2364

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 18, NO. 10, OCTOBER 2009

Correspondence
Image Denoising Using Trivariate Shrinkage Filter in the Wavelet Domain and Joint Bilateral Filter in the Spatial Domain
Hancheng Yu, Li Zhao, and Haixian Wang nicant ones. Therefore, one straightforward way of distinguishing information from noise in the wavelet domain is to threshold the wavelet coefcients. Of the various thresholding strategies, soft-thresholding is the most popular and has been theoretically justied by Donoho and Johnstone [1]. Accurate and efcient wavelet coefcients modeling, whether done explicitly or implicitly, is a critical component of image denoising. In [2], the wavelet coefcients were modeled with Gaussian density, and locally adaptive Wiener estimates were done for coefcient variances. To obtain better estimates of the signal variances, a doubly local Wiener ltering algorithm with directional windows was presented in [3]. Non-Gaussian bivariate distributions were introduced in [4] and [5] by taking into account the dependency among the wavelet coefcients, and then a nonlinear bivariate shrinkage function (BiShink) using the maximum a posteriori estimator (MAP) was derived. An image denoising algorithm based on Gaussian scale mixture model and Bayes least squares estimator (BLS-GSM) was presented in [6]. By exploiting the Steins unbiased risk estimator principle [7], Luisier et al. introduced a new interscale SURE-based approach to orthonormal wavelet image denoising (OWT-SURELET) [8]. Contributions from new various redundant wavelet techniques, such as double-density DWT [9], steerable pyramid [10], and dual-tree complex wavelet [11], [12], helped spawn many new developments. Although wavelet-based methods are efcient in image denoising, they are prone to producing salient artifacts such as low-frequency noise and edge ringing which relate to the structure of the underlying wavelet. In contrast, most algorithms based on the spatial domain can preserve more of the important image features and reduce low-frequency noise and edge ringing. In [13], Hirakawa et al. developed a model relating the noisy image to an ideal image in the total least squares (TLS) sense, taking into account the stochastic nature of the noise and allowing small perturbations in their system. In [14], Takeda et al. introduced kernel regression (KR) as a general framework for studying several efcient denoising algorithms. In [15], Awate et al. described a novel unsupervised, information-theoretic, adaptive lter (UINTA) that improved the predictability of pixel intensities from their neighborhoods by decreasing their joint entropy. In [16], Kervrann et al. presented spatial adaptive and patch-based approach for image denoising. Although the high quality of the output denoising image demonstrates the effectiveness of these spatial algorithms, they are usually too computationally demanding. Recently, Elad et al. developed an image denoising algorithm based on sparse and redundant representations of over learned dictionaries, where training and denoising used the K-SVD algorithm [17], [18]. Dabov et al. grouped 2-D image fragments into 3-D data arrays, and applied a collaborative ltering procedure [19], which consisted of 3-D transformation, shrinkage of the transformed spectrum, and inverse 3-D transformation. To surmount the problems of artifacts and the computational complexity, in this paper, we develop an efcient algorithm that combines the wavelet-based trivariate shrinkage lter and the spatial-based joint bilateral lter for image denoising. In the wavelet domain, the coefcients are modeled as trivariate Gaussian distribution based on the fact that the statistical dependencies among the coefcients especially among signicant coefcients are important for denoising, and then a trivariate shrinkage lter is derived by using the MAP estimator. Considering that the wavelet denoising image preserves most of the important image features, we use it as a reference image in the spatial

