonman[14], Liu[23], Moghaddam [26], Valizadeh [51]. A neuro-fuzzy
techniqe was given in [33]. In 2011 Bataineh, Abdulla and Omar [2] was proposed an
adaptive local binarization method for document images based on thresholding and
dynamic windows. Blayvas [5] method based on threshold surface, threshold surface
was determined by successive over relaxation as the solution of Laplace equation.
In recent Wen [52] proposed a effective binarization method for non-uniform illumi-
nated document images. Images was decomposed by the Curvelet transformation
and the curvelet coefficient were enhanced by non linear function then Otsu’s method
was combined for binarization. A learning built rules for document images produced
by cameras was given by [10]. Image was divided into several regions and decided by
learning process how to binarize each region. We may mention some other kind of
methods where text was extracted from degraded document images. Morphological-
based [12, 13, 15, 28, 38, 45], fuzzy logic based Sattar and Tay [46], Parbathi [34, 39]
based on an intensification operator on first and second type of intuitionistic fuzzy
sets. Leung [24] in 2005 presented a method based on Generalized Fuzzy Operator,
pre-processed with HE and POSHE for low contrast and low illumination document
images. The thresholding of document images is still an unsolved problem due to
different types of document degradations, such as uneven illumination, image con-
trast variation, bleeding-through, and smear. The high intensity variation within
2
 Sharmistha Bhattacharya et al./Ann. Fuzzy Math. Inform. x (201y), No. x, xx–xx

both the document background and foreground caused by degradations makes it

difficult to design an uniform classification method that correctly separates text and
background for all kinds of degraded document images. Mathematical morphol-
ogy [47, 48] offers a unified and powerful approach to numerous image processing
problems. It provides powerful tools for extracting geometrical structures and rep-
resenting shapes in many applications. Morphological feature extraction techniques
have been efficiently applied to character recognition and document analysis. It is a
problem solving tool for extraction of text from non-uniform illuminated document
images.
In this paper we first introduce fuzzy image membership intensification operator
and some of its application is given to extract text from images. For the case of
proposed method, we first used morphological Bottom-Hat transformation to the
image to remove non-uniform illumination from the image. Using fuzzy intensifica-
tion operator text is extracted from background. We then reconstructed the image
using fuzzy transformation. Lastly we remove background noise of the image and
enhance the contrast of the text.
The rest of the paper is organized as follows. In section 2, some preliminaries
are given. In section 3, the proposed method is introduce. Section 4 details the
experiment results. Then in section 5, we present our conclusion.

Bottom-Hat
Input Image Fuzzification
Transformation

Removal of Reconstraction of
Text Extraction
background noise Image

Contrast Defuzzification
enhancement of
text ( Output image)

Figure 1. Procedure of the proposed method.

3
 Md.Atikul Islam et al./Ann. Fuzzy Math. Inform. x (201y), No. x, xx–xx

(a). (b). (c).

(d). (e). (f).

Figure 2. Different kind of non-uniform illuminated document images.

2. Preliminaries
In this section the preliminaries necessary for further study are cited.
2.1. Review of binarization techniques. A threshold T (x, y) is a value such that

1, T (x, y) < I(x, y)
b(x, y) =
0, I(x, y) ≤ T (x, y)
(1)
where b(x, y)is the binarized image and I(x, y) ∈ [0, 1] be the intensity of a pixel at
location, (x, y) of the image I. It is clear that a fixed value of the threshold surface
T (x, y) = const cannot yield satisfactory binarization results for images obtained
under non-uniform illumination and/or with a non-uniform background. Otsu [29]
calculates a global threshold by accepting the existence of two classes, foreground
and background, and choosing the threshold that minimizes the interclass variance
of the thresholded black and white pixels. Rosenfeld and Kak [40] select global
threshold from the histogram of 2D image. They assume that gray values of each
object are possible to cluster around a peak of the histogram of 2D image and try
4
 Sharmistha Bhattacharya et al./Ann. Fuzzy Math. Inform. x (201y), No. x, xx–xx

