THRESHOLD CORRECTION OF DOCUMENT IMAGE
BINARIZATION FOR TEXT EXTRACTION
Hiroshi Tanaka, Yusaku Fujii and Yoshinobu Hotta
Fujitsu, 4-1-1 Kamikodanaka, Nakahara-ku, Kawasaki, Kanagawa 211-8588, Japan
Keywords: Adaptive binarization, Text extraction, Thresholding, Otsu binarization, Threshold correction, Background
noise, Niblack, Image resolution.
Abstract: In this paper, a simple threshold correction method for document image binarization for text extraction is
presented. This method enhances the binary image of characters, which is often adversely influenced by
neighboring strong pixels or background noise. The threshold correction method is based on a similar
method applied to ruled-line extraction presented by the author, and is claimed to be effective to text
extraction. The author also reveals the relationship between effectiveness of the method and the image
resolution.
1 INTRODUCTION
One of the most important objectives of document
image binarization is to extract text images from the
document background. In a simple document model,
each object is considered to be placed on the flat
surface of the document background. According to
this model, binarization can be considered as a two-
class discrimination problem for determining a
global threshold (Otsu, 1979). However,
complicated document images require adaptive
binarization, in which the local threshold is
calculated for each pixel. Such images have complex
designs, which cannot be expressed using two
classes; further, they could be severely degraded.
In the past, various adaptive binarization
methods have been proposed. Trier (Trier and Jain,
1995) compared several binarization methods on the
bases of thier character recognition accuracies and
concluded that Niblack’s method (Niblack, 1986)
yields the best result when the noise reduction
technique is applied. Sauvola (Sauvola et. al., 1997)
modified Niblack’s method using region analysis, in
which textual and nontextual regions were separated
from each other. Sauvola’s method has been the
most popular binarization method for document
images. These methods assume that pixels can be
classified into two classes among local neighbors.
In the recent years, we can also find a lot of
newly invented binarization methods that may
overcome some problems of conventional methods.
For example, DIBCO 2009, the Document Image
Binarization Contest held in ICDAR 2009, is a good
collection of the latest document binarization
methods (Gatos et. al., 2009). Although there are
great methods proposed in DIBCO 2009, most of
them focus on binarizing much degraded images
such as historical documents depending on the
image quality used in the contest (Fig. 1), and then
they require much computing cost.
Our document recognition system recognizes
binarized text images obtained by an adaptive
binarization method based on Niblack’s method
(Kamada and Fujimoto, 1999). As described later,
adaptive binarization methods, including those
developed by Niblack and Sauvola, have a problem.
Because these methods are based on the assumption
that local neighbors can be classified into two
classes, some pixels that have three or more pixel
classes in each local area are often dropped off. This
results in broken shapes of character images (Fig. 2)
and causes errors in character recognition. We solve
this problem by correcting the binarization threshold
with respect to the neighboring threshold surface.
This technique was once applied to ruled-line
extraction (Tanaka, 2009) and is also proved to be
effective to text extraction.
In Section 2, we describe the problems of
conventional methods and our solutions. In Section
3, we present experimental results. Finally in Section
4, we conclude the paper.
387
Tanaka H., Fujii Y. and Hotta Y..
THRESHOLD CORRECTION OF DOCUMENT IMAGE BINARIZATION FOR TEXT EXTRACTION.
DOI: 10.5220/0003396503870391
In Proceedings of the International Conference on Computer Vision Theory and Applications (VISAPP-2011), pages 387-391
ISBN: 978-989-8425-47-8
Copyright
c
 2011 SCITEPRESS (Science and Technology Publications, Lda.)

Figure 1: Sample images used in DIBCO 2009. (copied
from Gatos et al., 2009)
Figure 2: Broken text image.
2 BINARIZATION ALGORITHM
In our algorithm (Fig. 3), initial thresholds are
calculated for each pixel, and the pixels area initially
classified as foreground or background pixels. The
initial classification is used in the background
determination process to determine whether each
pixel can be a foreground pixel. For foreground
pixels, binarization thresholds are obtained by
correcting initial thresholds.
Figure 3: Binarization algorithm.
In the following subsections, the basic concept of
the adaptive binarization method, which is used to
calculate initial thresholds, is described. Background
noise reduction, which is a modification of Eikvil’s
method (Eikvil, 1991), will be described. Finally, the
method for the correction of binarization thresholds
is presented.
2.1 Initial Thresholding
The threshold values for each pixel are calculated in
the initial thresholding step using Niblack’s method
and the threshold surface is formed by the threshold
values. The threshold for a pixel is obtained using
the neighboring pixel values I (i, j), reverse values of
brightness, in the
w
w
×
local areas. The initial
threshold T is calculated using the following
formulas:
∑∑
=
w
i
w
j
wjiIm
2
/),(
(1)
∑∑
−=
w
i
w
j
wmjiI
222
/}),({ δ
(2)
δ
kmT
+
=
(3)
A binarized image is obtained by comparing
each pixel value with its associated threshold (Fig.
4). The value
δ
k in formula 3 is used to shift the
threshold value so as to reduce background noise
(Fig.5).