AbstractThis correspondence proposes an efcient algorithm for removing Gaussian noise from corrupted image by incorporating a wavelet-based trivariate shrinkage lter with a spatial-based joint bilateral lter. In the wavelet domain, the wavelet coefcients are modeled as trivariate Gaussian distribution, taking into account the statistical dependencies among intrascale wavelet coefcients, and then a trivariate shrinkage lter is derived by using the maximum a posteriori (MAP) estimator. Although wavelet-based methods are efcient in image denoising, they are prone to producing salient artifacts such as low-frequency noise and edge ringing which relate to the structure of the underlying wavelet. On the other hand, most spatial-based algorithms output much higher quality denoising image with less artifacts. However, they are usually too computationally demanding. In order to reduce the computational cost, we develop an efcient joint bilateral lter by using the wavelet denoising result rather than directly processing the noisy image in the spatial domain. This lter could suppress the noise while preserve image details with small computational cost. Extension to color image denoising is also presented. We compare our denoising algorithm with other denoising techniques in terms of PSNR and visual quality. The experimental results indicate that our algorithm is competitive with other denoising techniques. Index TermsImage denoising, joint bilateral lter, spatial domain, trivariate shrinkage lter, wavelet domain.

I. INTRODUCTION During acquisition and transmission, images are often corrupted by additive noise, which usually can be modeled as Gaussian most of the time. The main aim of an image denoising algorithm is thereby to reduce the noise level, while preserve the image features. Discrete wavelet transform (DWT) has received an increasing attention for image denoising in the eld of image processing. Up to now, a vast literature has emerged on image denoising based on DWT. The success of DWT is due to the fact that the wavelet-transformed images tend to be sparse (i.e., many coefcients are close to zero). In the wavelet domain, the noise is uniformly spread throughout the coefcients, while most of the image information is concentrated in few sigManuscript received October 16, 2008; revised June 01, 2009. First published July 06, 2009; current version published September 10, 2009. This work was supported in part by the Specialized Research Fund for the Doctoral Program of Higher Education of China under Grant 20070286030 and in part by the National Natural Science Foundation of China under Grants 60872073 and 60803059. The associate editor coordinating the review of this manuscript and approving it for publication was Dr. Michael Elad. H. Yu is with the School of Information Science and Engineering, Southeast University, Nanjing 210096, China, and also with the School of Information Science and Technology, Nanjing University of Aeronautics and Astronautics, Nanjing 210016, China (e-mail: yuhc@nuaa.edu.cn). L. Zhao is with the School of Information Science and Engineering, Southeast University, Nanjing 210096, China (e-mail: zhaoli@seu.edu.cn). H. Wang is with the Key Laboratory of Child Development and Learning Science of Ministry of Education, Southeast University, Nanjing 210096, China (e-mail: hxwang@seu.edu.cn). Color versions of one or more of the gures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identier 10.1109/TIP.2009.2026685

1057-7149/$26.00 2009 IEEE

Authorized licensed use limited to: Mepco Schlenk Engineering College. Downloaded on September 11, 2009 at 04:55 from IEEE Xplore. Restrictions apply.

YU et al.: ZHANG AND YE: CONTENT BASED IMAGE RETRIEVAL USING UNCLEAN POSITIVE EXAMPLES

2365

where Cn denes the covariance structure of n

Cn =
Fig. 1. Neighborhood of a central wavelet coefcient.

2 n

0 0

2 n

2 n

0 0

(7)

2 and n denotes the variance of noise. Let us continue on developing the MAP estimator given in (3), which is equivalent to

denoising. So, an efcient joint bilateral lter is developed. Compared with other spatial algorithms, this proposed spatial lter could suppress the noise while preserve image details with small computational cost. Furthermore, these two lters can be easily extended to capture interchannel correlations of multichannel image. So extension to color image denoising is also presented. II. TRIVARIATE SHRINKAGE FILTER AND JOINT BILATERAL FILTER A. Trivariate Shrinkage Filter in Wavelet Domain In this paper, the denoising of an image corrupted by additive independent white Gaussian noise will be considered. In the wavelet domain, there are strong dependencies between a coefcient and its four closest neighbors, especially between signicant coefcients that contain the edge information of image. For each central pixel x of the image (that is not a border pixel), we use a 3 2 3 neighborhood window as illustrated in Fig. 1. The four closest neighbors are fxE; xW ; xN ; xS g corresponding to two directions {East-West, NorthSouth}. If we use an orthogonal wavelet transform along the North-South direction, the problem can be formulated as

