Improved wolf algorithm on document images detection using optimum mean technique

Bulletin of Electrical Engineering and Informatics
Vol. 8, No. 2, June 2019, pp. 551~557
ISSN: 2302-9285, DOI: 10.11591/eei.v8i2.1426  551
Journal homepage: http://beei.org/index.php/EEI
Improved wolf algorithm on document images detection using
optimum mean technique
Wan Azani Mustafa1
, Mohamed Mydin M. Abdul Kader2
, Zahereel Ishwar Abdul Khalib3
1,2
Faculty of Engineering Technology, Universiti Malaysia Perlis, UniCITI Alam Campus, Sungai Chuchuh, 02100
Padang Besar, Perlis, Malaysia
3
Embedded, Network and Advanced Computing Research School of Computer and Communication Engineering
University Malaysia Perlis, 02600 Arau, Malaysia
Article Info ABSTRACT
Article history:
Received Dec 1, 2018
Revised Jan 31, 2019
Accepted Mar 1, 2019
Detection text from handwriting in historical documents provides high-level
features for the challenging problem of handwriting recognition. Such
handwriting often contains noise, faint or incomplete strokes, strokes with
gaps, and competing lines when embedded in a table or form, making it
unsuitable for local line following algorithms or associated binarization
schemes. In this paper, a proposed method based on the optimum threshold
value and namely as the Optimum Mean method was presented. Besides,
Wolf method unsuccessful in order to detect the thin text in the non-uniform
input image. However, the proposed method was suggested to overcome the
Wolf method problem by suggesting a maximum threshold value using
optimum mean. Based on the calculation, the proposed method obtained a
higher F-measure (74.53), PSNR (14.77) and lowest NRM (0.11) compared
to the Wolf method. In conclusion, the proposed method successful and
effective to solve the wolf problem by producing a high-quality
output image.
Keywords:
Binarization
Document
Mean
Optimum
Wolf
Copyright © 2019 Institute of Advanced Engineering and Science.
All rights reserved.
Corresponding Author:
Wan Azani Mustafa,
Department of Electrical Engineering Technology,
Faculty of Engineering Technology, Universiti Malaysia Perlis,
UniCITI Alam Campus, Sungai Chuchuh,
02100 Padang Besar, Perlis, Malaysia.
Email: wanazani@unimap.edu.my
1. INTRODUCTION
Recently, Document Image Binarization (DIB) is one of the first stages of the Document Image
Analysis (DIA) and the further recognition pipeline. It consists mainly of classifying the pixels of the image
into the background and foreground pixels [1-4]. At this stage, “binarization” will imply some cleaning and
enhancement, since detecting the foreground pixel will remove, for instance, ink spots, degradations or ink-
bleed through, besides it will recover lost strokes or other foreground information. The degraded image
quality such as contrast variation and illumination give a big effect on the binarization process [5, 6]. The
research areas of document binarization from old, degraded documents attract researchers because of the
challenges of segmentation of the foreground text from degraded background [7-12]. Binarization is also
known in the literature as thresholding, since a threshold value is applied to the pixels to saturate them in one
of the classes. Image binarization takes into account any images, but DIB is meant for documents where text
(and other graphics) are considered foreground and the background is the “common background of the image
[13, 14].

