You are on page 1of 5

Bonfring International Journal of Industrial Engineering and Management Science, Vol. 7, No.

2, May 2017 35

Decision Tree based Denoising Method for


Removal of Impulse Noise
K. Murugan, Dr.V.P. Arunachalam and Dr.S. Karthik

Abstract--- Impulse noise mostly causes alterations in the methods: Lower complexity techniques [3][4][5] and Higher
images. So we intend to do a denoising method for removal of complexity techniques[6][7][8]. The complexity of noise
random valued impulse noise. There are two methods involved removing algorithms based on the local window size, memory
in this scheme. To identify the noisy pixel, decision tree-based buffer and iteration times. In lower complexity techniques
impulse noise detector is used. To reconstruct the intensity fixed-size local window is used and it requires a few line
values of noisy pixels, an edge-preserving filter is used. To buffers and performs no iterations. So its computational
improve the cause of removal of impulse noise, an adaptive complexity is low. The higher complexity techniques use high
technology is used. By using this method we can yield an computations a complexity arithmetic operations, enlarging
effective performance compared to the existing methods. local window size or iterations. These techniques require long
Keywords--- Decision Tree, Denoising, Edge Preserving computational time and full frame buffer.
Filter, Impulse Noise.

I. INTRODUCTION

I MAGE processing is the study of any algorithm that takes


image as input and returns the output either in image or
parameters. It is used in various fields like medical images,
scanning process, for recognizing the face etc. Impulse noise
mostly causes alterations in the images. So removing the noise
is an important in image processing [1],[2]. Based on the noisy
pixels impulse noise is further divided into two methods: fixed
valued impulse noise and random-valued impulse noise. Those
are also called as salt-and pepper noise. In gray scale images
the range of [0,255] is assumed for random-valued impulse
noise. The effort for removing random-valued impulse noise is Fig. 1: A 3x3 Mask Centered on pi,j
very difficult due to random values of noisy pixels. Here we Decision tree is easy to understand but having great power
aimed to denoise the random-valued impulse noise in the in multiple variable analyses. It can be separated into pieces, a
affected image. decision–making process into easy decision, thus it is simple
Generally standard median filter [3] and its alternations [4] to interpret. To handle the salt-and-pepper noise there are
[5] are used for image denoising. But this may cause blur in many methods, some methods are more effective. By
the image so that noise–free pixels are altered. So, to keep analyzing this, an adaptive decision-tree-based denoising
away from the damage of noise-free pixels, an efficient method is used for denoising the random-valued impulse noise
method is used in this paper. Impulse detection and noise and to improve the quality of removal of impulse noise. In
filtering are the two steps involved in median filter. Impulse DTBDM addition and subtraction are used. From the
detector finds the noisy pixel and noise filter removes the corrupted image it removes the noise. This method gives
detected noisy pixel. By using this process we can avoid the efficient result compared to existing methods in terms of
damage. Luo proposed an alpha-trimmed mean-based method visual quality
(ATMBN)[6] which used the alpha-trimmed mean for impulse
detection and the noisy pixel value can be reconstructed by a II. THE PROPOSED METHOD
linear combination of its original value and the median of its Random–valued impulse noise looks attentively in this
local window. In [7], a different uac rank unlocks detect tir paper. 3*3 mask is assumed for denoising the image. The
(DRID) was presented. There are two types of denoising denoised pixel is placed at coordinate (i,j), it represented as p i,j
and fi,j for luminance value. In image denoising, the eight pixel
K. Murugan, Assistant Professor, Department of ECE, Tamilnadu values are split into two sets: WTophalf and Wbottomhalf.
College of Engineering, Karumathampatti, Coimbatore, India. E-mail:
WTophalf = {a, b, c, d} (1)
murugan88.kumar@gmail.com
Dr.V.P. Arunachalam, Professor and Director, SNS College of Wbottomhalf = {e, f, g, h} (2)
Technology, Coimbatore, India. E-mail:drvparunachalam@yahoo.com
Dr.S. Karthik, Professor and Dean, Department of CSE, SNS College of
Technology, Coimbatore, India. E-mail:profskarthi@gmail.com
DOI : 10.9756/BIJIEMS.8349

ISSN 2277-5056 | © 2017 Bonfring


Bonfring International Journal of Industrial Engineering and Management Science, Vol. 7, No. 2, May 2017 36

confirm the result. This similarity module compares the


relation between current pixel and its neighboring pixels. If
the result is true, pi,j is a noisy pixel; else, it is noise free.