to compute the location of valley or peaks directly from the histogram. Kittler and
Illingworth [18] present an algorithm that is based on the fitting of the mixture of
Gaussian distributions and it transforms the binarization problem to a minimum-
error Gaussian density fitting problem. A local technique Kapur et al. [19] is
the maximization of the entropy of the thresholded image and is interpreted as
indicative of maximum information transfer. The image foreground and background
are considered as two different signal sources, so that when the sum of the two class
entropies reaches its maximum, the image is said to be optimally thresholded.
2.1.1. Niblack’s Technique[27]. In this method the local threshold value T (x, y) at
(x, y) is calculated within a window of size w × w as :
T (x, y) = m(x, y) + kδ(x, y) (2)
where m(x, y) and δ(x, y) are the local mean and standard deviation of the pixels
inside the local window and k is a bias. The result is satisfactory at k = −0.2 and
w = 15. The local mean m(x, y) and standard deviation δ(x, y) adapt the value of
the threshold according to the contrast in the local neighborhood of the pixel. The
bias k controls the level of adaptation varying the threshold value.
2.1.2. Sauvola’s Technique[43]. In Sauvola’s binarization method, the threshold T (x, y)
is calculated using the mean m(x, y) and standard deviation δ(x, y) of the pixels
within a window of size w × w as:
δ(x, y)
T (x, y) = m(x, y)(1 + k( − 1) (3)
R
where R is the maximum value of the standard deviation (R = 128 for a gray-scale
document), and k is a bias, which takes positive values in the range [0.2, 0.5]. The
local mean m(x, y) and standard deviation δ(x, y) adapt the value of the threshold
according to the contrast in the local neighborhood of the pixel. When there is high
contrast in some region of the image, δ(x, y) ∼ R which, results in T (x, y) ∼ m(x, y).
This is the same result as in Niblack’s method. However, the difference comes in
when the contrast in the local neighbourhood is quite low. In that case the threshold
T (x, y) goes below the mean value thereby successfully removing the relatively dark
regions of the background. The parameter k controls the value of the threshold in
the local window such that the higher the value of k, the lower the threshold from
the local mean.
2.1.3. Bernsen’s Technique[4]. In this method the local threshold value T (x, y) at
(x, y) is calculated within a window of size w × w as:
T (x, y) = 0.5(IM ax(i,j) + IM in(i,j) ) (4)
where IM ax(i,j) and IM in(i,j) are maximum and minimum gray values within the
local window, provided contrast
C(x, y) = IM ax(i,j) − IM in(i,j) ≥ 15 (5)
In this method, the threshold is set at the mid range value, which is the mean of
the maximum and minimum gray values in a local window of size w × w. A value
of w = 31 gives the satisfactory results. However, if the contrast C(i, j) is below a
certain threshold (15), then that neighborhood is said to consist of only one class,
5
 Md.Atikul Islam et al./Ann. Fuzzy Math. Inform. x (201y), No. x, xx–xx

foreground or background, depending on the value of T (x, y) . There is no bias to

control the threshold value.
2.2. Mathematical Morphology. Mathematical Morphology is a geometric ap-
proach in image processing and analysis with a strong mathematical flavour. Orig-
inally it was developed as a powerful tool for shape analysis in binary and, later
grayscale images Mathematical morphology was firstly introduced by Matheron and
Serra[47, 48] as a methodology for image processing. The language of Mathematical
Morphology is set theory. As such morphology offers a unified and powerful ap-
proach to numerous image processing problems. Sets in mathematical morphology
represent objects in an image. Morphological operators transform the original image
into another image through the interaction with the other image of certain shape
and size, which is known as the structure element. By translating the structuring
element over an image and by applying basic set operations, such as intersections
or union, the basic morphological operations (Dilation, Erosion, Opening, Closing)
are obtained. Gray-scale mathematical morphology is a natural extension of binary
mathematical morphology into gray-scale images.
2.2.1. Structuring element. Structuring element in gray-scale morphology perform
the same basic functions as their binary counterparts: They are used as ”probes”
to given image for specific properties. This simple ”probe” is called structuring
element, and is itself a binary image (i.e., a subset of the space or grid. Here are
some examples of widely used structuring elements (denoted by B):
• Let B is an open disk of radius r, centered at the origin.
• Let B is a 3 × 3 square, that is, B = (−1, −1), (−1, 0), (−1, 1), (0, −1), (0, 0),
(0, 1), (1, −1), (1, 0), (1, 1).
• Let B is the ”cross” given by: B = (−1, 0), (0, −1), (0, 0), (0, 1), (1, 0).
2.2.2. Gray-scale morphological operations. In terms of gray-scale morphology ero-
sion and dilation are, respectively,
(GΘS)(x) = minyF {G(x + y) − S(y)} (6)
(G ⊕ S)(x) = maxyF,(x−y)E {G(x − y) + S(y)} (7)
2
where x, yZ are the spatial co-ordinates;G : E → R is the gray-scale image;
S : F → Z is the gray-scale structuring element; and x, yZ 2 are the domains of the
gray-scale image. The definitions of gray-scale opening and closing are, respectively,
G ◦ S = (GΘS) ⊕ S (8)
G • S = (G ⊕ S)ΘS (9)
2.2.3. Top-Hat and Bottom-Hat Transformation. Let G : E → R be a gray-scale
image. Let S : F → Z be a gray- scale structuring element. Then, the white top-hat
transform of G is given by
Tw (G) = G − G ◦ S (10)
where ◦ denotes the opening operation. The black top-hat transform (Botom-Hat
transform) of G is given by
Tb (G) = G • S − G (11)
where • is the closing operation.
6
 Sharmistha Bhattacharya et al./Ann. Fuzzy Math. Inform. x (201y), No. x, xx–xx