^ x(y) = arg max [log(pn (y 0 x)) + log(px (x))] : x

If px (x) is assumed to be strictly convex and differentiable, we get

(8)

C01 (y 0 x) 0 C01 x = 0: n x
Then the MAP estimator can be written as
0 0 ^ x = (Cn 1 + C01 )01 Cn 1y: x

(9)

(10)

Or we can estimate the true wavelet coefcient w from the North-South direction (trivariate shrinkage lter)

xNS ^
with

= hN yN +hhy + hS yS NS

(11)

y =x+n
where

(1)

are noise samples. Given the corrupted observation estimator for x is

x = (xN ; x; xS ) denote true wavelet coefcients, y = (yN ; y; yS ) denote the noisy observations of x, and n = (nN ; n; nS )

2 2 2 2 2 hN = N N (S + n )n 0 NS S N S n 2 + 2 )2 (2 + 2 ) h = (N n n S 2 2 +(2N NS S 0 2 )N 2 S NS 2 2 2 (2 + 2 ) 0 2 (2 + 2 )2 2 0N N S n S N n S 2 2 2 2 2 hS = S (N + n )S n 0 N NS N S n 2 + 2 )(2 + 2 )2 0 2 2 2 2 : hNS = h + (N n S n n NS N S n

(12)

y, the MAP
(2)

Similarly, we can estimate the true wavelet coefcient x from the EastWest direction

^ x(y) = arg max pxjy (xjy): x

Using Bayesian rule, one gets

h y + hy + hE yE xWE = W W ^ : hWE

(13)

^ x(y) = arg max[pn (y 0 x) 1 px (x)]: x

(3)

Therefore, the equation allows us to write this estimation in terms of the pdf of the true wavelet coefcient (px ) and the noise (pn ). We suppose trivariate Gaussian distribution for the coefcient and its neighbors in the North-South direction as

px (x) =

1 exp 0 1 xT Cx x 2 2 jCx j

(4)

When an image is corrupted by additive Gaussian noise, for most small wavelet coefcients, their local correlations are weak. Notice that in this case, the trivariate shrinkage lter is similar with the local Wiener lter in [2]. On the other hand, for few signicant coefcients that contain the most important information of the image, there are still strong dependencies between these coefcients and their neighbors. The trivariate shrinkage lter really takes effect in the denoising of these coefcients. Moreover, these coefcients can be estimated more accurately by combining the lter results of two directions. Specically, the direction that has stronger dependency between its coefcients and the central pixel has greater weights

2 N N N  NS N S 2 S S (5) Cx = N N  2 NS N S S S S 2 2 where N ,  2 , and S denote the local signal variance of xN , x and xS , respectively, and N , S and NS denote the local correlation coefcients which dene the dependence between xN and x, x and xS , xN and xS . We also let the noise pdf be

The matrix Cx denes the covariance structure of x

x= ^

max(jN j ; jS j) 1 xNS + max(jW j ; jE j) 1 xWE : ^ ^ max(jN j ; jS j) + max(jW j ; jE j)

(14)

pn (n) =

1 exp 0 1 nT Cnn 2 2 jCn j

(6)

Compared with the BLS-GSM [6] that uses Bayesian Least Squares estimator, our trivariate shrinkage lter estimates the true wavelet coefcient by using its closest neighbors and the local correlation coefcients directly. As we know, the trivariate shrinkage lter is not a very exact model especially when using redundant transformation, the covariance matrix of the noise coefcients is not a strict diagonal matrix as in the orthogonal transform case. However, the trivariate shrinkage lter is a convenient tool to model the relationship between the wavelet coefcients and it is computationally efcient.

