• Embed Doc
  • Copy Link
  • Readcast
  • Collections
  • CommentGo Back
Download
 
Anita Sahoo et al /International Journal on Computer Science and Engineering Vol.1(3), 2009, 217-221217
Fuzzy Weighted Adaptive Linear Filter for Color Image Restoration Using Morphological Detectors
Anita Sahoo
Department of Computer Science & EngineeringJSS Academy of Technical Education, NOIDA, Indiaanitasahu1@gmail.com
Dr. Manu Pratap Singh
Department of Computer ScienceICIS, Dr. B. R. Ambedkar University, Agra, Indiamanu_p_singh@hotmail.com 
 Abstract
— A novel scheme for restoration of color images ispresented. The proposed scheme has two phases; in the firstphase it detects the pixels corrupted with impulse noise byemploying the tools of mathematical morphology and the degreeof noisiness is labeled by a fuzzy membership function.Subsequent phase uses a linear weighted filter that attempts toreduce noise in an efficient manner. Experimental results indicatethat the proposed scheme provides a better restorationperformance than many state of the art filters used for the same.The scheme is capable of preserving color and image details,while suppressing impulse noise effectively.
 Keywords- Image Restoration, Impulse Noise, Morphological Operations, Fuzzy Functions, Adaptive Linear Filter, Images in RGB and Lu*v* color space.
I.
 
I
 NTRODUCTION
 With the wide use of color in many areas such asmultimedia applications, biomedicine, internet and so on, theinterest on the color perception and processing has beengrowing rapidly. Color, as we know is a powerful descriptor and helps us interpret and identify the objects in a picture moreclearly than a gray scale image. They provide the picture withan element of closeness with the real world which obviously,isn’t black and white. To add to it, there are thousands of shades our eyes can recognize thus elaborating the scope of this paper to all of these instead of restricting it to a few dozenshades of gray.Impulse noise is characterized by discontinuities in theform of very high or very low values in a dynamic range thusgiving it an overall granular kind of appearance. The standardmedian filter has been widely used for removing impulse noise.The problem with traditional linear or median filters and it’slike [1], [4] was that they treated all the pixels even theundisturbed ones in the same manner. Though they achievenoise suppression but at the expense of considerable distortionin image features. Improvements are the adapted switchingfilters [5], [6] [7]. Mathematical morphology has been widelyused in various image processing tasks. Morphological filtersare nonlinear signal transformations that locally modifygeometric features of signals. Jinsung Oh et.al [8], Deng, Z.F.et.al [9] and Z.P. Yin et.al [10] have shown the utilization of morphological filters in the perspective of impulse noisereduction from grayscale images. Jinsung Oh et.al [8] used aranked directional morphological operator, Deng, Z.F. et.al [9]used opening closing sequence filter and Z.P. Yin et.al [10]used conditional morphological operators. The morphologicaloperators used in these filters differ from each other by thestructuring elements they use or the sequence of operation they perform. The experiments have shown that the morphologicalfilters are able to clean impulse noise effectively, but then the biases that are introduced by the operators tend to lose imagedetails. Some of the two phased schemes utilize fuzzytechniques for noise detection. Like Wenbin Luo [11] used ahistogram based fuzzy noise detection technique. Then thefuzzy membership values are used in a modified median filter that filters out the noise. The filters we discussed so far areused for impulse noise reduction from gray scale images. Deng,Z.F et.al [12] used Median Controlled Adaptive RecursiveWeighted Median Filter for restoring color image corrupted byimpulse noise.In this paper, a new class of filter for color image processing is developed. The Morphological Operations inRGB color space are introduced, which are an extension of mathematical morphology from gray-scale images to color images. One of the key features of fuzzy logic is its ability todeal with the typical uncertainty, which characterizes any physical system. Indeed, fuzziness affects many aspects of image processing such as input signal can be noisy andincomplete. In the proposed scheme a fuzzy membershipfunction is used that gives each pixel a fuzzy flag indicatinghow much a pixel looks like an impulse. These fuzzy flags areused as weights corresponding to the pixel in a modifiedadaptive linear filter. The experimental results have beencompared with various other filters indicating the efficiency of the proposed technique. The measures such as Peak Signal to Noise ratio (PSNR) and Normalized Color Difference (NCD)are used to prove the noise reducing and detail preservingcapability of the proposed scheme.II.
 