 ISSN: 2302-9285
Bulletin of Electr Eng and Inf, Vol. 8, No. 2, June 2019 : 551 – 557
552
The general image binarization (or thresholding algorithms) could be divided into global, local
thresholding and mixing strategies. Global thresholding establishes a fixed threshold for all pixels of the
image, whereas local thresholding applies different thresholds for each pixel (using pixel local information).
Hybrid approaches try to compose both strategies: they use global and local information for thresholding.
Global thresholding techniques usually need less computation, and they work well in simple cases, though
fail in complex images. The most popular global image threshold is an Otsu’s method [15] which after
computing the gray level histogram, takes the optimal value that minimizes the inner class variance. There
are some refinements based on this method [16, 17]. Taking a closer look at document specific contributions,
Ntirogiannis et al. [18] proposed a combination of the global and local threshold for handwritten text
recognition. First, a background estimation is applied to binarize the image globally, then they work with
Connected Components (CCs), the approach takes stroke characteristics and discard components that do not
suit to the common stroke criteria. The above finding is consistent with the study by Moghaddam and Cheriet
[19]. The work in a similar manner since it extracts features such as stroke width and line height. It adapts
itself to the document structure because it combines a grid-based modeling and the estimated
background map.
In degraded documents, such historical records, traditional and general methods present poor
performances. Degraded documents involve not only binarization, however, the restoration of document parts
(mainly lost strokes) [20, 21]. In this purpose, usually, stroke characteristics are integrated into the
binarization. It is possible to find a vast set of contributions addressing these issues [22-24]. For example,
Moghaddam and Cheriet [25] treat binarization of degraded documents by combining different binarization
scales. Their work aims to restore lost strokes and eliminate ink bleed-through. In contrast, the study by Nina
et al. [26] indicated that a recursive Otsu’s thresholding could be used for binarization of historical
documents as well. In the same objective, another adaptive approach by the combination of image contrast,
gradients, and text shapes are estimated by Canny’s edge extractor was studied by Su et al. [27]. Besides,
Valizadeh and Kabir [28] suggest to extract useful characteristics in order to classify the pixels in background
or foreground; they use structural contrast taking into account information like the stroke width for feature
extraction. With the selected features, space is partitioned into several regions (clustering) where the pixel
within each region will be classified mixing the region information and the previous features extracted.
In this paper, a modification Wolf algorithm based on optimum mean was presented. The target of
this study is to improve the detection accuracy and blurred problem on binarization result and at the same
time to reduce the noise and any artefact by determining the optimum threshold value based on adjusting the
mean. In terms of the analysis stage, a few image quality assessment (IQA) such as an F-measure, Peak
Signal Noise Ratio (PSNR) and Negative Rate Metric (NRM) will be obtained. This paper is organized into
four sections. Section 2 explains the proposed method which focuses on Optimum Mean Technique, Section
3 shows the results performance in terms of the IQA. Section 4 describes the summary of the work.
2. RESEARCH METHOD
The optimum mean suggests based on a modification of Wolf method. Basically, the wolf method is
to localize artificial text in images and videos using a measure of accumulated gradients and morphological
processing. The quality of the localized text is improved by robust multiple frame integration [29]. However,
wolf method is not working properly on the image which contents many structures such as text image and
retinal image. The equation of wolf method as follows;
)
(
)
1
( M
m
R
s
a
aM
m
a
T 



 (1)
m and s is a mean and standard deviation, while a is a gain parameter. Then, M represent image size
minimum value of the gray levels and R is a maximum value of the standard deviations. Actually, the Wolf
method successful to segment the text region, however failed when dealing with the non-uniform and low
contrast input image. Besides, the output from the Wolf method also looks blurred, not clear and low quality
especially when performed on thin text.
In this paper, the objective is to determine the optimum threshold value which is likely to work
better for many types of low contrast document images. A slight modification of the Wolf threshold by using
adjustment of mean value was proposed. The optimum mean method was proposed to overcome this problem
by finding the maximum threshold value. The main advantage of the proposed method over Wolf is that it
considerably improves the image quality, reduced the blurred image and eliminate the noise.

Bulletin of Electr Eng and Inf ISSN: 2302-9285 
Improved wolf algorithm on document images detection using optimum mean technique (Wan Azani Mustafa)
553
By referring to (1), the mean value m will give a high effect on the threshold value as shown in
Figure 1. For example, according to Figure 1, if the mean value is less than average ( m -15) the result was
blurred and more information details were lost. Otherwise, if the mean value is more than the average ( m
+15) the result is better and improved. In this paper, the specific value (maximum mean) was found to
replace the normal mean value. The aim is to find the specific value more than the normal average.
Original Standard Mean
By using the original mean value,
the output image was satisfied.
However, a few thin texts were
loss and the image look blurred.
If the mean lower than
standard mean
(mean –15)
If modified the mean value of the
lower original mean, the output
images look more blurred and not
clear.
If the higher than standard
mean
(mean +15)
By applying mean value over than
original mean, the output image
shows more clear and also able to
detect the thin text
Figure 1. Comparison of resulting image with different mean value
Actually, the Wolf method failed to binarize the text image in low contrast condition because the
threshold value is not accurate. Therefore, the proposed method tends to improve the threshold value based
on an optimum mean calculation and applied to the low contrast image. The specific and maximum threshold
value was suggested caused if the threshold value is too high, it’s will introduce noise and artefact on the
resulting image. In this paper, the maximum mean is calculated in order to replace the original mean. The
maximum mean equation is depicted as follows;
 