Authorized licensed use limited to: Mepco Schlenk Engineering College. Downloaded on September 11, 2009 at 04:55 from IEEE Xplore. Restrictions apply.

2366

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 18, NO. 10, OCTOBER 2009

B. Joint Bilateral Filter in Spatial Domain As mentioned above, wavelet-based methods produce salient artifacts which relate to the structure of the underlying wavelet, and most spatial-based methods are usually too computationally demanding. The bilateral lter, as described in [20], applied a simple nonlinear lter to remove noise while retaining the sharpness of edges. Given a noisy ^ image u, at a pixel location (i; j ), the restored result si;j can be directly calculated by a weighted average of the intensities in its noisy neighborhood uk;l as follows:
si;j = ^
(k;l)

penalty function in [22]. In this paper, for simplicity, based on the reference image, the "i;j can be directly estimated as
"i;j = ^
(k;l)

(i;j )

hDk;l

hPk;l hDk;l

(uk;l
hPk;l

sk;l )

ref

(k;l)

(20)

(i;j )

and then we add this compensation term in the joint bilateral lter to suppress the spots
si;j = ^
(k;l)

(i;j )

hDk;l

hPk;l
1

uk;l

(i;j )

hDk;l

hPk;l
1

uk;l
0

(k;l)

(i;j )

hDk;l

hPk;l

(15)

(k;l)

(i;j )

hDk;l

hPk;l

 "i;j ^

[0; 1]:

(21)

where
N (i; j ) denote the set of points in the (2N + 1) 2 (2N + 1) window centered at (i; j ). The bilateral lter combines a classic lowpass function hD with an edge stopping function hP . They attenuate to smooth in regions of similar intensity while keeping edges intact, by heavily weighting those pixels that are both near the central pixel spatially and similar to the central pixel radiometrically. The main problem of the classic bilateral lter in image denoising is that the edge stopping function hP could not be estimated accurately based on the noisy image. Since the ash image contains a much better estimate of the true high-frequency information than the ambient image, Petschnigg et al. [21] presented a joint bilateral lter to compute the edge-stopping function using the ash image in the such ash/noash image denoising. Also, considering that the wavelet-based denoising image preserves most of the important image features, we can use it as a reference image, like the ash image in [21]. In this way, the edge-stopping function could be estimated more accurately
hPk;l = exp
0

In the joint bilateral lter, the parameters  and P vary with respect to the reference image quality and the noise level. Ideally, if noise level is so low that the reference image is very close to the original image,  = 1 and P = 1=1 could be good choice. However, in fact, the noise level is not xed and the reference image is generated from the noisy image by using the wavelet trivariate shrinkage lter. Experimental results show that good range for the two parameters are  2 [0:6; 0:8] and P 2 [3; 12]. In this paper, for simplicity, we set  = 0:7 and let
P = 3 +
p

n :

(22)

C. Denoising Algorithm As shown in Fig. 2, our denoising algorithm, named TRISF-JBF, is composed of two phases based on the wavelet trivariate shrinkage lter and the spatial joint bilateral lter, respectively. In the wavelet phase, in order to achieve better denoising results, the trivariate shrinkage lter is used twice. It is noticeable that the estimates of the local correlation coefcient  and the signal variance  2 play an important role in our trivariate shrinkage lter. In the rst TRISF stage, taking W for example, it can be estimated as
(k;l)

sk;l

ref

si;j

ref

P

(16)

where sref is the wavelet denoising reference image. The parameter P determines the threshold to penalize the neighboring pixels which vary greatly in intensity from the central pixel according to the reference image. The hD is determined by the distance between the neighbor and the central pixel, where a Gaussian function is usually used
hDk;l = exp
0

Wi;j = ^

(i;j )

yk;l

yk;l01
1

(k;l)

(i;j )

2 yk;l

2 yk;l01

(23)

(k

i) + (l
2 2D

j)