P
ROPOSED
S
CHEME
 The proposed method uses a sequence of morphologicaloperations for noise classification. A 5X5 flat structuringelement of value one has been used for morphological noisedetection. The flat structure of it reduces the complexity of theoverall technique. After the sequence of morphologicaloperations, a fuzzy membership function as shown in figure(1), is used to find how much noisy a pixel is. Then thesemembership values corresponding to any pixel is used as theweight in the filter. Generally in the fixed small window sizefilters, the amount of noise density filtered will be very less, for filtering high density noise the window size of the filter may
ISSN : 0975-3397
 
Anita Sahoo et al /International Journal on Computer Science and Engineering Vol.1(3), 2009, 217-221218increase. This may lead to blurring in the output images. Inorder to overcome this, here we use an adaptive weighted linear filter that restores the color image effectively.
 A.
 
 Morphological Noise Detection
A color pixel at (x, y) in a given domain of color image can be described by a color vector (R(x, y), G(x, y), B(x, y)) inRGB color space, where
2
),(
 Z  y x
∈
. In order to process thecolor image, it is first decomposed into its three componentimages. Now each component image is treated as a separateintensity image. All the operations are performed on thesecomponent images independently. Noise detection employing asequence of morphological operations is performed in twostages.
 B.
 
Stage-I of Noise Detection
This involves the application of basic morphologicaloperations such as Dilation and Erosion on the image functionf(x, y) that is the value of the component (R, G or B) under consideration at position (x, y). Considering, b(x, y) as thestructuring element, Dilation
)(
b f 
⊕
and Erosion
)(
b f 
Θ
 operations can be represented by equation (1) and (2)respectively.
}),( ;)(),(),(*),(max{),)((
b f n
 Dt s Dt  ys xt sbt  ys x f  y xb f 
∈∈−−−−=⊕
(1)
}),( ;)(),(),(*),(min{),)((
b f n
 Dt s Dt  ys xt sbt  ys x f  y xb f 
∈∈++++=Θ
(2), where Df and Db denote the domain of the image f and thedomain of the structuring element b, respectively. As shown,the above operations find out the minimum and maximumcolor component pixels in their neighborhood. Now as the saltand pepper impulse noise itself implies significantdiscontinuities in their neighborhood, they can be detectedcomparing the results of dilated or eroded image with theoriginal one. If the values after dilation or erosion turn out to bethe same as before, the pixel is classified as the noisy one withthe noise flag set. Else it is considered a noise free pixel. Usingthis concept a binary array B(x, y) containing the noise flagscan be derived as explained in equation (3).
⎩⎨⎧Θ=⊕= =
otherwise y xb f  y x f or  y xb f  y x f   y x B
0),)((),(),)((),(1 ),(
 
(3)
 However, there are some issues with this. Sometimes theedge pixels or other such sharp discontinuity areas whichotherwise form a part of the image itself are also misclassifiedas being noisy pixels. Therefore, a second stage of noiseclassification is also employed.
C.
 
Stage-II of Noise Detection
Based on the basic operations of Erosion and Dilation,Opening
)(
b f 
o
and Closing
)(
b f 
•
operations are definedas given in equation (4) and (5) respectively.
),)()((),)((
 y xbb f  y xb f 
⊕Θ=
o
(4)
),)()((),)((
 y xbb f  y xb f 
Θ⊕=•
(5)Combining the opening-closing operators, the localcharacteristic at position (x, y) is determined as in equation (6).
),(2),)((),)(( ),(
 y x f  y xb f  y xb f   y x D
−•+=
o
(6)Whilst the local characteristic value D for a pixel is more, itshows that it differs from its neighbourhood significantly.Suppose, the pixel under consideration is an edge pixel, thenthe connecting edge pixels in its neighborhood also have thesimilar characteristics. Using these concepts, the pixel withhigher local characteristic value is suspected to be an impulsenoise. To decide, how much noisy a pixel is we make use of afuzzy membership function. The binary array B(x, y)containing the noise flags, is then revised to contain the fuzzyflags that indicate the degree of noisiness.
 D.
 
Computation of Fuzzy Weights
A fuzzy system is characterized by fuzzy variables whichare members of a fuzzy set. A fuzzy set based on the concept of  partial membership is a generalization of a classical set. Let F be a fuzzy set defined on the universe of discourse U. Thefuzzy set is described by the membership
)(
uF 
µ 
that maps U tothe real interval [0, 1]. A membership of value 0 signifies thefact that the element
U u
∈
does not belong to the set F, amembership of value 1 signifies that the element
U u
∈
 belongs to the set F with full certainty. Whereas, amembership of any real value from 0 to 1 represents theelement u to be a partial member of the set F.In the proposed scheme, a fuzzy membership function asgraphically shown in figure [2] is employed. The universe of discourse is the local characteristic values D as computed usingthe equation (6). For any value
 Dd 
∈
the fuzzy membershipvalue is computed as,
⎪⎪⎩⎪⎪⎨⎧≥<>−−≤=
maxmaxminminmaxminmin
1ddd0
d d if d d or d d if  d d d if 
d 
µ 
(7), where
min
d 
and
max
d 
are two preset parameters. Massiveexperiments indicate that for most of the cases, images can be
ISSN : 0975-3397
 