 
y
x
mean
Omean ,
max
 (2)
where, )
,
(
max y
x is the maximum intensity of the overall mean image and mean is the standard mean.
Normally, the method was processed on 3x3 windowing size and contain a low of the mean value in the
whole images. So, we select the maximum mean value to put into the final algorithm. The final proposed
algorithm is;
)
(
)
1
( M
O
R
s
a
aM
O
a
T mean
mean 



 (3)
Based on the modified algorithm, the blur problem and unsegment thin text can be solved. Then, it
will increase the binarization result and produced high image quality. In order to evaluate the effectiveness of

 ISSN: 2302-9285
554
the proposed method and compared to the Wolf methods, 14 document images experimented and the results
are explained in the following section.
3. RESULTS AND ANALYSIS
In this study, 14 document images with non-uniform image background intensity due to uneven
illumination were tested [30]. The document dataset is publicly image provided by
http://utopia.duth.gr/~ipratika/HDIBCO2012/benchmark. Every input image shows the different pixel size
and the major problem images is caused by the fade ink and degrade of background. In this experiment, all
images were processed in grayscale images. The proposed process was performed on the default windowing
size, which is 3×3 [31, 32]. In binarization stage, a small windowing size will give a better result compared to
the larger size [33]. A few mathematical statistics such as variance and the standard deviation was obtained
before applying the final proposed equation. Figure 2 illustrated the comparison resulting image for Wolf
method and the proposed method. Based on observation, the wolf method failed in order to detect the thin
text, especially on illumination and low contrast problem input image. However, after applying the
modification proposed method was increased the resultant performance. The resulting image shows that more
clearly and improved the image quality.
Original Ground Truth Wolf Optimum Mean
Figure 2. Comparison of Wolf method and Optimum mean method on the document image

555
In order to prove the performance method, a few selected Image Quality Assessment (IQA) such as
an F-measure, PSNR and NRM was calculated. The quality of the segmented image is determined based on
the pixels similarity of the resultant segmented image against the manually segmented image. The
mathematical for F-measure, PSNR and NRM as follows;









