Fig. 4: Non Smooth Region in Lena


Isolation Module
In Isolation Module the values of the pixel in a smooth
region should be slightly varying. The difference between its
neighboring pixel values are less. If the values, edges, or
blocks in this region are noisy and the value should also be not
same. From this, we can find out whether current pixel is an
Fig. 2: The Dataflow of DTBDM
isolation point by watch attentively the smoothness of its
The components for DTBDM are decision- tree based surrounding pixels.
impulse noise detection and edge preserving filter. By using
the correlation between pixel pi,j and the neighboring pixels Tophalf_diff= TopHalf_max - TopHalf_min (3)
and the decision tree, the detector determines whether pi,j is a
noisy pixel. If the pi,j is noisy pixel filter generates the Bottomhalf_diff= BottomHalf_max - (4)
reconstructed value. If pi,j is noise free pixel the value remains BottomHalf_min
same.
True, if(Tophalf_diff ≥Th_IMa)
DecisionI= or(BottomHalf_diff≥Th_IMa)

False, otherwise (5)

IM_TopHalf
Fig. 3: Smooth Region in Lena
True,if(|fi,j-TopHalf_max|≥Th_IMb)
A. Decision Tree based Impulse Detector or (|fi,j -TopHalf_max|≥Th_IMb)
To find out pi,j is a noisy pixel or noise free pixel, the =
correlations between pi,j and its neighboring pixels are taken False, otherwise (6)
for consideration. By analyzing this process it separates into IM_BottomHalf
many ways comparing the relation between current pixel and
its neighboring pixels, finding whether the current pixel on a True,if(|fi,j-BottomHalf_max|≥Th_IMb)
fringe, to watch attentively the degree of isolation at current
= or (|fi,j -BottomHalf_max|≥Th_IMb)
pixel.
In decision tree based impulse detector they are three False, otherwise (7)
modules-Isolation module (IM), Fringe Module (FM),
Similarity Module(SM). By using these three modules a true, if (IM_TopHalf= true)
decision tree is implemented. For analyzing the status of pi,j
decision tree is used. They are different equations for different
Or(IM_BottomHalf=true)
modules. First module is to find whether the pixel value is in
a smooth region. If the pixel is in smooth region, we decide Decision II =
that the current pixel belongs to noisy free. Else, if the pixel is false, otherwise. (8)
not in smooth region, it is a noisy pixel or just situated on the Noisy pixel has less similarity compared with the
edge. neighboring pixels. Noisy pixels are called isolation point.
The difference between the noisy pixel and its neighboring
To confirm the result, fringe module is used. If the current
pixels value is large from this we first find out the maximum
pixel is placed on an edge, the result will be noise free; else,
and minimum luminance values in Wtophalf called as Top Half-
the result will be noisy pixel. In case isolation module and
max, Top Half-min, and find the difference between them,
fringe module cannot able to find the current pixel belongs to
called as Top Half-diff, like that for Wbottomhalf , named as
noisy free or not. In this situation similarity module is used to
Bottom Half-min Bottom half-max and the difference is called

ISSN 2277-5056 | © 2017 Bonfring


Bonfring International Journal of Industrial Engineering and Management Science, Vol. 7, No. 2, May 2017 37

as BottomHalf-diff. These two are compared with a threshold nine values in ascending order and obtain the fourth, fifth and
Th-IMa this is used for deciding whether the surrounding sixth values which are near to median. The fourth value is
region belongs to a smooth area. named as 4th in Wi,j and fifth values as Median InWi,j and sixth
values called as 6th in Wi,j