(17)

and the signal variance is estimated as

i;j = ^
2

The parameter D characterizes the spatial behavior of the bilateral lter, but with the changes of the noise level and the size of
N , D is not easy to select. So we dene a simpler spatial weight function
hDk;l =

1
M
(k;l)

yk;l
(i;j )

n ^
+

(24)

1 1 + jk
0

ij + jl

jj

(18)

If we use the joint bilateral lter in (15) directly, there are clear noise spots in the output denoising image, especially when the noise level is high. Let
hDk;l hPk;l hDk;l

"i;j =

(k;l)

(i;j )

(uk;l
hPk;l

sk;l ) :

where (g )+ = max(0; g ),
N 2N (i; j ) denote the set of points in the (2NV + 1) 2 (2NH + 1) directional window centered at (i; j ) and M denotes the size of the window. Taking into account that these parameters are estimated from noisy coefcients, the estimated results are not guaranteed to be satisfactory. In [3], a doubly estimate was presented to solve this problem: based on the rstly denoising image which contains much less noise and is taken as the pilot image, the local correlation coefcient and the signal variance of the wavelet coefcient can be estimated more accurately in the second TRISF stage
+ Wi;j ^

(k;l)

(19)
=

(k;l)

(i;j )

(i;j )

xk;l ^

xk;l01 ^
1

If "i;j = 0, the joint bilateral lter performs well in suppressing noise. The equation, however, is not completely established, which is the main reason of noise spots. To overcome this dilemma, Elad presented a

(k;l)

(i;j )

x2 ^k;l
2

x2 01 ^k;l

(25)

i;j = ^

2+

1
M
(k;l)

xk;l : ^
(i;j )

(26)

Authorized licensed use limited to: Mepco Schlenk Engineering College. Downloaded on September 11, 2009 at 04:55 from IEEE Xplore. Restrictions apply.

YU et al.: ZHANG AND YE: CONTENT BASED IMAGE RETRIEVAL USING UNCLEAN POSITIVE EXAMPLES

2367

Fig. 2. Block-diagram of TRISF-JBF algorithm.

In the spatial phase, the joint bilateral lter uses the previously generated wavelet restoration results as a reference image, and then we get the nal recovery image. In the wavelet TRISF phase, another problem needed to be considered is how to select window for local covariance and estimation. In the rst TRISF stage, the energy clustering in the three oriented subbands are distributed approximately along the horizontal, vertical, and diagonal directions respectively for orthogonal wavelet transform, the signal variances and the correlation coefcient are estimated from the noisy coefcients. So we use relatively large directional window which is determined primarily by three factors: the type of subband, the size of subband and the noise level

III. EXTENSION TO COLOR-IMAGE DENOISING A naive approach to denoising multichannel images is simply to apply a grayscale denoising method for each channel separately. However, this solution is far from being optimal, due to the presence of potentially strong common information between the various channels. One conceivable strategy is to extend grayscale denoising method by efciently exploiting interchannel correlations, such as C-KVSD [23], C-BM3D [15], C-SURELET [24], and ProbShrink-MB [25]. One advantage of our trivariate shrinkage lter and joint bilateral lter is that they can be easily extended to capture some correlation between the channels. So they are also efcient in color image denoising. We consider a natural RGB image with additive Gaussian noise in each channel. Taking the red channel in RGB space for example, in the wavelet domain, we extend the trivariate shrinkage lter (14) from intrascale two directions to three directions, by adding an interscale direction that considers the dependencies between the wavelet coefcients of different channels. See (28), shown at the bottom of the page, where the ^GB denotes the estimator from RGB direction. In the spatial space, we also extend the edge-stopping function in (16) from intracannel to intercannel. See equation (29), shown at the bottom of the page, where ref R , ref G and ref B are three wavelet denoising reference images in RGB color space, RG and RB denote the local correlation coefcients between R and G, R and B cannel, p n. respectively, and let P = 0 5 + In our color denoising algorithm, to achieve better results, we use the rst wavelet trivariate shrinkage lter in luminance-chrominance (YUV) color space, and then we use the second wavelet trivariate shrinkage lter and spatial joint bilateral lter in RGB color space.