FP
TP
TP
FN
TP
TP
FP
TP
TP
FN
TP
TP
measure
F
2 (4)
2




















TN
FP
FP
FN
TP
FN
NRM (5)
MSE
S
PSNR
2
10
log
10
 (6)
TP, FP, FN and TN denote the True Positive, False Positive, False Negative and True Negative
values, respectively [34, 35]. Mathematically, the highest of F-measure and PSNR show the good
binarization result [36, 37]. Oppositely, the lowest of NRM should be present the good image quality.
Table 1 presents the comparison result between Wolf method and Optimum Mean method based on IQA.
Table 1. Result of Image Quality Assessment (IQA) in terms of F-measure PSNR, NRM and MPM
Image
F-measure PSNR NRM
Wolf Optimum Mean Wolf Optimum Mean Wolf Optimum Mean
1 70.81 78.12 15.79 15.66 0.22 0.08
2 28.90 76.28 11.49 14.19 0.42 0.14
3 26.86 81.47 12.68 15.69 0.42 0.04
4 78.95 82.46 17.68 17.29 0.17 0.05
5 72.54 65.52 15.83 12.07 0.21 0.03
6 59.42 69.90 14.12 12.59 0.28 0.04
7 27.06 60.98 12.51 13.76 0.42 0.25
8 81.31 71.35 16.40 12.44 0.15 0.03
9 37.71 84.67 10.54 14.84 0.38 0.11
10 22.53 80.93 10.58 14.63 0.44 0.13
11 12.06 84.77 10.93 15.57 0.47 0.06
12 28.17 74.29 13.18 15.81 0.42 0.17
13 18.29 49.86 13.58 14.59 0.45 0.32
14 69.86 82.89 16.66 17.62 0.23 0.05
Average 45.32 74.53 13.71 14.77 0.33 0.11
According to the Table 1, the F-measure and PSNR show the increment result after applying the
proposed method. Based on the average value, the proposed method achieved an F-measure (74.53) and the
PSNR (13.71) compared to the Wolf method which is an F-measure (45.32) and the PSNR (13.71). In terms
of increment percentage, the proposed method scored 64.45 % (F-measure) and 7.73 % (PSNR). Besides, the
proposed method also gave the lowest NRM (0.11) compared to the Wolf method (0.33). From this data, we
can see that the proposed method was successful to improve the binarization performance of Wolf method
4. CONCLUSION
Nowadays, digital image processing is very important in order to analyse the information on the
image. Thousands of valuable historical documents stored on the shelves of national libraries throughout the
world are waiting to be scanned in order to facilitate access to the information it contains. In this paper, a new
approach based on a modification of Wolf method was present. The main goal of the study was to determine
the optimum threshold value using the maximum mean. The second objective is to solve the blur problem
and improved the detection performance. Based on the numerical result, the proposed method obtained the
highest PSNR (14.773), F-measure (74.539) and lowest NRM (0.111) compared to the Wolf method.
Summary, the proposed method effective and successful to solve the Wolf problem and also increased the
binarization result.

