Content uploaded by Ramaprasad Poojary
Author content
All content in this area was uploaded by Ramaprasad Poojary on Jan 31, 2017
Content may be subject to copyright.
SSRG International Journal of Electronics and Communication Engineering (SSRG-IJECE) – Volume 3 Issue 6 – June 2016
ISSN: 2348 – 8549 www.internationaljournalssrg.org Page 1
Image Denoising using Decision Tree Based
Method
Remya Ravi Nair#1, Ramaprasad Poojary#2
#1M.Tech Student, School of Engineering & IT, Manipal University, Dubai, UAE-345050
Abstract—Images are often degraded by various
kinds of noises. Therefore, image denoising plays an
important role in image processing. In this paper,
Decision Tree Based Method has been proposed for
denoising of images. This algorithm comprises of
three stages Isolation Module (IM), Fringe Module
(FM) and Similarity Module (SM) for the detection of
noisy pixel followed by an edge preserving filter to
remove noise from images. The proposed algorithm
has been tested for different set of images. It has been
shown in the paper that proposed method outperforms
the conventional denoising algorithm.
Keywords—Image Denoising, Noise Addition,
Decision Tree, Noise Filter, Noise Removal
I.INTRODUCTION
An image may be defined as a two-dimensional
function, f(x, y), where x and y are the spatial (plane)
coordinates, and the amplitude for f is called the
intensity or grey level of the image at that point [1].
Images are often corrupted by various noises usually
when visual information is transmitted in the form of
digital images [2]. Hence the images obtained after
transmission may often be corrupted with noise. [1][2].
Noise hides the important details and properties of an
image. To enhance the image quality, we have to
remove noise from images without any loss of
information. Image denoising is one such powerful
method which is employed to remove the noise
wherein many methods have been adopted to remove
noise from corrupted images [3]. But preserving edges
and details in the filtering process is a main problem
[4]. These noises appear in images by various ways
such as due to faulty scanner, faulty digital cameras,
transmission channel errors, corrupted storage media,
etc. Impulse noise in an image occur due to bit errors
in transmission or may be induced during the signal
acquisition stage [5]. Impulse noise can therefore be
classified into two types, Fixed-Valued impulse noise
and Random-Valued impulse noise [5]. The Fixed
Valued impulse noise can also be called as ―Salt and
Pepper Noise‖ this is mainly because the pixel value
of a noisy pixel will either be minimum or maximum
value in a grayscale image [6]. The main idea of this
paper is to remove noise completely from the image
and also to preserve the edges of an image.
II. LITERATURE SURVEY
In this paper we make a short overview of the most
popular image denoising methods. Generally different
filters are used for eliminating different noises like
mean filter, Gaussian filter, etc., these are discussed in
detail as given below. Earlier methods such as alpha-
trimmed mean in which the method uses alpha-
trimmed mean only for the detection of impulse noise
instead of pixel value estimation [2]. Y Dong,
proposed a work based on a new directional weighted
median filter in which a new impulse detector was
used. A combination of weighted filter was also done
to get a directional weighted filter. However this
method had a limitation with the median filter where
the failure of the filter affected the performance of the
whole system [7]. Petrovic proposed a work based on
switching scheme with two cascaded detectors and
two estimators. But this work was limited since both
detectors are based on estimators of location and scale
median as well as it’s a high complex method and a
less efficient method [8]. C. Y Lien proposed a work
based on an efficient denoising scheme using VLSI
architecture. The VLSI architecture designing with a
processing rate of about 200 MHz was done. But the
drawback of this method was that the edges of an
image will not be preserved very well [9]. P.E Ng
proposed a method on switching median filter that
incorporates with an impulse noise detection design.
This paper was proposed to effectively denoise the
corrupted images. But a major drawback was of poor
visual quality that occurred as well as the SNR value
was very low [10]. P.Y Chen proposed a method that
was based on adaptive median filters in which the
algorithm was used for removing the salt and pepper
noise in which an impulse noise detector to detect the
noisy pixels was also done. But the main limitation of
this design was that it concentrated mainly on the
edges of an image as a result the noisy pixels present
in the other parts of an image were not clearly
removed a finally resulted with the blurring effect [11].
