IMPROVED ADAPTIVE BINARIZATION TECHNIQUE FOR

DOCUMENT IMAGE ANALYSIS

Lal Chandra, Puja Lal, Raju Gupta, Arun Tayal

Newgen Software Technologies Limited, A-6, Satsang Vihar Marg

Qutab Institutional Area, New Delhi-110067, India

Dinesh Ganotra

GGSIP University, Kashmere Gate, Delhi-110006, India

Keywords: Binarization, Thresholding, Gamma Correction, Reverse Video, Contrast Stretch, ICR, OCR.

Abstract: Technology of image capturing devices has graduated from Black & White (B&W) to Color, still majority

of document image analysis and extraction functionalities work on B&W documents only. Quality of

document images directly scanned as B&W is not good enough for further analysis. Moreover, nowadays

documents are getting more and more complex with use of variety of background schemes, color

combinations and light text on dark background (reverse video) etc. Hence an efficient binarization

algorithm becomes an integral step of preprocessing stage. In proposed algorithm we have modified

Adaptive Niblack's Method (Rais et al., 2004) of thresholding to make it more efficient and handle reverse

video cases also. The proposed algorithm is fast and invariant of factors involved in thresholding of

document images like ambient illumination, contrast stretch and shading effects. We have also used gamma

correction before applying the proposed binarization algorithm. This gamma correction is adaptive to

brightness of document image and is found from predetermined equation of brightness versus gamma.

Based upon result of experiments, an optimal size of window for local binarization scheme is also proposed.

1 INTRODUCTION

Binarization of image is the central part to many

image processing applications like form processing,

check processing etc. The binarization process

computes the threshold value that differentiates

objects and background pixels.

A number of methods have already been

proposed for image binarization, most of them being

specific to their respective applications. The

accuracy of these image binarization or thresholding

techniques depends on the quality of the scanned

image and the nature (such as background, prints,

data cell sizes, contrast stretch) of the document. If

the input gray image is of good quality then the

output of most binarization technique can be used

for good analysis of image. However in real time

scenario, the quality and nature of input image may

not always provide desirable accuracy with most

binarization methods.

There are many factors that complicate the

thresholding scheme like ambient illumination,

variation of gray levels within the object and the

background, inadequate contrast etc. In case of

noisy background the binarization can become a

challenging job.

To improve binarization, (Gonzalez and Richard,

2005) use equation (6) for setting gamma of the

image and Adaptive Thresholding Technique (Rais

et al., 2004; Niblack, 1990). The basis of

binarization method is the calculation of appropriate

threshold. A thresholding selection method (Otsu,

1979) suggests minimizing the weighted sum of

variances of the objects and background pixels to

establish an optimum threshold. A fixed value of

threshold cannot give satisfactory results in case of

images with illumination variance. In document

image analysis, logical and semantic content

conservation is needed. Global threshold cannot

conserve these kinds of contents. Contrary to global

method, local method is an adaptive approach in

317

Chandra L., Lal P., Gupta R., Tayal A. and Ganotra D. (2007).

IMPROVED ADAPTIVE BINARIZATION TECHNIQUE FOR DOCUMENT IMAGE ANALYSIS.

In Proceedings of the Second International Conference on Computer Vision Theory and Applications - IFP/IA, pages 317-321

 SciTePress

which a threshold value is determined over a small

region. The Niblack's method (Niblack, 1990) uses

mean and standard deviation of image to compute

threshold over a small region (75 × 75 window).

Dynamic Thresholding (Bemsen, 1986) uses mean

and standard deviation of the image along with the

contrast to compute the threshold. Sauvola et al.

(1997) presented a modified Niblack's method

(Niblack, 1990), which uses adaptive contribution of

standard deviation in determining local threshold.

Gray-level Thresholding (Parker, 1991) is done by

computing local threshold value by classifying

object and background pixels and then using region

growing technique to produce the binarized image.

In our proposed algorithm we have modified

Adaptive Niblack's Method (Rais et al., 2004) of

thresholding to make it more efficient and handle

reverse video cases also. The proposed algorithm is

fast and invariant of factors involved in thresholding

of document images like ambient illumination,

contrast stretch and shading effects. We have also

used gamma correction before applying the

proposed binarization algorithm. Gamma correction

is adaptive to brightness of document image and is

found from predetermined equation of brightness

versus gamma. Based upon result of experiments, an

optimal size of window for local binarization

scheme is also proposed.