Definition 2.1 ([35]). An image G of size m×n and L gray levels k ranging between
0 to L − 1 can be consider as an array of fuzzy singletones, each having a value of
membership denoting its degree of brightness relative to some brightness lavels. For
an image G, we can write in the notion of fuzzy sets as
G = {< Gk (i, j), µGk (i, j) > |i = 1, 2, ...m, j = 1, 2, ...., n} (12)
where Gk (i,j) is the value of G at position (i, j) at any gray level k, µGk (i, j)
denotes the degree of brightness possessed by the gray level intensity Gk (i, j) of the
(i, j)th pixel.

3. Text Extraction by proposed method

Before entering into Our proposed method we introduce Fuzzy Intensification
Operator (CRFO) and we give some of its application to remove background noise
of document images.

3.1. Fuzzy intensification operator CRFO[3]. In a low-contrast image most

of the gray level lies in the lower luminance range (typically under 0.5 on closed
interval [0,1] ) and fewer gray level lies in the upper luminance range (typically
above 0.5 on closed interval [0,1]). A higher contrast in an image can be achieved
by darkening the gray level in the lower luminance range and brightening the ones
in the upper luminance range. Moreover in the case of over-contrast, over-bright
images all gray level lies in the lower and upper luminance range respectively. Fuzzy
membership values of such images lies in a tiny interval which can be seen from
Fig.3(c). To get a good contrast enhanced images we must distribute membership
values of the image in lower and upper luminance range. Here we introduce CRFO
to distribute membership values from small interval to larger interval in [0,1] for the
case of low contrast, high contrast and over bright images, So that the maximum
contrast enhancement is possible. We define Contrast Intensification Fuzzy Operator
(CRFO) for enhancement of low contrast images and to remove background noise
of images. A mathematical expression of such functionCRFO:[0, 1] → [0, 1] can be
expressed as
2
CRF O (µGk (i, j)) = µGk (i, j) = {1 − e−αµGk (i,j) , 0 ≤ µGk (i, j) ≤ 1
/
(13)

where µGk (i, j) denotes the degree of brightness possessed by the gray level or colour
/
level intensity Gk (i, j) of the (i, j) th pixel, µGk (i, j) is the modified membership
value of µGk (i, j) by CRFO and α is an intensification parameter which can range
from 1 to infinity.
/
Then we generate new gray level or colour level Gk (i, j) for Gk (i, j) by the
following way
  σ

/ /
Gk (i, j) = (L − 1) 1 − 1 − µGk (i, j) τ (14)

where L is the highest gray level or colour level of the image, σ > 0 and τ ≥ 1 are
arbitrary parameters. Here image enhancement is done by applying the equation
(13)and (14) on images by setting suitable parameters.
7
 Md.Atikul Islam et al./Ann. Fuzzy Math. Inform. x (201y), No. x, xx–xx

3.1.1. Property. When α → ∞,


/ 1, if µGk (i, j) 6= 0
µGk (i, j) =
0, if µGk (i, j) = 0

(15)

3.1.2. Property. When α ≥ 1,


/ (0, 1], if µGk (i, j) 6= 0
µGk (i, j) =
0, if µGk (i, j) = 0

(16)
In fuzzy image processing a binarized image is obtain by changing the member-
ship values from closed interval [0, 1] to {0, 1}(0 for forground and 1 for background.
In Fig.3 and Fig.4 we show extraction of text from document images. The images of
Fig.3 is obtained by capturing in low-illumination condition by 2 mega pixel camera.
Fig.3(a)shows the document image of dimension 600 × 323 is in high contrast con-
dition. It can be seen from the Fig.3(c) membership values of the image lies in tiny
interval [0,0.1] and no membership values in upper luminance range. So to extract
text form the dark background we removed high contrast of the image by applying
CRFO operator at parameter α = 6000, σ =80 and τ = 2. Fig.3(b) shows the
image after extraction of text where as Fig.3(d)(membership histogram) shows after
binarization membership values are only 0 or 1. Now consider the document image
(d) of Fig.3 of dimension 796 × 726.The document image consist of gray background
with text. Fig.3(g) shows the image has no dark pixel, no white pixel and maximum
membership values lie in [0.5,0.8] and few membership values in the range[0.1,0.6].
Binarized image is shown in Fig.3(f) after removing noisy background.
Extraction of text by preserving the colour information is always a challenging
task.Many binarization algorithms was proposed earlier for extraction of text from
colour document images based on thresholding. But extraction of text preserving
the colour information is still unsolved problem for low-contrast colour document
images. Fig.4(a), (e) shows two colour document images of dimension 435 × 246
and 472 × 468 pixel respectively. The first image consist of dark background with
colour and black text. In Fig.4(c) the red, green and blue colour lines shows mem-
bership histogram of red green and blue colour channel. Here maximum number
of membership value lies in [0.3,0.7] for each channel red, green and blue and not
a single membership value greater than 0.7. Hence the image is in low-contrast
condition. The resulting image of Fig.4(b) is obtained by giving α = 30, σ = 5
and τ = 2. Fig.4(d)shows the membership histogram of extracted image(b). Here
it can seen from the graph that after processing the image maximum membership
values distributed from closed interval [0.3,0.7] to the upper and lower luminance
range . Hence a good text extracted image preserving the colour information. Text
extraction of another document image with noisy background is shown in Fig.4(e),
(f), (g), (h).Text colour in this case is black, so removal of background noise is much
easier than the previous one.
8
Sharmistha Bhattacharya et al./Ann. Fuzzy Math. Inform. x (201y), No. x, xx–xx