Anita Sahoo et al /International Journal on Computer Science and Engineering Vol.1(3), 2009, 217-221219restored with high quality with the proposed algorithm when15
min
=
d 
and25
max
=
d 
. For all noise-free pixels,
d 
µ 
is set tozero.As discussed earlier, these noise detection techniquesdefined in two stages are to be applied on the three color components of the image independently. The sequence of above described operations is used to realize accurate noisedetection. Once the noise is detected we move to the next phase, where the noise will be eliminated by applying anadaptive weighted linear filter.
Figure 1: A Continuous Fuzzy Membership Function
 E.
 
 Noise Elimination
Adapted switching filters improve the efficiency by freeingthe filter from the constraints imposed by a fixed window size.Adaptive switching filters have the capability to alter thewindow size according to the presence of noise affected pixelswithin the filter window. The adaptive mean filter used in thescheme, determines the filtering window size for everydetected noise pixel based on the number of its neighboringnoise-free pixels. For each detected noise pixel (x, y), thefiltering window with the size of (2Lf+1)*(2Lf+1) centeredabout it is used. Starting with Lf=1, the filtering window isiteratively extended outwards by one pixel in its four sides untilthe number of noise-free pixels within this window is at leastone. The output of the adaptive mean filter applied at position(x, y) is obtained by:
),()1(),(*)h(x,
 y x f  y xm y
d d 
µ µ 
α 
−+=
(8), where m
α
(x, y) is the alpha-trimmed mean valuecalculated as:
⎣ ⎦
⎣ ⎦⎣ ⎦
∑
−+=
=
y) p(x,),( 1y) p(x,
)(y) p(x,2-y) p(x, 1y)(x,m
α α α 
α 
 y x pk 
k F 
(9), where p(x, y) is the number of noise pixels within thefiltering window,
α
is the trimming parameter (0<
α
<0.5) and F(k) denoted the kth data item in increasing order samples of noise free pixels. After each component of the color image inRGB space have been processed by means of the abovementioned techniques, they are concatenated into one,rendering a single restored color image.III.
 
E
XPERIMENTAL
R 
ESULTS
 The proposed method has been simulated usingMATLAB7. For the experimentation a 5X5 flat structuringelement is chosen, while15
min
=
d 
and25
max
=
d 
, and
α
has been considered to be 0.3. A number of color images have beencorrupted by impulse noise with the noise density varying from0.1 to 0.9 and the proposed scheme has been applied for therestoration. The results have been compared with some of therecent filters [1], [2], [3], [4], [12] used for impulse noisereduction from color images. The measure used for comparisonis the Peak Signal to Noise Ratio (PSNR), computed as
⎟⎟ ⎠ ⎞⎜⎜⎝ ⎛ =
 MSE PSNR
210
255log10
 
(10)
 
 MN oo MSE where
 M i N  j ji ji
∑∑−=
= =
11,',
,
 
(11)M and N denote the image dimensions.
',
 ji
o
and
 ji
o
,
show theoriginal and restored image pixels located at (i,j) respectively.Since RGB is not a perceptually uniform space in the sense thatdifferences between colors in this space do not correspond tocolor differences perceived by humans, the restoration errorsare often analyzed using the perceptually uniform color spaces.In this work, the CIE LUV color space is used and thenormalized color difference (NCD) defined as:
∑
=
−+−+−=∆
 N i xioi xioi xioi
vvuu L L  N  E 
12**2**2**
)()()( 1
 
(12)
∑
++∆=
2*2*2*
)()()(
oioioi
vu L E  N  NCD
 
(13)
 
where L* represents lightness values and u*,v*chrominance values corresponding to original oi and restored xisamples expressed in the CIE LUV color space. Various filtersas mentioned earlier are applied on Lena.jpg and Zelda.jpg.Figure (3) and (5) shows the comparison of these filters withthe proposed scheme based on PSNR value. It isstraightforward to see from the figures that the proposedmethod provides significantly higher PSNR values than other compared filters. Subjective evaluation done in figures (2), (4),(6) and (7) further states the noise reduction capability of thescheme in different types of color images. Normalized Color Differences (NCDs) computed for different images are shownin figure (8). Lower NCD values show better preservation. Thescheme is able to preserve color attributes significantly evenfor the images with high variations.
µ
d
0.501d
min
d
max
D
ISSN : 0975-3397
of 00

Leave a Comment

You must be to leave a comment.
Submit
Characters: ...
You must be to leave a comment.
Submit
Characters: ...