Iyad proposed a paper filtering technique with
Boundary Discriminative Noise Detection (BDND)
algorithm. The method related to switching median
filtering i.e., a main classification of this filter
technique by filtering the noisy pixels and leaving
other pixels intact. But the drawback of this method
was that the filtering using BDND itself has many
limitations as a result it tends to poor visual clarity of
the image [12]. Xuming Zhang proposed a filtering
SSRG International Journal of Electronics and Communication Engineering (SSRG-IJECE) – Volume 3 Issue 6 – June 2016
ISSN: 2348 – 8549 www.internationaljournalssrg.org Page 2
technique based on a two stage algorithm which is
known as switching based adaptive weighted mean
filter. In this method the pixels are being replaced with
the average value of the neighbouring pixels. Hence
the main drawback was the blurring effect of the
image since averaging is the process that is carried out
[13]. Dong-Hyuk Shin proposed a filtering method
using Gaussian filter. However the method was
limited since it is a very simple filtering technique
averaging of pixels was the main idea as a result the
output image produced was a blurred image and also
was not efficient for high noise densities [14]. P. D.
Hanlon proposed a filtering technique based on
kalman filter bank. In this method multiple model
adaptive estimation algorithm was used. The main
limitation was that the whole process was time
consuming, details were all reduced and filtering was
not that efficient [15]. Based on basic concepts that
are available in the literature, here we present an
efficient Decision-Tree-Based method (DTBM) that
can overcome the limitations faced in the area of
image denoising as we have discussed.
I. PROPOSED DECISION TREE BASED METHOD
Here a 3x3 mask is used for denoising process [5]. Let
us consider the coordinates(i,j) where the image pixel
that is required to be denoised has been located and it
is denoted as P(i,j) and its luminance value is named
as f(i,j) as shown in Fig. 1. The mask is divided into
two sets: WTopHalf and WBottomHalf. They are
given as
WTopHalf = {a,b,c,d}. ------------------------- (1)
WBottomHalf = {e,f,g,h} ------------------------- (2)
In order to determine whether the pixel is noisy or not
the Decision Tree Based Impulse Detector is used. In
this method, the determination of the pixel, whether it
is noisy or not is done by using the modulator which
consists of three different modules namely: Isolation
Module (IM), Fringe Module (FM) and Similarity
Module (SM).
A. Isolation Module
Isolation Module is used to make a decision whether
the pixel value is in a smooth region or not. If the
result is negative, we can conclude the fact that the
pixel under consideration belongs to be in a noisy-free
area. Or else if the result ispositive, this indicates that
the pixel under consideration can be a noisy pixel or it
is just situated at the edge of the object in an image. In
the given dataflow diagrams, THD, BHD, TH Max,
TH Min, BH Max, BH Min denotes Top Half
difference, Bottom Half difference, Top Half max,
Top Half min, Bottom Half max, Bottom Half min
respectively.
Where,
THD=TH Max – TH Min-----------------(3)
BHD=BH Max – BH Min ----------------(4)
At final stage, we make a temporary decision i.e. p(i,j)
belongs to a suspected noisy pixel or is noisy free.
(a) (b) (c)
Fig. 1 Dataflow of Decision in IM
B. Fringe Module
Suppose a case occurs, when the pixel is located at the
edge of an image pixel, then the Isolation Module may
detect it as a noisy pixel. In order to deal with this case,
four directions are defined with the help of arrow
markings, E1 to E4, as shown in Fig. With the
calculation of the absolute difference between f(i, j)
and the other two pixel values along the same
direction, we can determine whether there is an edge
or not in this method of image denoising. The
Dataflow of the Fringe Module is therefore as shown
below:
E1 E2 E3
E4
Fig. 2 The Directions in Fringe Module
Fig. 3 Dataflow of FM
C. Similarity Module
This is termed as the last module in decision tree
based denoising method. In the noise free areas,
generally, will be marked with the luminance values
on the window (W) and the median will always be
located at the centre whereas the impulse will be
located near to one of its ends. Thus in case if an
extreme big or small value occurs, this may show the
presence of noisy signals. Hence nine values will
SSRG International Journal of Electronics and Communication Engineering (SSRG-IJECE) – Volume 3 Issue 6 – June 2016
ISSN: 2348 – 8549 www.internationaljournalssrg.org Page 3
always be sorted in ascending order and the 4th, 5th
and 6th values will be obtained. These are always
close to median in mask (W). In order to perform the
operation, we may define with the following variables.