a. b.

c. d.

e. f.

g. h.

Figure 3. (a),(e) Original image. Text extraction by CRFO (b)

at α =6000, σ =80 and τ = 2 (f) α = 15, σ = 25 and τ = 2.
Membership histogram : (c), (g) Before extraction of text; (d), (h)
After extraction of text.

9
 Md.Atikul Islam et al./Ann. Fuzzy Math. Inform. x (201y), No. x, xx–xx

a. b.

c. d.

e. f.

g. h.

Figure 4. (a), (e) Original image. Text extraction by CRFO :

(b) at α =30, σ =5 and τ = 2 (f) α = 20, σ = 30 and τ = 2.
Membership histogram : (c), (g) Before extraction of text; (d), (h)
After extraction of text.
10
 Sharmistha Bhattacharya et al./Ann. Fuzzy Math. Inform. x (201y), No. x, xx–xx

3.2. Proposed method. Our new method includes seven steps (i) Image prepara-
tion. (ii) Removal of non-uniform illumination and uneven object.(iii) Fuzzification
of image.(iv) Text Extraction. (v) Re-construction of image. (vi) Background noise
removal. (vii) Contrast enhancement of the text.
(i) Image preparation. The low-contrast non-uniform illuminated images
shown in Fig.2 were acquired by camera of resolution 1600 × 1200, and the ex-
perimental image consisted of 600 × 450 pixels. To test our method, different non-
uniform illumination effects were applied on printed document. The document was
composed of text region and blank region. To test our method, the text and blank
regions were all under non-uniform illumination. Some patterns of low-contrast non-
uniform illuminated images were used shown in Fig. (2). These images had a low
contrast ratio in a certain region, which makes global and local method generally
a bad result. To show the step by effect we consider an image I(i, j) of dimension
600 × 450 pixels shown in Fig.5(a).
(ii) Removal of non-uniform illumination and uneven object.Global and
local threshold techniques, which are often the first step in object measurement,
cannot be applied to unevenly illuminated images. To overcome this problem we
used here morphological operator to even out the lighting in the image. One principal
application of Top-Hat and Bottom-Hat transforms is in removing objects from an
image by using an SE in the opening and closing that does not fit the objects to
be removed. The difference then yields an image with only the removed objects.
We here used Bottom-Hat transformation to extract black pixels from image. Al
the images shown in Fig.2 are consist of gray background with black text. A gray
scale Bottom-Hat transformation only visualise those pixels whose intensity values
are nearer to zero. The structuring element used here is either disk of certain radius
or square. The extraction of text depend on the structuring element, as big as the
structuring element text will be extracted more as a bright object. Fig.5(b) shows
the image after step (ii). Here structuring element is taken as a disk of radius 15.
(iii)Fuzzification of image.A gray-scale image can be consider as a fuzzy set
[35]and also colour image can be consider as a fuzzy set [16]. There are so many
method for fuzzification of an image . We describe image I(i, j) in fuzzy environment
as follows
I = {< Ik (i, j), µIk (i, j) > |i = 1, 2, ...m, j = 1, 2, ...., n} (17)
where Ik (i, j) is the intensity value of I at position (i, j) at any gray level or colour
level k ranges from 0 to 255, µIk (i, j) = IkL−1 (i,j)
denotes the degree of brightness
possessed by the gray level or colour level intensity Ik (i, j) of the (i, j)th pixel.
(iv) Text Extraction.After pre-processing with morphological transformation
and after fuzzification of intensity values we get image with dark background and
text as a white object. Since one main characteristic of CRFO operator is that
as parameter α grows up its gives modified membership values nearer to 1 except
membership values near about zero. That is, membership values increases as α
increase except for 0. To extract text as maximum as possible we then apply CRFO
defined in eqn.(13) on the resultant image. We modified membership values of image
I(i, j) at gray or colour level k as follows:
2
µIk (i, j) = {1 − e−αµIk (i,j) , 0 ≤ µIk (i, j) ≤ 1
(i)
(18)
11
 Md.Atikul Islam et al./Ann. Fuzzy Math. Inform. x (201y), No. x, xx–xx