T
Figure 4: Thresholding by Niblack’s method.
Figure 5: Binarization example.
2.2 Background Noise Reduction
When a background surface has small fluctuations,
many tiny noises will be extracted from the
background because the value
δ
k is so small that
even minute noises cannot be suppressed (Fig. 6).
To reduce these noises, the value of background
flatness is calculated, and noises in the flat region
are eliminated.
We calculate the flatness value on the bases of
Eikvil’s method, which is described as follows. In
the neighboring
w
w
×
local areas, each pixel can
be classified according to the initial threshold
(formula 3), and the mean values of each class, i.e.,
μ
1
and μ
2
, can be obtained. The flatness value F of
the central pixel of the local area is defined as
Gra
y
-scale
image
Initial
thresholding
Background
determination
Threshold
correction
Binar
y
image
I(i,j) B(i,j)
T(i,j)
(a) step boundary (b) slope boundary (c) flat slope
gray
level
x
gray
level
x
binary
image
gray
level
x
(a) original image 1 (b) binarized image 1
(c) original image 2 (d) binarized image 2
VISAPP 2011 - International Conference on Computer Vision Theory and Applications
388

dF −−= ||
21
μ
μ
, where d is a predefined
parameter, and the pixel is considered to be flat
when F is negative. For reference, Bernsen (Bernsen,
1986) has defined F as follows:
djiIjiIF
w
ji
w
ji
−−= )},({min)},({max
,
,
(4)
And Sauvola has proposed a flatness measure called
“transient difference” to distinguish between flat and
nonflat regions.
Although Eikvil’s method is effective for most
images, it encounters problems when binarizing low
contrast images. Fig. 7 is an example of a test in
which we applied Eikvil’s method to extract ruled-
lines. As shown in Fig. 7, some of the ruled-line
pixels are dropped with the normal d parameter (e.g.,
10), but there should be a lot of noise pixels with the
lower parameter d (e.g., 2). We had proposed a new
method using adaptive parameter to improve ruled-
line binary image (Tanaka, 2009), in which the
parameter d is set to lower value only when the
mask shown in Fig. 8 detects straightness feature.
Although the adaptive parameter improves the
ruled-line images, it is not always effective for text
images because the straightness feature is not
essential for text images. Then we adopt another
noise remove method which removes small CCAs as
noise pixels (Sauvola et.al., 1997). It reduces some
of the remaining noise pixels as shown in Fig. 9.
Figure 6: Background noise.
Figure 7: Background decision adaptation.

areaA
areaB
areaC
Figure 8: Mask for horizontal line.
(a)grayimage
(b)lower'd’value (c)removenoise
Figure 9: Small CCA removal.
2.3 Threshold Correction
As adaptive binarization methods, such as those
developed by Niblack and Sauvola, assume that a
local area contains pixels of two classes, local areas
containing more than two classes of pixels often
encounter classification failures. Fig. 10 shows the
effect of neighboring pixels on the binarization
threshold. When a local area contains a third class of
pixels (black pixels in the figure), the threshold
shifts because of the change in the mean value m
(see formula 3). This causes lacking pixels in the
text images in Fig. 2.
The effect of the third class of pixels can be
prevented if the local area is shifted away from this
class as shown in Fig. 11. This shift is equivalent to
using the binarization threshold of the shifted
position at a shift width S from the current pixel. We
define the threshold conversion using a shift
operation as follows.
)},({min),(
,
ljkiTjiT
S
Slk
++=
′
−=
(5)
Fig. 12 explains the detail and a modification of
threshold correction. Fig. 12(a) is an ideal binary
image which has two classes (black and white) of
pixels. By scanning, it may be converted to an image
which has three or more pixel values such as Fig.
12(b) because of quantization errors. Fig.12 (b) also
expresses the neighboring area in which threshold
values are used to correct the threshold of current
position (Formula 5). As we think that the pixels in
the correction area should have similar pixel values,
the correction area is narrowed as shown in Fig.
12(c). Using this modification, threshold correction
works like Fig. 12(d),(e).
By the modification, formula 5 is also modified
as follows, where e is a predefined parameter and we
set a value 20 for e in the experiment.
ejiIljkiIwhen
ljkiTjiT
S
Slk
<−++
++=
′
−=
|),(),(|
)},({min),(
,
(6)
By carrying out threshold conversion using
formula 6, the result of binarization of the image
(a) gray image (b) d=10 (c) d=2 (d) adaptive
gray
level
x
image
threshold
foreground pixel background noise
binary
image
THRESHOLD CORRECTION OF DOCUMENT IMAGE BINARIZATION FOR TEXT EXTRACTION
389

shown in Fig. 13(a) changes from that shown in Fig.
13(b) to Fig. 13(c). Fig. 14 shows another example.
Please note that S must be smaller than w/2, or the
current pixel should be outside the shifted local
window.
Figure 10: Affection of neighboring pixels.