NV = 1 + round NH = 1 + round L L   

256 L ( + 256

L ( +

) )
(27)

where V and H denote the vertical and horizontal sizes of the subband respectively and is the orientation factor. For horizontal subband, let H = 2 and V = 0, for vertical subband, let H = 0 and V = 2, and for diagonal subband, let H = 1 and V = 1. For the dual-tree complex wavelet subbands that oriented at other different angles, we still can classify each subband to these three directions roughly. In the second wavelet transform, the signal variances are estimated from the pilot image with much less noise. So, small size
N (3 2 3) could be a good choice. In the spatial JBF phase, if the noise power level ( n ) is higher than 15, we use large size
N (11 2 11 or 13 2 13). Otherwise we use small size
N (5 2 5 or 7 2 7).

; jS j) ^ j 1 ^ + max(j j ;  ^ xR = max(jN jmax(j1NxNS+j)max(jW j;WjE j) ExWEmax(jG j;GjBjj)B j) 1 xGB ^ j ; j S + max(j ; j j) +

(28)

ref R ref R ref G 0 ref G ref B ref B i;j hPk;l = exp 0 sk;l 0 si;j + RGPsk;l+ RGs+ RB+ RB sk;l 0 si;j )  (1

(29)

Authorized licensed use limited to: Mepco Schlenk Engineering College. Downloaded on September 11, 2009 at 04:55 from IEEE Xplore. Restrictions apply.

2368

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 18, NO. 10, OCTOBER 2009

TABLE I COMPARISON OF SOME GRAYSCALE IMAGE DENOISING METHODS IN PSNR

TABLE II COMPARISON OF SOME COLOR IMAGE DENOISING METHODS IN PSNR

Fig. 3. Denoised results in restoring corrupted image Lena ( = 30). (a) The original image. (b) Noisy input image. (c) Output from BLS-GSM. (d) Output from K-SVD. (e) Output from BM3D. (f) Output from TRISF-JBF.

IV. EXPERIMENTAL RESULTS In this section, the proposed algorithm is evaluated and compared with many other existing techniques. Extensive experiments are conducted on four standard grayscale test images and four standard color test images with distinctly different features, corrupted by simulated additive Gaussian white noise at different power levels. The parameters of each method have been set according to the values given by their respective authors in the corresponding referred papers. A. PSNR Comparisons We rst compare our TRISF-JBF with four popular grayscale image denoising techniques: Portillas BLS-GSM [6] with an 8-orientations full steerable pyramid wavelet, Hirakawas total least squares (TLS) algorithm [13], Elads K-SVD [18] and Dabovs BM3D [19]. For color image denoising, we compare our C-TRISF-JBF with three existing denoising techniques: Luisiers C-SURELET [24], Pi uricas Probz Shrink-MB [25], and C-BM3D. The PSNR values of the KR method are missing because we dont know how to optimize the parameters. In order to improve the denoising performance of our TRISF-JBF algorithm, in the TRISF phase, we use the dual-tree complex wavelet with near-symmetric 13,19 tap lters at level 1 and Q-Shift 14,14 tap

lters at higher levels [12]. From Tables I and II, it is noted that our results are competitive with the existing best denoising techniques, especially for the color image denoising. B. Visual Quality In Figs. 3 and 4, the output images are processed with the proposed and other state- of-the-art image denoising algorithms. The fragments of the grayscale Lena and the color Baboon are corrupted by additive Gaussian white noise. Although there is no consensual objective way to judge the visual quality of a denoised image, two important criteria are widely used: the conservation of image edges and the visibility of processed artifacts. The proposed algorithm (TRISF-JBF) generally outputs smoother surfaces in homogeneous regions, and preserves sharper edges in detailed regions. While all denoising algorithms in this section suffer from artifacts under severe noise, by contrast, our TRISF-JBF and BM3D exhibit relatively few numbers of artifacts. Compared with BM3D, TRISF-JBF can preserve more of the image details. C. Computational Time It is also important to evaluate the various denoising algorithms from a practical point of view: the computational time. In evaluating the