So by setting larger value of parameter α (which ranges from 1 to infinity) we

get image with modified membership values, which leads extraction of text from the
image shown in Fig.5(c). Here parameter α is taken as 150.
(v) Re-construction of image.After step. (iv)we get image which has text
as a white object.Sice our main aim is to binarized the image, so that the text can
extract easily. We reconstruct the resulting image applying fuzzy transformation
such a way that the text will converted in black mode and background will be in
white mode.The fuzzy transformation is defined as follows:

(ii) (i) λ (i)

µIk (i, j) = 1 − (µIk (i, j)) γ , 0 ≤ µIk (i, j) ≤ 1 (19)
(i)
Here µIk (i, j) are membership values defined in (18), λ ≥ 0 and γ > 0 are
arbitrary parameter. Fig.5(d) is the reconstructed image obtain by setting λ = 1
and γ = 2. The image of Fig.5(d) shows that there are large number of background
noise with text. So to get noise free image we remove noise from the image by the
following way.
(vi) Background noise removal.To remove background noise we again apply
CRFO operator as follows:

(ii)
(iii) −βµI (i,j) (ii)
µIk (i, j) = {1 − e k , 0 ≤ µIk (i, j) ≤ 1 (20)
Here β is an intensification parameter which ranges from 1 to infinity. Fig.5(e)
shows the result after step(vi)with parameter β = 15.
(vii) Contrast enhancement of the text.To give more contrast in the image
we then apply fuzzy transformation which is defined as follows:

(iv) (iii) ω (iii)

µIk (i, j) = 1 − (1 − (µIk (i, j)) δ ), 0 ≤ µIk (i, j) ≤ 1 (21)
Here ω ≥ 0 and δ > 1 are arbitrary parameter. Fig.5(f) is the final binarized
image obtain by setting ω = 3, δ = 2 and generating new gray level by defuzzifying
the image as follows:

/ (iv)
Gk (i, j) = (L − 1) µIk (i, j) (22)
where L is the maximum intensity value of the image I.

4. Result and discussion:

In this section to show the effectiveness of our method we compared with seven
different methods from the literature like Otsu’s method, Kapur’s method, Kit-
tler’s method, Mean threshold method, Median threshold method, Niblack’s method,
Sauvola,s method. We worked under Wolfram Mathematica version 8 with windows
7 operating systems, 2 GB RAM, Pentium 2.30 GHz Dual core processor. The re-
sults of our method and other methods mention here are shown in Fig.6, 7, 8, 9,
10, 11. To test the generality of proposed method, not only the low-contrast non-
uniform illumination images but also other type of grayscale and colour degraded
images some results were shown in Fig.12, 13.
12
 Sharmistha Bhattacharya et al./Ann. Fuzzy Math. Inform. x (201y), No. x, xx–xx

(a). (b). (c).

(d). (e). (f).

(g). (h).

Figure 5. Step by step result of proposed method:(a)Original

image, (b)After Bottom-Hat transformation, (c) After applying
CRFO,(d) After reconstruction, (e) After applying CRFO, (f) After
contrast enhancement of the text, (g) Membership histogram before
extraction of text, (h) membership histogram after extraction of
text.
13
 Md.Atikul Islam et al./Ann. Fuzzy Math. Inform. x (201y), No. x, xx–xx

(a). (b). (c).

(d). (e). (f).

(g). (h). (i).

(j). (k).

Figure 6. (a) Original image. Extraction of text by: (b) Otsu’s

method, (c) Kapur’s method, (d) Kittler’s method, (e) Mean thresh-
old method, (f) Median threshold method, (g) Niblack’s method,
(h) Sauvola’s method, (i) Our method. Membership histogram: (j)
Before extraction of text, (k) After extraction of text.
14
 Sharmistha Bhattacharya et al./Ann. Fuzzy Math. Inform. x (201y), No. x, xx–xx

(a). (b). (c).

(d). (e). (f).

(g). (h). (i).

(j). (k).

Figure 7. (a) Original image. Extraction of text by: (b) Otsu’s

method, (c) Kapur’s method, (d) Kittler’s method, (e) Mean thresh-
old method, (f) Median threshold method, (g) Niblack’s method,
(h) Sauvola’s method, (i) Our method. Membership histogram: (j)
Before extraction of text, (k) After extraction of text.
15
 Md.Atikul Islam et al./Ann. Fuzzy Math. Inform. x (201y), No. x, xx–xx

(a). (b). (c).