2 NIBLACK’S METHOD

Niblack’s method (Niblack, 1990) is a local

thresholding method that adapts the threshold

according to the local mean and local standard

deviation over a specific window size around each

pixel location. The local threshold at any pixel (i, j)

is calculated by equation (1)

,,, jijiji

kMT

(1)

Where

and

, ji

are the mean and variance

of a window in the image respectively. The size of

the window depends upon the application. The value

of the weight 'k' is used to control and adjust the

effect of standard deviation due to changes in

object's features. Niblack’s algorithm suggests the

value of 'k' to be −0.2.

Niblack's algorithm suffers from the problem of

local thresholding by providing details in the

binarized images that may not be required in

processing. Niblack's method uses fixed value of the

weight 'k' which is not the optimum value.

3 ADAPTIVE NIBLACK’S

METHOD

Adaptive Niblack Method (Rais et al., 2004) offers

improvement over original Niblack's method

(Niblack, 1990). It not only depends upon image's

local statistics characteristics but also considers the

global statistics. This algorithm calculates 'k'

dynamically for each pixel and thresholding is done

using Niblack's method. The normalized difference

between global and local mean provides information

about the illumination difference for each pixel

window with respect to global illumination.

Equation (2) provides a reasonable value for

factor 'k', but it fails to adapt to changes in images

with different contrast.

),max(

−

(2)

Here M is the global mean of the image and

the local mean computed on each window.

The use of standard deviation of image and local

window improves the result. This algorithm uses the

interrelation of global and local characteristics and

sets the threshold based on the relative change of

local and global mean and standard deviation. The

effect of standard deviation remains same on

different images having different local illumination

and contrast.

Adaptive Niblack method uses eq (2) for images

which do not have large variation in contrast. For

images with large variation in contrast eq (3) is

used.

),max(

)(

03.0

jiji

σσ

−

−=

(3)

4 PROPOSED ALGORITHM

This section describes the proposed algorithm along

with improvements over adaptive Niblack's

algorithm. We have used images with 256 gray

levels, scanned at 300dpi.

VISAPP 2007 - International Conference on Computer Vision Theory and Applications

318

4.1 Image Quality Improvement

Equation (4 and 5) gives an appropriate gamma

(Gonzalez and Richard, 2005) for each pixel in the

image. We devised these curves by manually

adjusting gamma for 150 document images and

optimising the Newgen Software Tech. (2005) Find

Data Cell (FDC) engine’s results. These equations

depend on the global mean of the image. In this

algorithm we propose two equations for calculating

gamma. Though quadratic equation is proposed,

however our experimental results demonstrate that

linear equation is sufficient for proper binarization.

Quadratic equation may be used in cases where

linear equation is giving unsatisfactory results.

However it may be time consuming. Figure 1

represents the plot of linear and quadratic equation

with respect to gamma and mean of the image.

Figure 1: Manually adjusted gamma vs. mean intensity of

original images.

2.3012.0

−= M

(4)

1412.0

00028.0 −+−= MM

(5)

Here M is mean of original image. We set

gamma for all gray pixels of the image. We use

)/1(

(6)

convention for gamma change, where I is the

image’s intensity.

4.2 Binarization

The proposed binarization algorithm is based on

Adaptive Thresholding Technique (Rias et al.,

2004). The algorithm is executed in three steps:

1. If the selected pixel's gray value is greater than

240 then saturate the pixel as white (gray value 255)

or if the pixel’s gray value is less than 30 then

saturate the pixel as black (gray value 0).

2. If the condition to execute first step is not

satisfied then set a window of size 15×15 and

calculate the mean. If the difference between

selected pixel’s gray value and mean is less than 10

then set the pixel as white. This operation is specific

for removing the gray background and preserving

the text and other components present over it.

3. For remaining pixels apply adaptive thresholding

technique. With the use of 15×15 window, the

algorithm requires less computation time without

deteriorating performance.

Figure 2(a) shows input image with dark

background and figure 2(b) is the output image after

processing with proposed algorithm.

(a) (b)

Figure 2: Execution of step 2 for dark gray background

image. (a) Dark background image (b) Image after

background removal.

4.3 Performance of the Proposed

Algorithm

Figure 3 shows the performance of the proposed

algorithm over adaptive thresholding technique.

Figure 3: Performance of proposed algorithm.

We executed both algorithms on 150 different

images. The images were off-the-shelf available

IMPROVED ADAPTIVE BINARIZATION TECHNIQUE FOR DOCUMENT IMAGE ANALYSIS

319

bank account opening forms. They had large

variation in terms of background, prints, data cell

sizes, illumination, contrast stretch etc. About

30,000 different cells were selected on them and

Newgen Software Tech. (2005) FDC engine was

used to detect intelligent character recognition (ICR)

cells and optical character recognition (OCR) cells.