shiftwidth
Figure 11: Shifted local window.
S
(a)idealbinaryimage (b)threshold
correctionarea
(c)modifiedarea
(
d
)
thresholdcorrection
(
e
)
modifiedarea
S S
(Similarvaluepixels)
Gray
level
x
S
’
S
(Correctionarea)
Gray
level
x
Figure 12: Modification of correction area.
(a)grayimage (b)binaryimage
(c)binaryimage(improved)
Figure 13: Improved binary image.
Figure 14: Improved binary image (another example).
3 EVALUATION
We tested our method using a document image set
consists of 17 kinds of document images with 4
resolutions (150, 200, 300 and 600 dpi); 68 images
in total. Table 1 shows the result of character
recognition with recall and precision rate. Although
the absolute values do not imply anything because
they depend on the test images, the results of the
developed method are relatively more accurate than
those of the conventional method. It is observed that
almost one-third of the recognition errors are
corrected using the proposed method.
In addition, the proposed method can be seen
effective especially for lower resolution images. It is
because the dropped pixels (Fig. 2) are mainly low
contrast part of the text pixels, and are often found in
low resolution images caused by quantization errors.
Table 1: Character recognition rates.
conventional method proposed method
recall precision recall precision
150 dpi 94.91% 93.75% 97.21% 97.21%
200 dpi 97.05% 95.41% 98.37% 98.12%
300 dpi 97.94% 97.11% 98.01% 97.17%
600 dpi 97.98% 97.57% 98.23% 97.66%
average 96.97% 95.71% 97.96% 97.54%
4 CONCLUSIONS
In this paper, we have described a threshold
correction method that can be used to improve the
quality of text image binarization and lead to higher
accuracy of character recognition.
This method was originally developed to
enhance ruled-line extraction (Tanaka 2009), and is
revealed to be effective for text extraction. In the
case of ruled-lines, we used the straight-line feature
of ruled-lines to distinguish line pixels from
background noise. On the other hand, we use
another noise reduction method to suppress noises
for text binarization.
value
hist.
threshold
b
ac
k
ground
line
value
hist.
threshold
back
ground
line
neighboring
pixels
(a) gray image (b) binary image
(c) binary image (improved)
VISAPP 2011 - International Conference on Computer Vision Theory and Applications
390

According to the experimental result, we have
found that the proposed method is especially
effective for lower resolution images. The reason of
it can be explained that the pixel dropping problem
is often caused by quantization errors which is
mainly seen in lower resolution images. In addition,
it can be considered that the same errors may occur
even in high resolution images when character size
is very small.
As the next step, we will investigate the
relationship between the effectiveness of our method
and the character sizes in detail. It is because we
expect our method to improve the recognition
accuracy for higher resolution images.
REFERENCES
Otsu, N., 1979. A Threshold Selection Method from Gray-
level Histograms. In IEEE Trans. Systems, Man, and
Cybernetics, vol.9, no.1, 1979, pp. 62-66.
Trier, O. D., and Jain, A. K., Goal-directed Evaluation of
Binarization Methods. In IEEE Trans. PAMI. vol.17,
no.12, Dec. 1995, pp.1191-1201.
Niblack, W., An Introduction to Digital Image Processing.
In Prentice Hall, Englewood Cliffs, N. J., 1986,
pp.115-116.
Sauvola, J., Seppanen, T., Haapakoski, S., and Pietikainen,
M., Adaptive Document Binarization. In Proc. 4th.
ICDAR, Ulm, Germany, Aug. 1997, pp.147-152.
Gatos, B., Ntirogiannis, K., and Pratikakis, I., ICDAR
2009 Document Image Binarization Contest (DIBCO
2009). In Proc. 10th. ICDAR, Barcelona, Spain, Jul.
2009, pp.1375-1382.
Kamada, H., and Fujimoto, K., High-speed, High-
accuracy Binarization Method for Recognizing Text in
Images of Low Spatial Resolutions. In Proc. 5th.
ICDAR, Bangalore, India, Sep. 1999, pp.139-142.
Tanaka, H., Threshold Correction of Document Image
Binarization for Ruled-line Extraction. In Proc. 10th.
ICDAR, Barcelona, Spain, Jul. 2009, pp.541-545.
Eikvil, L., Taxt, T., and Moen, K., A Fast Adaptive
Method for Binarization of Document Images. In Proc.
1st. ICDAR, Saint-Malo, France, Sep. 1991, pp.435-
443.
Bernsen, J., Dynamic Thresholding of Gray-level Images.
In Proc. 8th. ICPR, Paris, France, Oct. 1986, pp.1251-
1255.
THRESHOLD CORRECTION OF DOCUMENT IMAGE BINARIZATION FOR TEXT EXTRACTION
391