(d). (e). (f).

(g). (h). (i).

(j). (k).

Figure 8. (a) Original image. Extraction of text by: (b) Otsu’s

method, (c) Kapur’s method, (d) Kittler’s method, (e) Mean thresh-
old method, (f) Median threshold method, (g) Niblack’s method,
(h) Sauvola’s method, (i) Our method. Membership histogram: (j)
Before extraction of text, (k) After extraction of text.
16
 Sharmistha Bhattacharya et al./Ann. Fuzzy Math. Inform. x (201y), No. x, xx–xx

(a). (b). (c).

(d). (e). (f).

(g). (h). (i)

(j). (k).

Figure 9. (a) Original image. Extraction of text by: (b) Otsu’s

method, (c) Kapur’s method, (d) Kittler’s method, (e) Mean thresh-
old method, (f) Median threshold method, (g) Niblack’s method,
(h) Sauvola’s method, (i) Our method. Membership histogram: (j)
Before extraction of text, (k) After extraction of text.
17
 Md.Atikul Islam et al./Ann. Fuzzy Math. Inform. x (201y), No. x, xx–xx

(a). (b). (c).

(d). (e). (f).

(g). (h). (i).

(j). (k).

Figure 10. (a) Original image. Extraction of text by: (b) Otsu’s
method, (c) Kapur’s method, (d) Kittler’s method, (e) Mean thresh-
old method, (f) Median threshold method, (g) Niblack’s method,
(h) Sauvola’s method, (i) Our method. Membership histogram: (j)
Before extraction of text, (k) After extraction of text.
18
 Sharmistha Bhattacharya et al./Ann. Fuzzy Math. Inform. x (201y), No. x, xx–xx

(a). (b). (c).

(d). (e). (f).

(g). (h). (i).

(j). (k).

Figure 11. (a) Original image. Extraction of text by: (b) Otsu’s
method, (c) Kapur’s method, (d) Kittler’s method, (e) Mean thresh-
old method, (f) Median threshold method, (g) Niblack’s method,
(h) Sauvola’s method, (i) Our method. Membership histogram: (j)
Before extraction of text, (k) After extraction of text.
19
 Md.Atikul Islam et al./Ann. Fuzzy Math. Inform. x (201y), No. x, xx–xx

(a). (b).

(c). (d).

(e). (f).

(g). (h).

(i). (j).

(k). (l).

Figure 12. (a), (e), (i) Original image. (b), (f), (j) Extraction of
text by the Proposed method. Membership histogram:(c), (g), (k)
Before extraction of text;(d), (h), (l) After extraction of text.

20
 Sharmistha Bhattacharya et al./Ann. Fuzzy Math. Inform. x (201y), No. x, xx–xx

(a). (b).

(c). (d).

(e). (f).

Figure 13. (a), (e), (i) Original image. (b), (d), (f) Extraction of
text by the Proposed method.

21
 Md.Atikul Islam et al./Ann. Fuzzy Math. Inform. x (201y), No. x, xx–xx

5. Conclusions
In this paper we have proposed text extraction method for the case of low-
contrast non-uniform illuminated document images based on morphological oper-
ator combined with Fuzzy Operator (CRFO). Comparative study shows that the
proposed method has achieved best performance for extraction of text over the
other method for non-uniform illuminated document images.We conclude that the
proposed method attains the most improvement in readability of characters in low-
contrast non-uniform illuminated document images.

Acknowledgements. I would like to thank UGC, New Delhi,India for the

financial support to doing this work under the scheme of Maulana Azad National
Fellowship for Minority students to pursue M.Phil/Ph.D Degree 2010-11.