Fig. 5: The Difference between Noisy and Neighboring Pixels Fig. 6: Four Direction in DTBDM
in Lena Maxi,j = 6th in Wi,j + Th_SMa
th
Fringe Module Mini,j = 4 in Wi,j + Th_SMa. (14)
To find out the status of pixel Pi,j , Maxi,j, Mini,j is used. For
If Pi,j has a large difference with neighboring pixels, it
getting effective decision, we do some changes as
called as noisy pixel or just placed on an edge. It is hard to
decide that it is edge or noisy pixel. In this case, we define Maxi,j if(Maxi,j≤ MedianInWi,j
four directions, from E1 to E4. For example if we take E1 as Nmax= +Th_SMb)
direction to determine whether there is an edge or not it is MedianInWi,j
determined by calculating the absolute difference between fi,j . +Th_SMb , otherwise (15)
False, if (|a –fi,j|≥ Th_FMa)
FM_E1 = or (|h –fi,j|≥ Th_FMa)
or (|a-h|≥Th_FMb) Mini,j if(Mini,j≤ MedianInWi,j
True, otherwise (9) Nmin = -Th_SMb)
MedianInWi,j
False, if (|c–fi,j|≥ Th_FMa) -Th_SMb , otherwise (16)
FM_E2 = or (|f –fi,j|≥ Th_FMa)
Or (|c-f|≥Th_FMb) To decide that Pi,j is a noisy free pixel, if fi,j is between
Nmax and Nmin . To build the reconstructed value, edge
True, otherwise (10)
preserving filter is used. Else, the output will be the original
value of fi,j. The equation is ,
True, if (fi,j≥Nmax) or (fi,j Nmin)
False, if (|b–fi,j|≥ Th_FMa) Decision IV=
FM_E3 = or (|g –fi,j|≥ Th_FMa) false, otherwise (17)
Or (|b-g|≥Th_FMb) In the proposed method, threshold affects denoised images
True, otherwise (11) for a better detection result. However it is not easy to derive an
optimal threshold through analytic formulation. The fixed
False, if (|d–fi,j|≥ Th_FMa) values of thresholds make our algorithm simple. According to
our experimental results the threshold Th_IMa, Th_IMb,
FM_E4 = or (|e –fi,j|≥ Th_FMa) Th_FMa, Th_FMb, Th_SMa and Th_SMb are all predefined
Or (|d-e|≥Th_FMb) values and set as 20, 25,40,80,15 and 60, respectively.
True, otherwise (12) B. Edge-preserving Image Filter
Edge preserving image filter is used to filter out the noisy
False, if(FM_E1)or(FM_E2) pixel.In this D1 to D8 are taken as eight directional differences
because it is to reconstruct the noisy pixel value. The noise
Decision III = or (FM_E3) or (FM_E4) free pixels are used to avoid possible misdetection. Directions
passing through the suspected pixels are discarded to reduce
True, otherwise (13) misdetection. So that Maxi,j , Mini,j are used, to find out the
values of d,e,f,g and h are likely corrupted if pixel is affected
Similarity Module by noise, we need not to be taken the direction including the
suspected pixel. In second block, noisy pixels are d,e,f,g,h and
The final module is similarity module. The median is
no edge can be processed. Luminance values is fi,j is calculated
placed in the center but the impulse value is placed near the
as (a+b × 2+ c)/4. In third block the smallest one (dmin) is
pixel. If there are extreme big or small values that implies the
chosen among the eight directional difference.
possibility of noisy signals. From this analysis, we arrange

ISSN 2277-5056 | © 2017 Bonfring


Bonfring International Journal of Industrial Engineering and Management Science, Vol. 7, No. 2, May 2017 38

If f^i,j obtain the correct edge, it will situate at the median


of b ,d,e and g because of the spatial relation and the
characteristic of edge preserving . Otherwise, the values of f̄ i,j
will be replaced by the median of four neighboring pixels
(b,d,e and g). We can express f̄ i,j as,
f̄ i,j = Median( f^i,j, b,d,e,g) (20)

