Real-color Image Denoid and Enhanced Synchronously Bad on Wavelet
Transform
Han Li-na , Geng Guo-hua, Xiong Jie
Institute of Visualization TechnologyˈNorthwest University (NWU)Xi’an, China
Abstract
Real color images are images with noi. The application of gray-scale image enhancement methods in the three-channel of images (R,G,B), on the one hand, incread noi, on the other hand, tended to have a color deviation, which impacted the effect of enhancement. Therefore in view of the difference of human visual system nsitivity of the hue, saturation, luminance affects the image noi differently on the hue, saturation, and luminance. It propod a method of image denoid and enhanced in HSV space. In this paper, according to the result of influence of V and S by noi , we transform a real color image from RGB space to HSV space at first, then maintain hue channel unchanged, decompo v channel into high frequency and low frequency information using wavelet transform. And then it directl
y compress dynamic range of the j-scale low frequency information bad on a new Illumination—Reflectance Model ,us Byes-estimation threshold to denoi in all scale high frequency information. Saturation channel is denoid by some formula from the relationship of S, V and R, G, B. At last, it transforms into a real color images from HSV space to RGB space. Experiments show that image has a good dynamic compression range and slight color deviation, which improve the visual effects and the true effects of color.
1. Introduction
Becau of the color images widespread application in the multimedia, biomedical, internet, video etc, people pay more attention to color images processing. However, with the impact of various factors or conditions, the color images are often the feature of dark color, low contrast, non-highlight details, and noi, which not only affects the visual effects, but also difficultly identify and distinguish some images. Therefore, it often needs to enhance images before dealing with them. Over the years, it is propod that image enhancement algorithms are for most of the gray-scale image enhancement, but for few of color image enhancement. If we directly apply the gray-scale image enhancement algorithms to the three-channel of color images for enhancing them, color may decline. The reason is that the high correlation of R,G,B, the respectively unreasonable enhancemen
t of the red, green and blue channels, it will result deviation from the original color.
The noi in the color images will be enhanced as well as color images are enhanced. Retinex only has a certain color image enhancement, but is powerless to do anything denoising. Most papers related to the image denoising either talk about gray-scale image or not take denoising and enhancement into account in company. In recent years the application of wavelet theory is a very wide range of technologies[1~3]. Becau of its ability to lf-adaptive lection of low-frequency or high-frequency signal, it has a good time-frequency characteristic, we u wavelet transform for color images to enhance and denoi at the same time.
So the key problems of color image enhancement are that how to keep Hue Invariability and which method is suitable for image enhancement when Hue Invariability is kept and how to accomplish color images enhancement and denoising synchronously.
The work prented in this paper is transforming RGB color space to HSV color space, remaining H channel, respectively enhancing and denoising the S channel and V channel. The dynamic range of color images, the different impact on color saturation and luminance by noi are all considered.
2. HSV color space and the transformation formula
HSV color space that evolved three-dimensional color space CIE describes colors by urs’ intuitive methods. It is not only more suitable for description of human color perception than in RGB space, but also the effective paration of the color, the saturation and brightness of the color image ,which has brought great convenience for follow-up enhancement.
RGB space to HSV space conversion formula:
)
,
,
max(
max B
G
R
)
,
,
min(
psoc
min B
鳄鱼英语G
vdm
R
°
°
¯
°
°
®
¯
®
t
u
,
360
,
60
H
H
H
H
H
H
max
min)
(max
S
)
,
,
max(B
G
R
V
B
G讲座免费下载
R
H
G
R
B
H
R
圣诞节英文祝福语G B
H
max
,
min)
(max
)
(
4
max
,
min)
(max
)
(
2
max
,
min)
(max
)
(片假名翻译
2009 International Conference on Artificial Intelligence and Computational Intelligence
3. Illumination—reflection model and wavelet transformation
3.1. Illumination—reflection model
The foundations of Retinex are the color invariance
and the illumination-reflectance model. As the illumination-reflectance model reveals the physical theory of imaging, An image can be described
as:, Where
is the illumination component of the image and determined by all light sources in scene.