References
[1] A. S. Abutaleb, Automatic thresholding of gray-level pictures using two- dimensional entropy,
Comput. Vision Graph. Image Process. 47(1) (1989) 22–32.
[2] B. Bataineh, S. N. H. S.Abdullah and K. Omar, An adaptive local binarization method for doc-
ument images based on a novel thresholding method and dynamic windows, Pattern Recogn.
Lett. 32 (2011) 1805–1813.
[3] S. Bhattacharya and A. Islam, Contrast Removal Fuzzy Operator in image processing, Ann.
Fuzzy Math. Inform. 6(2)(2013) 401–413.
[4] J. Bernsen, Dynamic thresholding of gray-level images, in:Proceedings of the Eighth Interna-
tional Conference on Pattern Recogn., Paris, IEEE Computer Society Press, France (1986)
1251–1255.
[5] I. Blayvas, A. Bruckstein and R. Kimme, Efficient computation of adaptive threshold surfaces
for image binarization, Pattern Recogn. 39 (2006) 89–101.
[6] I. Blayvas, A. Bruckstein, R. Kimme, Efficient computation of adaptive thresh- old surfaces
for image binarization, Pattern Recogn. 39 (2006) 89–101.
[7] H. D. Cheng, Y. H. Chen, Fuzzy partition of two-dimensional histogram and its application
to thresholding, Pattern Recogn. 32(5) (1999) 825–843.
[8] L. Cinque, S. DiZenzo, S. Levialdi, Image thresholding using fuzzy entropies, IEEE Trans.
SMC 28(1) (1998) 15–23.
[9] Q. Chen, Q. S. Sun, P. A. Heng and D. S. Xia, A double-threshold image binarization method
based on edge detector, Pattern Recogn. 41(4) (2008) 1254–1267.
[10] C. –H. Chou, W. –H. Lin and F. Chang, A binarization method with learning-built rules for
document images produced by cameras, Pattern Recogn. 43 (2010) 1518–1530.
[11] H. S. Don, A noise attribute thresholding method for document image binarization, IJDAR
4(2) (2001) 131–138.
[12] M. P. Dhore, V.M.Thakre and K.V. Kale, A novel morphological technique for segmentation
of document images in NIST database, Int. J. Mach. Intell. 3(5) (2011)172-180.
[13] A. Farahmand, A. Sarrafzadeh and J.Shanbehzadeh, Document noises and removal meth-
ods, Preceeding of the International Multi Conference of Engineers and Computer Scientists
(IMECS2013)(1)(March 13-15, 2013).
[14] Gorman, Binarization and multithresholding of document images using connectivity, Graph.
Mods. Image Process. 56(6) (1994) 494–506.
[15] B. Gangamma, S. Murthy and A. B. Sing, Restoration of degraded historical document images,
Journal of Emerging Trends in Computing and Information Sciences, 3(5) (2012) 792–798.
[16] H. J. Jimmermann, Fuzzy Set Theory and its Applications, second ed. Kluwer Academic
Publishers Dordrecht, 1991.
[17] S. H. Kwon, Threshold section based on cluster analysis, Pattern Recogn. Lett. 25(9) (2004)
1045–1050.
22
 Sharmistha Bhattacharya et al./Ann. Fuzzy Math. Inform. x (201y), No. x, xx–xx

[18] J. Kittler, J. Illingworth, On threshold selection using clustering criteria, IEEE Trans. Systems
Man Cybernet 15 (1985) 652–655.
[19] J. N. Kapur, P. K. Sahoo and A. K. Wong, A new method for gray-level picture thresholding
using the entropy of the histogram, Comput. Vision Graphics Image Process. 29(3) (1985)
273–285.
[20] I. K. Kim, D. W. Jung, R. H. Park, Document image binarization based on topographic
analysis using a water flow model, Pattern Recogn. 35 (2002) 265–277.
[21] S. J. Lu and C. L. Tan, Binarization of badly illuminated document images through shading
estimation and compensation, Int. Conf. on Document Anal. and Recogn. (2007)312–316.
[22] Y. Li, C. Y. Suen and M. Cheriet, A threshold selection method based on multiscale and
gray level co-occurrence matrix analysis, in:ICDAR’05:Proceedings of the Eighth International
Conference on Document Analysis and Recognition, IEEE Computer Society, Washington, DC,
USA (2005) 575–579.
[23] Y. Liu, S. N. Srihari, Document image binarization based on texture characteristics, IEEE
Trans. Pattern Anal. Mach. Intell. 19(5) (1997) 540–544.
[24] C. -C. Leung , K. -S. Chan, H. -M. Chan and W. -K. Tsui, A new approach for image enhance-
ment applied to low-contrast-low-illumination IC and document images, Pattern Recogn. Lett.
26 (2005) 769–778.
[25] J. M. Monte, R. D. Lins and R. Wachenchauzer, A new and efficient algorithm to binarize
document images removing back-to-front interference, J.UCS 14(2) (2008) 299–313.
[26] R. F. Moghaddam, M. Cheriet, A multi-scale frame work for adaptive binarization of degraded
document images, Pattern Recogn. 43 (2010) 2186–2198.
[27] W. Niblack, An Introduction to Digital Image Processing, Prentice-Hall, Englewood Cliffs,
NJ(1986) 115–116.
[28] S. Nomura, K. Yamanaka, T. Shiose, H. Kawakami and O. Katai, Morphological preprocessing
method to threshold degraded word images, Pattern recogn. Lett. 30(8) (2009) 729–242.
[29] N. Otsu, A threshold selection method from gray level histogram, IEEE Trans. on Sys. Man.
Cyber. 191 (January 1978 )62–66.
[30] H. -H. Oh, Kil-TaekLim, Sung-IlChien, An improved binarization algorithm based on a water
flow model for document image within homogeneous backgrounds, Pattern Recogn. 38 (2005)
2612–2625.
[31] N. Papamarkos, A technique for fuzzy document binarization, in:Proceedings of the 2001 ACM
Symposium on Document Engineering, Atlanta, Georgia, USA, ACM (2001) 152–156.
[32] J. R. Parker, Gray level thresholding in badly illuminated images, IEEE Trans. on Pattern
Analysis and Machine Intelligence 13 (1991) 813–819.
[33] Papamarkos, A neuro-fuzzy technique for document binarization, Neural Comput. Appl., 12(3-
4) ( 2003) 190–199.
[34] R. Parvati, S. Jayanti, N.Pallaniappan and S.Devi, Intuitionistic Fuzzy Approach to En-
hance text documents, Proceedings 3rd IEEE International Conference on Intelligent Sys-
tems(IEEE1S,06) (2006) 733–737.
[35] S. K. Pal, R. A. King, Image enhancement using smoothing with Fuzzy Sets, IEEE Trans.
Sys. Man Cybern. SMC-11 (1981) 494–501.
[36] M. A. Rami’rez-Ortego’n, E. Tapia, Transition pixel:a concept for binarization based on edge
detection and gray intensity histograms, Pattern Recogn. 43 (2010) 1233–1243.
[37] M. A. Rami’rez-Ortego’n and E. Tapia, et al., Transition thresholds and transition operators
for binarizatio nand edge detection, Pattern Recogn. 43 (2010) 3243–3254.
[38] H. Rajace, Enhancement of document images by morphological operator, 12th International
conference of computer society of Iran, 2007.
[39] P. Rangasamy, J. Kuppannan, K. T. Atanassov , and G. Gluhchev, Role of fuzzy and intu-
itionistic fuzzy contrast intensification operators in enhancing images, Twelfth Int. Conf. on
IFSs, Sofia, NIFS 14 (2) ( 2008) 59–66.
[40] A. Rosenfeld and A. C. Kak, Digital Picture Processing,2nd ed. New York: Academic, 1982.
[41] M. Sezgin and B. Sankur, Survey over image thresholding techniques and quantitative perfor-
mance evaluation, Journal of Electronic Imaging 13(1) (2004) 146–165.