Maxij =6th in Wij+ Th_SMa --------------(5)
Minij =4th in Wij - Th_SMa --------------(6)
Calculation of Nmin is always same as Nmax along
with some minor modifications which can be dealt.
Threshold often affects the overall performance of the
denoising algorithms. According to our experimental
results, the thresholds Th IMa, TH IMb, Th FMa,
ThFMb, Th SMa and Th SMb are all predefined
values and are set as 20, 25, 40, 80, 15, and 60,
respectively. These are the values that are mainly
considered in this method of denoising. The dataflow
diagram of similarity module is as shown in fig.
Fig. 4Dataflow of SM
II. PROPOSED EDGE PRESERVING IMAGE FILTER
The noisy pixels are determined with the decision
tree based noise detector and these pixels are
reconstructed using an edge preserving image filter,
since edge preservation is considered to be one of the
important considerations in these denoising algorithms.
Hence, for the reconstruction of noisy pixels in an
image 8 different directional differences will be
considered i.e. from D1 TO D8. All the unaffected
pixels will not be considered or even included in the
denoising process. The possible misdetection of the
pixels will all be avoided and for all the case
explained in this method, the unaffected pixels are not
considered at all, for this and , are
included inorder to determine whether the values of d,
e, f, g and are likely to be corrupted or not respectively.
Suppose, if d, e, f, g and h are all suspected to be noisy
pixels, and no edge can be processed, the value of Pi,j
is always equal to the weighted average of luminance
values for the three previously denoised pixels and can
hence be calculated with this equation of (a+b x
2+c)/4.
V. SIMULATION RESULTS AND DISCUSSIONS
The characteristics and performance of decision
tree based denoising method was tested and verified
by using MATLAB – GUI. Simulation in MATLAB
environment is done with an adopted method of
browsing of an image into which noise with varying
intensities are being applied and hence tested and
verified. This noisy image obtained is then fed into the
denoising process. The simulation results are as
follows:
Fig. 5Noisy Image
Fig. 6 Denoised Image using DTBM
As the values of PSNR increases for the images the
filtering technique is more accurate and performs
extremely well. PSNR values are calculated using the
equation:
SSRG International Journal of Electronics and Communication Engineering (SSRG-IJECE) – Volume 3 Issue 6 – June 2016
ISSN: 2348 – 8549 www.internationaljournalssrg.org Page 4
As a result the clarity of images often increases with
respect to the values of PSNR.
VI. CONCLUSIONS
Image denoising using Gaussian and median filter
was also performed and compared. From these two
filters the image was able to be browsed from the
system and varying intensities of noise were added to
the original image in order to denoise and thus show
the efficiency of the above mentioned filters. From the
image filtering operations performed by Gaussian and
median, it was then compared with the proposed
method. Standard Deviation for different values
ranging from 20% to 50% is being tabulated. The
result Denoising Method filtering perform better than
any other techniques that have been dealt so far. The
filtering techniques like Gaussian and Median filtering
(all performed for 3X3 window) always show good
performances only for low noise density i.e., for low
noise that is being applied to the image. All the
performances that were done for these filters could be
increased, which is possible, if the window size is
increasedbut a drawback is that as the window size
increases the image tends to show more and more
blurring effect (blurriness) of an image. Hence,
Decision Tree Based Method filtering always shows
the best performances for all noises.
TABLE 1. PSNR VALUES FOR DIFFERENT IMAGES
VII. ACKNOWLEDGEMENTS
I would like to express my deepest appreciation to all
those who provided me the possibility to complete this
report. A special gratitude I give to my final year
project guide, Mr. Ramaprasad Poojary whose
contribution in stimulating suggestions and
encouragement helped me to coordinate my project
especially in writing this paper.
REFERENCES
[1] R. C. Gonzalez and R. E. Woods, Digital Image Processing.
Pearson Education, Upper Saddle River, New Jersey, 2007
[2] W. K. Pratt, Digital Image Processing. New York: Wiley-
Interscience, 1991
[3] W. Luo, ―An efficient detail-preserving approach for removing
impulse Noise in images,‖ IEEE Signal Process.Lett. vol. 13,
no. 7, pp. 413-416, July 2006
[4]I. Aizenberg, and C. Butakoff, ―Effective impulse detector based
on rank-order criteria,‖ IEEE Signal Process. Lett. vol. 11, no.