Where is the reflection component of the image and determined by all objects in scene[4,6]. ),(y x f ),(y x r ),(),(y x i y x f u f ),y 1),(0 y x r (0x i The illumination
is characterized by slow change, and reflection caud by mutations, in
particular the edge of some objects. If we do the Fourier
transform of the image, lect the appropriate filter function, it will compress the high dynamic range of
color image or contrast enhanced images, but noi in the
high-frequency information are also incread and The
images enhanced by Retinex will have ‘halos’ and the details may be also blurred. ),(y x i ,(y x r )
3.2. Wavelet transformation
When people look at a scene with bright lights, our eyes be able to automatically adjust themlves to obver the details in multi-scales. The multi-resolution of wavelet is similar to the function of human eyes.
With the continuous wavelet transform of image, The high-frequency parts of wavelet transform can prerve the details of image, the low-frequency prerve the slow-varying parts determining the dynamic range of images. That is, through wavelet transform the dynamic range of image decided b
y low-frequency parts be effectively compresd the image energy, the details of image by high-frequency parts effectively avoid the ‘halos’ of image enhancement. The details must be loss when the low-frequency parts are attenuated, but some details are stored in the high-frequency parts very well. Using inver wavelet transform (IWT) the reconstructed images will have more details.
purpleImage decompod by multi-scale wavelet as shown in Fig.1 (the three-scale) [3].The scales of wavelet transform
is decided by the imagery details which we need.
3.3. Wavelet transformation and illumination—reflection mode
There is the relationship of (7) in 2-D multi-resolution analysis [3].
11 j j j W V V (7) Where is which is the low frequency part of
image decompod by wavelet transform,
is compod of , and which are the high frequency parts.j V j cA )(1h j cD 1 j W )(1
v j cD )(1d j cD j is the scale level of wavelet transform,ƻˇ is synthesis of operation symbols. 1
1
j j j V V W (8) j
j j j j j j W W V W W V V 11111)( (9)
1210W W W W V V j j j (10) We can assume: 121W W W W W j j (11) W V V j 0 (12)
Compared (12) with illumination—reflection mode, As is the constant imagery part and the illumination imagery part can be also taken as constant imagery part, can be known as . Becau W
is compod of all detail parts of image and the reflectance imagery part can be also taken as variance imagery part which is mainly compod of the details of image, W can be known as . The illumination-reflectance model can be approximately described by wavelet transform as :
j V ),(y x i j V ),(y x i r ),(y x r ),(y x ),(),(),(y x W y x V y x f j (13)
4. Noid color image enhanced and denoid
Images are more or less generally contain noi, the noi can be divided into Gaussian noi, Poisson noi, particles noi[1,2,3]. In this paper, take additive Gaussian noi as an example.
世博会 英文E
E E
'
' ' B B G
G R R '''
Where
E
'is Gaussian nois 2
e,),0(V N ,
',B is
the value o B with noi in RGB color space. From (1)(2)(3)(4)(5)(6)(14),the following equation is as following:
','G R f G R ,,
H H '
)
',','max())
',','min()',','(max('B G R B G R B G R S Fig.1. Image decompod by wavelet (three scales) )',','max('B G R V
Where 'H ,',
is the value of S 'V H ,S ,V channel with noi in HSV color space.
The following conclusions can be drawn From (15)(16)(17), the nois of image have an effect on saturation and luminance, and have no effect on hue.
Therefore bad on the human visual system for different nsitivity of H,S,V and the noi to different impact on the H,S,V, it propod the denoid and enhanced method in the HSV color space. The method steps are as: maintain hue channel, denoi and enhance the V channel to adjust the whole image of the contrast and dynamic range, denoi S channel to enhance color resolution.
4.1. V channel denoid and enhanced
Through The multi-scale Wavelet decomposition, the image details can be prerved in the scale of the high-frequency parts, the same is noi. So the work is de-noising for the high frequency noi, while the dynamic range of compression for low frequency parts, Then getting the value of V channel by inver wavelet transform.