23
 Md.Atikul Islam et al./Ann. Fuzzy Math. Inform. x (201y), No. x, xx–xx

[42] Y. Solihin and C. Leedham, Integral ratio: A new class of global thresholding techniques for
handwriting images, IEEE Trans. on Pattern Analysis and Machine Intelligence 21(8) (1999)
761–768.
[43] J.Sauvola, M.Pietikainen, Adaptive document image binarization, Pattern Recogn. 33(2)
(2000) 225–236.
[44] B. Su, S. Lu, and C. L. Tan, Binarization of historical handwritten document images using
local maximum and minimum filter, International Workshop on Document Analysis Systems
(2010) 303–314.
[45] J. Said, M. Cherict and C. Suena, Dynamical morphological processing:a fast method for base
line extraction, ICDAR (1996) 8–12.
[46] F. Sattar and D. B. H. Tay, Enhancement of document images using multiresolution and fuzzy
logic techniques, IEEE Signal process. Lett. 6 (1992) 249-252.
[47] J. Serra, Mathematical Morphology, Academic Press, London, UK 1, 1982.
[48] J. Serra, Image Analysis and Mathematical Morphology,Theoritical Advances, Academic Press,
London, UK, 1982.
[49] O. D. Trier, T. Taxt, Improvement of integrated function algorithm for binarization of docu-
ment images, Pattern Recogn. Lett. 16(3) (1995) 277–283.
[50] M. Valizadeh, E. Kabir, An adaptive water flow model for binarization of degraded document
images, IJDAR, Springer-Verlag, 2012 (20 January).
[51] M. Valizadeh, N. Armanfard and M. Komeili, et al., A novel hybrid algorithm for binarization
of badly illuminated document images, Computer Conference, Tehran (1-4) (2009).
[52] J. Wena, S. Li and J. Sun, A new binarization method for non-uniform illuminated document
images, Pattern Recogn. 46 (2013) 1670–1690.
[53] S. D. Yanowitz, A. M. Bruckstein, A new method for image segmentation, Comput. Vision
Graph. and Image Process. 46(1) (1989) 82–95.

(Mrs)Sharmistha Bhattacharya (Halder) (halder− [email protected])

Reader,Department of Mathematics, Tripura University, Suryamaninagar, Agartala-
799022, Tripura, India

Md. Atikul Islam ([email protected])

Research scholar, Department of Mathematics, Tripura University, Suryamaninagar,
Agartala-799022, Tripura, India.

24