The comparison of FDC results is given in table 1.

Table 1: Comparison of FDC engine.

Forms with 29089 ICR

cells

Detected ICR

cells

% Accuracy

Original forms 17742 61.0

Forms with gamma

correction

27119 93.0

Forms binarized with

proposed algorithm

28592 98.0

4.4 Results and Analysis

Figure 4 and 5 show the comparison of document

images with white background and gray background

respectively. Experiments were conducted on

images with varying background scanned with

scanner AVISION830C.

Figure 4(a) is a scanned document image. There

is not much difference between print and

background gray levels.

(a)

(b)

(c)

Figure 4: Comparison of output images with light lines

and text gray value with respect to background. (a)

Original Gray level image (b) Binarized through adaptive

thresholding (c) Binarized through proposed Algorithm.

Figure 4(b) is the binarized image through

adaptive thresholding technique. The ICR and OCR

cells are clear but the text quality is not so good and

also the background is still noisy. While as shown in

figure 4(c), the output of proposed algorithm, all

ICR and OCR cells are clear and the text quality is

good as compared to figure 4(b) and the background

noise is also removed.

(a)

(b)

(c)

Figure 5: Comparison of output images with non-

continuous gray background (a) Gray Background Image

(b) Binarized through adaptive thresholding technique (c)

Binarized through proposed algorithm.

A document image over gray background is

shown in figure 5(a). In this case the gray

background is non-uniform. The figure 5(b) is

the binarized image through adaptive

thresholding technique. Figure 5(c) is the

binarized image through proposed algorithm. In

this case the image quality is improved over

existing algorithm.

VISAPP 2007 - International Conference on Computer Vision Theory and Applications

320

4.5 Flowchart of Proposed Algorithm

Flowchart of the algorithm is shown in fig 6.

Figure 6: Flow chart of the proposed algorithm.

5 CONCLUSION

In this paper, we have proposed an algorithm of

binarization for document image analysis. This

algorithm was tested over images with large

variation in terms of background, prints, data cell

sizes, illumination, contrast stretch etc and it gave

satisfactory results. The result shows that it had

improved document image binarization significantly

over adaptive Niblack’s algorithm (Raise et al.,

2004) .We have also shown that a local window size

of 15 × 15 is the appropriate option for the

thresholding scheme as it provides significant time

optimization without deteriorating performance. The

use of local and global statistics has made our

algorithm strong and robust. Though targeted here

for document image analysis, it will be a good

candidate for other kind of applications as well

scene processing and image segmentation.

REFERENCES

Rais, N. B., Hanif, M. S., Taj, I. A., 2004. “Adaptive

Thresholding Technique for Document Image

Analysis.”, Proc. INMIC 2004, 8

Int. Multitopic

conf, 61-66 IEEE.

Gonzalez, R. C., Richard, E. W., 2005. “Digital Image

Processing”, Pearson Prentice-Hall Inc.

Niblack, W., 1990. “An Introduction to Digital Image

Processing”, Prentice-Hall Inc.

Otsu, N, 1979. "A threshold selection method from gray

level histograms." IEEE Tran. on Sys. Man. Cyber.

Bemsen, J., 1986. "Dynamic thresholding of gray-level

images". Proc. 8th ICPR, Paris. 1251-1255.

Sauvola, J., Seppanen, T., Haapakoski, R. Pietikainen, M.,

1997. "Adaptive Document Binarization." Int. Conf.

Doc. Ana. Rec., 147-152.

Parker, J.K., 1991. "Gray level thresholding in badly

illuminated images". IEEE Trans. Patt. Ana. Mac.

Intell. Volume 13, 813-819.

Newgen Software Technologies Ltd.

www.newgensoft.com/2005/products/omniextract.htm

I(i,j)

I(i,j)>240 OR

I(i,j)

Set a Window of

15×15

I(i,j)-M(i,j) < 10

M (i,j) ← mean(i:i+15,j:j+15)

Std ← Standard deviation (Image)

std(i,j) ← local standard deviation

k (i,j)

T(i,j) ← M – k(i,j)* std

Apply T(i,j) on I(i,j)

I(i,j) = 255

Set T(i,j) = 128

Yes

Check for

Com

letion

Yes

Increment

by one

Pixel

M ← Mean (Image)

Binarized

Ima

IMPROVED ADAPTIVE BINARIZATION TECHNIQUE FOR DOCUMENT IMAGE ANALYSIS

321