3, pp. 363-366, Mar. 2004
[5] Shiby. S, PG Scholar, Asha Sunil, Asst.Professor‖ A Decision
tree and Conditional Median Filter Based Denoising for
impulse noise in images,‖ ISSN: 2248-9622, Vol. 4, Issue
7(Version 3), pp.83-88, July 2014
[6] Sajan P. Philip, S. P. Prakash, Dr. S. Valarmathy,‖ An Efficient
Decision Tree Based Architecture for Random Impulse Noise
Removal in images,‖ International Journal of Advanced
Research in Electrical,
[7] Y. Dong and S. Xu, ―A new directional weighted median filter
for removal of random-valued impulse noise,‖ IEEE Signal
Process.Lett. vol. 14, no. 3, pp.193–196, 2007
[8]N. I. Petrovic and V. Crnojevic, ―Universal impulse noise filter
based on genetic programming,‖ IEEE Trans. Image Process.,
vol. 17, no. 7, pp. 1109- 1120, July 2008
[9] C. Y Lien, C.H Huang, P.Y Chen, Y. F Lin ―An Efficient
Denoising Architecture for Removal of Impulse Noise in
Images, IEEE Transactions on Computers, March 2012
[10]P.E. Ng and K.K. Ma, ―A Switching Median Filter with
Boundary Discriminative Noise Detection for Extremely
Corrupted Images,‖IEEE Trans. Image Processing, vol. 15,
no. 6, pp. 1506-1516, June 2006
[11] P.-Y. Chen and C.-Y. Lien, ―An Efficient Edge-Preserving
Algorithm for Removal of Salt-and- Pepper Noise,‖ IEEE
Signal Processing Letters, vol. 15, pp. 833-836, Dec. 2008
Image
Standard Deviation
(SD=0.2)
Standard Deviation
(SD=0.3)
Standard Deviation
(SD=0.4)
Standard Deviation
(SD=0.5)
Name
Gaussi
an
Medi
an
DTBD
M
Gaussi
an
Medi
an
DTBD
M
Gaussi
an
Medi
an
DTBD
M
Ga
ussi
an
Median
DTBD
M
Image1
24.38
28.85
37.36
19.91
23.54
34.61
16.10
19.03
32.41
13.
10
15.23
30.34
Image2
24.87
26.96
33.75
19.77
22.78
31.33
15.96
18.96
29.15
13.
12
15.21
27.18
Image3
24.88
26.45
34.19
19.80
23.01
32.34
16.33
19.00
31.33
13.
14
15.31
29.98
Image4
24.98
27.09
35.52
19.98
24.66
33.54
17.43
19.78
32.35
13.
45
15.98
30.78
SSRG International Journal of Electronics and Communication Engineering (SSRG-IJECE) – Volume 3 Issue 6 – June 2016
ISSN: 2348 – 8549 www.internationaljournalssrg.org Page 5
[12] Iyad F. Jafar, Rami A. AlNa'mneh ; Khalid A. Darabkh,‖
Efficient Improvements on the BDND Filtering Algorithm
for the Removal of
High-Density Impulse Noise,‖ IEEE Transactions on Image
Processing (Volume: 22, Issue: 3), March 2013
[13] Xuming Zhang, Youlun Xiong,‖ Impulse Noise Removal
Using Directional Difference Based Noise Detector and
Adaptive Weighted Mean Filter‖ IEEE Signal Processing
Letters (Volume:16 , Issue: 4 ), April 2009
[14] Dong-Hyuk Shin, Rae-Hong Park ; Seungjoon Yang ; Jae-Han
Jung,‖ Block-based noise estimation using adaptive Gaussian
filtering,‖ IEEE Transactions on Consumer
Electronics (Volume:51 , Issue: 1 ), Feb. 2005
[15]P. D. Hanlon, P. S. Maybeck,‖ Multiple-model adaptive
estimation using a residual correlation Kalman filter bank,‖
IEEE Transactions on Aerospace and Electronic
Systems (Volume:36 , Issue: 2 ), Apr 2000