4.1.1. Denoising for high frequency of V channel
Wavelet threshold de-noising is the processing that ts threshold for the wavelet decomposition coe
fficients of the high-frequency to achieve de-noising. The two choice of threshold are hard threshold and soft threshold[3].
°¯°® t O O O k j k j k j k j k
j W W W W ,,,,,,0),)(sgn(ˆ (16)
x n G
O
/22
w (17)
Where is noi standard deviation, 2
n
w x
G is signal standard deviation.
Becau soft threshold method (16) overcomes oscillation in the signal reconstruction caud by hard threshold, wavelet soft-threshold method and Byes estimation threshold (17) is suggested to deal with noi.
By applying the algorithms above which deal with noi, the result showed in the fig.2 (a) (b) (c) (d). Fig.2 (a) was the V channel of original image. Fig.3 (b)(c)(d) was clouds parts of the V channel of original ,noid and denoid image in magnification 200% in order to clearly e the detail comparison. By comparison of Fig.2(b) (c),we could find the gauss noi in the image (c) be clearly en around the clouds and mountain. By
comparison of the three figures, we could find that the
noi was reduced obrvably.
4.1.2. Compressing the dynamic range of V channel for low frequency denoising for high frequency of V channel
According to illumination-reflectance model, we will compress the dynamic range of image as long as we attenuate the energy of the low frequency of image. Fig 2(a) is V channel of original image. Fig 4 is the imagery Fourier transforms. As the energy of image mainly distributes in the four corners of Fig5, it can be attenuated by lowpass frequency filter. Gaussian lowpass filter is described as (19), Gaussian lowpass filter shown as Fig 6 is able to attenuate the energy of image and is beneficial to the image enhancement [3,5].
e rH v u H D v u D c u ),()
/),((2
02
Where M ǃ is the size of image, r N H is the ratio coefficient and 45.0 rH , is a constant which determines the inclined plane of filter, is cut-off
frequency,c 0D 2
2)()2
(),v M
u v u
2
N
(D DŽ
Fig.4.Noid image Fig.3.Original image (PSNR=31.8dB)
Fig.6.Gaussian lowpass filter
4.2. Denoising for S Channel
According to (5)(6)(16)(17),equation (20) is
HVWDEOLVKPHQW ġ
'
)',','max(),,max(),,max()),,min(),,(max()
',','max()),,min(),,(max()
',','max())',','min()',','(max('V V
S B G R B G R B G R B G R B G R B G R B G R B G R B G R B G R B G R S
(c=3,rH=0.45,
Fig.5. Imagery Fourier transform D0=0.5u median(median(D))
(a) (b) (c) (d)Fig.2.V channel of original image (a) and clouds parts of V channel of original, noid and denoid image in magnification 200% (b) (c) (d)
Equation (20) show the algorithms which deal with
noi in S channel of image, the result showed in the fig.7
(a) (b) (c) (d). Fig.7(a) was the S channel of original
image. Fig.7 (b)(c)(d) was clouds parts of the S channel
of original ,noid and denoid image in magnification
200% in order to clearly e the detail comparison. By
comparison of Fig.7(b) (c) (d), we could also find that the
gauss noi in the image (c) be clearly en around the
clouds and mountain and the noi was reduced
obrvably.