III. RESULTS AND DISCUSSION


The performances of the proposed DTBDM filter can be
compared with the Median filter to compare and estimate the
performances of this method. The peak signal to noise ratio is
used to compare the performance of DTBDM and Median
filter.
Visual Analysis
Fig. 7: Dataflow of Edge Preserving Image Filter
The equations are,
D1 = |d-h|+|a-e|
D2 = |a-g|+|b-h|
D3 = |b-g|×2
D4 = |b-f|+|c-g| (18)
D5 = |c-d|+|e-f|
D6 = |d-e|×2 Fig. 9: Noise Affected Image
D7 = |a-h|×2
D8 = |c-f|×2

(a+d+e+h)/4, if Dmin=D1
(a+b+g+h)/4, if Dmin=D2
(b+g)/2, if Dmin=D3
fˆi,j = (b+c+f+g)/4, if Dmin=D4
(c+d+e+f)/4, if Dmin=D5
(d+e)/2, if Dmin=D6 Fig. 10: Denoised Image Using DTBDM Method
(a+h)/2, if Dmin=D7 Fig 10. Shows the denoised image using DTBDM method.
(c+f)/2, if Dmin=D8 (19) It shows that the resolution of the image is increased while
In last block, smallest directional difference exist an edge denoising. Table 1 shows the qualitative analysis to validate
in its direction. Hence, it is treated as f^i,j. the effectiveness of the DTBDM, we compared it with the
median filter and by using different densities of images. The
result shows that the DTBDM performs well both in denoising
and edge preservation.
Table 1: Performance Comparison between two Methods
METHOD DENSITY PSNR
0.02 37.3798
Median Filter 0.05 32.6026
0.08 30.8020
DTBDM 0.02 56.9211
0.05 53.2693
0.08 49.9693

IV. CONCLUSION
A decision tree based denoising algorithm for efficient
removal of random valued impulse noise is proposed in this
paper. The approach uses the decision tree based impulse
Fig. 8: Eight Directional Differences of DTBDM
detector to detect the noisy pixel and employs an effective

ISSN 2277-5056 | © 2017 Bonfring


Bonfring International Journal of Industrial Engineering and Management Science, Vol. 7, No. 2, May 2017 39

method to locate edges with adaptive skill; the quality of the


reconstructed image is notably improved. Our extensive
experimental results shows that the performances of our
proposed technique is better than the previous median filter
methods and comparable to median filter it has better visual
quality. Therefore it is very suitable to apply many real time
applications.

REFERENCES
[1] R.C. Gonzalez and R.E. Woods, Digital Image Processing, Pearson
Education, 2007.
[2] W.K. Pratt, Digital Image Processing, Wiley-interscience, 1991.
[3] T. Nodes and N. Gallaghar, “Media Filters: Some Modifications and
Their Properties”, IEEE Trans. Acoustics, Speech, Signal Processing,
Pp. 984-993, 1982.
[4] SJ. Ko and YH. Lee, “Center Weighted Median Filters and Their
Applications to Image Enhancement”, IEEE Trans. Circuits Systems,
Vol. 38, No. 9, Pp. 984-993, 1991.
[5] T. Sun and Y. Neuvo, “Detail-Preserving Median Based Filters in
Image Processing”, Pattern Recognition Letters, Vol. 15, Pp. 341-347,
1994.
[6] W. Luo, “An Efficient Details-Preserving Approach for Removing
Impulse Noise in Images”, IEEE Signal Processing Letters, Vol. 13,
No. 7, Pp. 413-416, 2006.
[7] I. Aizenberg and C. Butakaff, “Effective Impulse Detector on Rank-
Order Criteria”, IEEE Signal Processing, Letters, Vol. 11, No. 3,
Pp.363-366, 2004.
[8] H. Yu, L. Zhao and H. Wang, “An Efficient Procedure for Removing
Random Valued Impulse Noise in Images”, IEEE Signal Processing
Letters, Vol. 15, Pp. 922-925, 2008.

ISSN 2277-5056 | © 2017 Bonfring

You might also like