 ISSN: 2302-9285
556
ACKNOWLEDGEMENTS
The author would like to acknowledge support from the Short Term Grant under a grant number of
9001-00576, Universiti Malaysia Perlis (UniMAP)
REFERENCES
[1] B. M. Singh, R. Sharma, D. Ghosh, and A. Mittal, “Adaptive binarization of severely degraded and non-uniformly
illuminated documents,” Int. J. Doc. Anal. Recognit., pp. 393–412, 2014.
[2] E. T. Zemouri, Y. Chibani, and Y. Brik, “Restoration based Contourlet Transform for historical document image
binarization,” Int. Conf. Multimed. Comput. Syst. -Proceedings, vol. 0, pp. 309–313, 2014.
[3] H. Z. Nafchi, R. F. Moghaddam, and M. Cheriet, “Phase-based binarization of ancient document images: Model
and applications,” IEEE Trans. Image Process., vol. 23, no. 7, pp. 2916–2930, 2014.
[4] A. Sehad, Y. Chibani, and M. Cheriet, “Gabor Filters for Degraded Document Image Binarization,” Proc. Int. Conf.
Front. Handwrit. Recognition, ICFHR, vol. 2014-Decem, pp. 702–707, 2014.
[5] W. A. Mustafa and H. Yazid, “Image Enhancement Technique on Contrast Variation : A Comprehensive Review,”
J. Telecommun. Electron. Comput. Eng., vol. 9, no. 3, pp. 199–204, 2017.
[6] W. A. Mustafa, H. Yazid, and S. Yaacob, “A Review : Comparison Between Different Type of Filtering Methods
on the Contrast Variation Retinal Images,” in IEEE International Conference on Control System, Computing and
Engineering, 2014, pp. 542–546.
[7] W. A. Mustafa, H. Yazid, and M. Jaafar, “An Improved Sauvola Approach on Document Images Binarization,” J.
Telecommun. Electron. Comput. Eng., vol. 10, no. 2, pp. 43–50, 2018.
[8] E. Ahmadi, Z. Azimifar, M. Shams, M. Famouri, and M. J. Shafiee, “Document image binarization using a
discriminative structural classifier,” Pattern Recognit. Lett., vol. 63, pp. 36–42, 2015.
[9] N. Mitianoudis and N. Papamarkos, “Document image binarization using local features and Gaussian mixture
modeling,” Image Vis. Comput., vol. 38, pp. 33–51, 2015.
[10] W. A. Mustafa and H. Yazid, “Illumination and Contrast Correction Strategy using Bilateral Filtering and
Binarization Comparison,” J. Telecommun. Electron. Comput. Eng., vol. 8, no. 1, pp. 67–73, 2016.
[11] W. A. Mustafa and H. Yazid, “Background Correction using Average Filtering and Gradient Based Thresholding,”
J. Telecommun. Electron. Comput. Eng., vol. 8, no. 5, pp. 81–88, 2016.
[12] W. A. Mustafa and H. Yazid, “Contrast and Luminosity Correction Based On Statistical Region Information,” Adv.
Sci. Lett., vol. 23, no. 6, pp. 5383–5386, 2017.
[13] W. A. Mustafa and M. M. M. A. Kader, “Binarization of Document Images: A Comprehensive Review,” J. Phys.
Conf. Ser., vol. 1019, no. 012023, pp. 1–9, 2018.
[14] W. A. Mustafa and M. M. M. A. Kader, “Binarization of Document Image Using Optimum Threshold
Modification,” J. Phys. Conf. Ser., vol. 1019, no. 012022, pp. 1–8, 2018.
[15] N. Otsu, “A Threshold Selection Method from Gray-Level Histograms,” in IEEE Transactions On Systrems, Man,
And Cybernetics, 1979, vol. 20, no. 1, pp. 62–66.
[16] A. D. Brink, “Thresholding of digital images using two-dimensional entropies,” Pattern Recognit., vol. 25, no. 8,
pp. 803–808, 1992.
[17] J. Kittler and J. Illingworth, “On threshold selection using clustering criteria,” IEEE Trans. Syst. Man. Cybern., vol.
SMC-15, no. 5, pp. 652–655, 1985.
[18] K. Ntirogiannis, B. Gatos, and I. Pratikakis, “A combined approach for the binarization of handwritten document
images,” Pattern Recognit. Lett., vol. 35, pp. 3–15, Jan. 2014.
[19] Reza Farrahi Moghaddamn and M. Cheriet, “AdOtsu: An adaptive and parameterless generalization of Otsu’s
method for document image binarization,” Pattern Recognit., vol. 45, pp. 2419–2431, Jun. 2012.
[20] W. A. Mustafa, H. Aziz, W. Khairunizam, Z. Ibrahim, S. Ab, and Z. M.Razlan, “Review of Different Binarization
Approaches on Degraded Document Images,” in IEEE International Conference on Computational Approach in
Smart Systems Design and Applications (ICASSDA), 2018, pp. 1–8.
[21] W. A. Mustafa, W. Khairunizam, Z. Ibrahim, S. Ab, and Z. M.Razlan, “Improved Feng Binarization Based on
Max-Mean Technique on Document Image,” in IEEE International Conference on Computational Approach in
Smart Systems Design and Applications (ICASSDA), 2018, pp. 1–6.
[22] J. Amudha, N. Pradeepa, and R. Sudhakar, “A survey on digital image restoration,” in Procedia Engineering, 2012,
vol. 38, pp. 2378–2382.
[23] C. C. Fung and R. Chamchong, “A review of evaluation of optimal binarization technique for character
segmentation in historical manuscripts,” in 3rd International Conference on Knowledge Discovery and Data
Mining, WKDD 2010, 2010, pp. 236–240.
[24] J. Kaur and R. Mahajan, “A Review of Degraded Document Image Binarization Techniques,” Int. J. Adv. Res.
Comput. Commun. Eng., vol. 3, no. 5, 2014.
[25] R. Farrahi Moghaddam and M. Cheriet, “A multi-scale framework for adaptive binarization of degraded document
images,” Pattern Recognit., vol. 43, no. 6, pp. 2186–2198, Jun. 2010.
[26] O. Nina, B. Morse, and W. Barrett, “A Recursive Otsu Thresholding Method for Scanned Document Binarization,”
in IEEE Workshop on Applications of Computer Vision (WACV), 2010, pp. 307–314.
[27] B. Su, S. Lu, and C. L. Tan, “Robust document image binarization technique for degraded document images,”
IEEE Trans. Image Process., vol. 22, no. 4, pp. 1408–1417, 2013.
[28] M. Valizadeh and E. Kabir, “Binarization of degraded document image based on feature space partitioning and