4.3. The Float Chart of Our Algorithm Other
Recommendations
5.Experiment and analysis
From the Fig.9~Fig.23 shown, our methods are
simulated in MATLAB and applied to Fig.9 ,Fig.15 and
Fig.21. Fig.10, Fig.16 and Fig.22 are involved Gaussian
noi with variance 0.0002. The effect of enhancement is
distinctly visible in the images. Let us examine the
Fig.9~Fig.14.The noi be en mainly around the cloud
or the sky. This is becau human eyes are less nsitive Fig.8. the float chart of our algorithm
Fig.9.Original Image Fig.10.Noid Image
(PSNR=37.6dB)
Fig.11.Denoid Image
by Wavelet
(PSNR=40.1dB)
Fig.21.Original Image Fig.22.Noid Image
(PSNR=37.8dB)
Fig.23.Denoid and
Enhanced Noid
image (PSNR=30.4dB)
Fig.14.Denoid and
Enhanced Noid
image(PSNR=31.7dB)
Fig.12.Enhanced Image Fig.13.Enhanced
Noid Image
(PSNR=27.6dB) Fig.7.S channel of original image (a) and clouds parts of S channel of
original, noid and denoid image in magnification 200% (b) (c) (d)
Fig.16.Noid Image
(PSNR=38.1dB)
Fig.17.Denoid Image
by Wavelet
Fig.15.Original Image
(PSNR=41.4dB)
Fig.18.Enhanced Image Fig19.Enhanced
Noid Image
(PSNR=25.2dB)
Fig.20.Denoid and
Enhanced Noid
学习英语软件image(PSNR=27.1dB)
to high and low brightness of the background details of the texture than to middle-high brightness. For the comparison of denoising and enhancement results, we u the Peak Signal to Noi Ratio (PSNR) of the images[7,8,9]. The PSNR of Fig.19 is smaller than of Fig.16 which illustrated the nois
e was also incread at the same time; The PSNR of Fig.14 is bigger than of Fig.13 which illustrated the noi was truly decread; The PSNR of Fig.20 is bigger than of Fig.19 which illustrated that the noi was truly decread. Our denoising and enhancement algorithm in this paper are ud together from Fig.15~Fig.20 and Fig.21~Fig.23, we can clearly e that the Fig.20 and Fig.23 are better than their original images. They not only have the moderate dynamic range, but also have the well details. Moreover the visual results show us more vivid and limpid images than before for the human eyes.
6.Conclusions
This paper prents color image processing method of denoising and enhancement together, which does retain color impartiality and has a good enhancement results. The HSV color space was lected and divided into H channel, S channel, V channel. The H channel kept invariable, For S channel and V channel, we ud wavelet decomposition and deduced equation in order to denoi and Gaussian lowpass filter to compress dynamic range. The algorithm is effective. In future, we should explore more effective color space from human visual system, at the same time, also determine the type of image noi, do further rearch for different noi. Acknowledgment
This rearch is supported by NSF of China under National Natural Science Foundation of key projects grant number 607360.
References
[1]Ruan qiuqi. Digital Image Processing, Publishing Hou of
Electronics Industry, Beijing 2001..
[2]R.C. Gonzalez, Digital Image Processing, Prentice Hall,
Englewood Cliffs, NJ, 2002.
[3]Yan Jingwen.Digital Image Processing (MATLAB
Version).Nationnal Defen industry Press,Beijing 2007. [4] D.J. Jobson, Z. Rahman, G.A. Woodell, “ A multi-scale
环球留学网retinex for bridging the gap between color images and the
human obrvation of scenes”, IEEE Trans. Image Process. ,1997,Vol.6,No.7,pp 965–976.
[5]M.J. Seow, V.K. Asari, “Homomorphic Processing System
and Ratio Rule for Color Image Enhancement”, Proceedings of the IEEE International Joint Conference on
Neural Networks, Budapest, Hungary, 2004,Vol.7,No.6,pp
2507–2511.
[6]Zia-ur Rahman,Daniel J.Jobson,Glenn A.Woodell.“Retinex
Processing for Automatic Image Enhancement”.Journal of
Electronic Imaging ,2004,Vol.13,No.1,pp 100-110
[7]Han Xiaowei. “the Rearch on Color Image Processing
Key Technology”.Northeastern university,2005.1.1 [8]R.C. Gonzalez, Digital Image Processing Using MATLAB,
Prentice Hall, 2004, 205-206.
[9] E. Kornatowski, and K. Okarma, “Probabilistic Measure of
Color Image Processing Fidelity”, Journal of Electrical Engineering,Vol.59,No.1,2008,pp.29-33.