Authorized licensed use limited to: Mepco Schlenk Engineering College. Downloaded on September 11, 2009 at 04:55 from IEEE Xplore. Restrictions apply.

YU et al.: ZHANG AND YE: CONTENT BASED IMAGE RETRIEVAL USING UNCLEAN POSITIVE EXAMPLES

2369

Fig. 4. Denoised results in restoring corrupted image Baboon ( (d) Output from C-BM3D. (e) Output from C-TRISF-JBF.

= 25). (a) The original image. (b) Noisy input image. (c) Output from ProbShrink-MB.

computational complexity of our algorithm, we consider wavelet and spatial stages. The number of operations per pixel is approximately

2Cw + 12klMW + 5MS

(30)

where Cw denotes the number of arithmetic operations required for wavelet transform, k is the wavelet factor (k = 4 in dual-tree complex wavelet, and k = 1 in orthogonal wavelet), l is the direction factor in wavelet phase (l = 3 for color image, and l = 2 for grayscale image), MW is the size of the window for the double signal variance and correlation coefcient estimation in wavelet phase, and MS is the size of the window for the joint bilateral lter in spatial phase. They all depend strongly on the noise level. The approximate execution time (for denoising a 512 2 512 grayscale image by using Matlab on a Windows XP-based personal computer with an Intel Pentium IV 2.40-GHz CPU and 512-MB RAM) is about 50 s. V. CONCLUSION AND FUTURE WORK In this paper, an efcient algorithm is proposed for removing noise from corrupted image by incorporating a wavelet-based trivariate shrinkage lter with a spatial-based joint bilateral lter. To demonstrate the superior performance of the proposed method, extensive experiments have been conducted on several standard test images to compare our method with many other well known techniques. It is known that the white Gaussian noise assumption is not always accurate for real noisy images. The neighborhood-based bilateral lter is also effective for removing non-Gaussian noise. So, we are currently considering extending this work to real image denoising. ACKNOWLEDGMENT The authors would like to thank the anonymous referees for their recommendations which improve the paper greatly.

REFERENCES
[1] D. L. Donoho and I. M. Johnstone, Adapting to unknown smoothness via wavelet shrinkage, J. Amer. Statist. Assoc., vol. 90, no. 432, pp. 12001224, Dec. 1995. [2] M. K. Mihak, I. Kozinsev, K. Ramchandran, and P. Moulin, Lowcomplexity image denoising based on statistical modeling of wavelet coefcients, IEEE Signal Process. Lett, vol. 7, no. 6, pp. 300303, Jun. 1999. [3] P. L. Shui, Image denoising algorithm via doubly local Wiener ltering with directional windows in wavelet domain, IEEE Signal Process. Lett, vol. 12, no. 10, pp. 681684, Oct. 2005. [4] L. Sendur and I. W. Selesnick, Bivariate shrinkage functions for wavelet-based denoising exploiting interscale dependency, IEEE Trans. Signal Process., vol. 50, no. 11, pp. 27442756, Nov. 2002. [5] L. Sendur and I. W. Selesnick, Bivariate shrinkage with local variance estimation, IEEE Signal Process. Lett, vol. 9, no. 12, pp. 438441, Dec. 2002.