557
classification,” Int. J. Doc. Anal. Recognit., vol. 15, no. 1, pp. 57–69, 2012.
[29] C. Wolf, J. Jolion, R. F. V Rr-, I. De Lyon, and A. A. Einstein, “Extraction and Recognition of Artificial Text in
Multimedia Documents,” Pattern Anal. Appl., vol. 6, no. 4, pp. 309–326, 2003.
[30] W. A. Mustafa and M. M. M. A. Kader, “Document Image Database ( 2009 - 2012 ): A Systematic Review,” J.
Phys. Conf. Ser., vol. 1019, no. 012024, pp. 1–7, 2018.
[31] K. Saeed and M. Albakoor, “Region growing based segmentation algorithm for typewritten and handwritten text
recognition,” Appl. Soft Comput., vol. 9, no. 2, pp. 608–617, Mar. 2009.
[32] J. Ruiz-del-Solar and J. Quinteros, “Illumination compensation and normalization in eigenspace-based face
recognition: A comparative study of different pre-processing approaches,” Pattern Recognit. Lett., vol. 29, pp.
1966–1979, 2008.
[33] S. Y. W.Azani Mustafa, Haniza Yazid, “Illumination Normalization Of Non-Uniform Images Based On Double
Mean Filtering,” in IEEE International Conference on Control System, Computing and Engineering, 2014, pp.
366–371.
[34] R. Hedjam, H. Z. Nafchi, R. F. Moghaddam, M. Kalacska, and M. Cheriet, “ICDAR 2015 contest on MultiSpectral
Text Extraction (MS-TEx 2015),” Proc. Int. Conf. Doc. Anal. Recognition, ICDAR, pp. 1181–1185, 2015.
[35] I. Pratikakis, B. Gatos, and K. Ntirogiannis, “ICDAR 2013 document image binarization contest (DIBCO 2013),”
in Proceedings of the International Conference on Document Analysis and Recognition, ICDAR, 2013, pp. 1471–
1476.
[36] W. A. Mustafa and H. Yazid, “Conversion of the Retinal Image Using Gray World Technique,” J. Biomimetics,
Biomater. Biomed. Eng., vol. 36, pp. 70–77, 2018.
[37] W. A. Mustafa and M. M. M. A. Kader, “A Comparative Study of Automated Segmentation Methods for Cell
Nucleus Detection,” Malaysian Appl. Biol., vol. 47, no. 2, pp. 125–129, 2018.
BIOGRAPHIES OF AUTHORS
Wan Azani Mustafa received degree in Biomedical Electronic Engineering (2013) and PhD in
Mechatronic Engineering (2017) from University Malaysia Perlis (UniMAP). Registered on
Board Engineer Malaysia (BEM) on 2014. Published more than 40 academic articles and 1 book.
Now, working as a senior lecturer at University Malaysia perlis, Malaysia. Current interest
includes image processing, biomechanics, intelligence system, and control system.
Mohamed Mydin M. Abdul Kader received Master of Science In Electrical Engineering from
Universiti Sains Malaysia (USM). Recognition as a Professional Technologist member from
Malaysia Board Of Technologists (MBOT) ( 2018 ). Now, working as a lecturer at University
Malaysia perlis, Malaysia. Current interest includes image processing, control system, Robotic
and Automation.
Zahereel Ishwar Abdul Khalib received Ph.D in Computer Engineering, UniMAP MSc. in Real
Time Software Engineering, UTM B.Sc in Computer Engineering, California State University,
Sacramento, USA.Now, working as a senior lecturer at School of Computer & Communication
Engineering, University Malaysia perlis, Malaysia. Research interest Image and Signal
Processing, Real Time, SystemMulticore Programming and Multimedia Streaming.

Improved wolf algorithm on document images detection using optimum mean technique

Recommended

More Related Content

What's hot (18)

Similar to Improved wolf algorithm on document images detection using optimum mean technique (20)

More from journalBEEI (20)

Recently uploaded (20)

Improved wolf algorithm on document images detection using optimum mean technique