[6] J. Portilla, V. Strela, M. J. Wainwright, and E. P. Simoncelli, Image denoising using scale mixtures of gaussians in the wavelet domain, IEEE Trans. Image Process., vol. 12, no. 11, pp. 13381351, Nov. 2003. [7] C. Stein, Estimation of the mean of a multivariate normal distribution, Ann. Statist., vol. 9, pp. 11351151, 1981. [8] F. Luisier, T. Blu, and M. Unser, A new SURE approach to image denoising: Interscale ortonormal wavelet thresholding, IEEE Trans. Image Process., vol. 16, no. 3, pp. 593606, Mar. 2007. [9] I. W. Selesnick, The double density DWT, in Wavelets in Signal and Image Analysis: From Theory to Practice, A. Petrosian and F. G. Meyer, Eds. Boston, MA: Kluwer, 2001. [10] W. T. Freeman and E. H. Adelson, The design and use of steerable lters, IEEE Pattern Anal. Mach. Intell., vol. 13, no. 10, pp. 891906, Sep. 1991. [11] N. G. Kingsbury, Image processing with complex wavelets, Phil. Trans. Roy. Soc. Lond. A, pp. 25432560, Sep. 1999. [12] N. G. Kingsbury, Complex wavelets for shift invariant analysis and ltering of signals, Appl. Compt. Harmon. Anal., pp. 234253, May 2001. [13] K. Hirakawa and T. W. Parks, Image denoising using total least-squares, IEEE Trans. Image Process., vol. 15, no. 9, pp. 27302742, Sep. 2006. [14] H. Takeda, S. Farsiu, and P. Milanfar, Kernel regression for image processing and reconstruction, IEEE Trans. Image Process., vol. 16, no. 2, pp. 349366, Feb. 2007. [15] S. P. Awate and R. T. Whitaker, Unsupervised, information-theoretic, adaptive image ltering for image restoration, IEEE Trans. Pattern Anal. Mach. Intell., vol. 28, no. 3, pp. 364376, Mar. 2006. [16] C. Kervrann and J. Boulanger, Optimal spatial adaptation for patchbased image denoising, IEEE Trans. Image Process., vol. 15, no. 10, pp. 28662878, Oct. 2006. [17] M. Elad and M. Aharon, Image denoising via learned dictionaries and sparse representation, IEEE Comput. Vis. Pattern Recognit., Jun. 2006. [18] M. Elad and M. Aharon, Image denoising via sparse and redundant representations over learned dictionaries, IEEE Trans. Image Process., vol. 15, no. 12, pp. 37363745, Dec. 2006. [19] K. Dabov, A. Foi, V. Katkovnik, and K. Egiazarian, Image denoising by sparse 3D transform-domain collaborative ltering, IEEE Trans. Image Process., vol. 16, no. 8, pp. 20802095, Aug. 2007. [20] C. Tomasi and R. Manduchi, Bilateral ltering for gray and color images, in Proc. Int. Conf. Computer Vision, 1998, pp. 839846. [21] G. Petschnigg et al., Digital photography with ash and no-ash image pairs, in Proc. SIGGRAPH, pp. 664672. [22] M. Elad, On the bilateral lter and ways to improve it, IEEE Trans. Image Process., vol. 11, no. 10, pp. 11411151, Oct. 2002. [23] J. Mairal, M. Elad, and G. Sapiro, Sparse representation for color image restoration, IEEE Trans. Image Process., vol. 17, no. 1, pp. 5369, Jan. 2008. [24] F. Luisier and T. Blu, SURE-LET multichannel image denoising: Interscale orthonormal wavelet thresholding, IEEE Trans. Image Process., vol. 17, no. 4, pp. 482492, Apr. 2008. [25] A. Pi urica and W. Philips, Estimating the probability of the presence z of a signal of interest in multiresolution single- and multiband image denoising, IEEE Trans. Image Process., vol. 15, no. 3, pp. 645665, Mar. 2006.

Authorized licensed use limited to: Mepco Schlenk Engineering College. Downloaded on September 11, 2009 at 04:55 from IEEE Xplore. Restrictions apply.