Вы находитесь на странице: 1из 4

Image Denoising using Wavelet Thresholding and Model Selection

Shi Zhong Dept. of ECE, Univ. of Texas at Austin szhong@ece.utexas.edu


ABSTRACT This paper describes wavelet thresholding for image denoising under the framework provided by Statistical Learning Theory aka Vapnik-Chervonenkis (VC) theory. Under the framework of VCtheory, wavelet thresholding amounts to ordering of wavelet coefficients according to their relevance to accurate function estimation, followed by discarding insignificant coefficients. Existing wavelet thresholding methods specify an ordering based on the coefficient magnitude, and use threshold(s) derived under gaussian noise assumption and asymptotic settings. In contrast, the proposed approach uses orderings better reflecting statistical properties of natural images, and VC-based thresholding developed for finite sample settings under very general noise assumptions. A tree structure is proposed to order the wavelet coefficients based on its magnitude, scale and spatial location. The choice of a threshold is based on the general VC method for model complexity control. Empirical results show that the proposed method outperforms Donohos level dependent thresholding techniques and the advantages become more significant under finite sample and non-gaussian noise settings. 1. INTRODUCTION

Vladimir Cherkassky Dept. of ECE, Univ. of Minnesota cherkass@ece.umn.edu


The methodology is presented in next section, followed by empirical results. Finally, we present the conclusions. 2. 2.1 VC-theory VC-theory has recently emerged as a general theory for estimating data dependencies from finite samples. It provides a framework for model selection called structural risk minimization (SRM). Under SRM, a set of possible models (each model may consist of one or more basis functions) are ordered according to their complexity. The set, called a structure in SRM, consists of a group of nested subsets Sk such that (1) S1 S 2 S k where each element Sk has finite VC-dimension (the complexity measure in VC-theory) of hk. A structure is designed to provide an ordering of its elements according to their complexity. Model selection can be done by choosing the minimal analytic upper bound (VC-bound) of the prediction risk provided for each element by SRM. For detailed formulation and explanation of VC-bound see [10]. A simplified formula [3] derived for signal estimation (regression) is ln n 1 (2) R p r e d R e m p (1 p p ln p + ) 2n where R e m p is the empirical risk, R pred is the estimated prediction risk, n is the number of signal samples, p( = hk n ) is a complexity parameter. This inequality holds with probability ( 1 1 / n ). A straightforward implementation of SRM is to construct each element Sk in the structure as a linear combination of n k basis functions, in which case the complexity of each element Sk is simply h k = n k + 1 [10]. 2.2 Wavelet thresholding Wavelet thresholding for image denoising involves two steps: 1) taking the wavelet transform of an image (i.e., calculating the wavelet coefficients); 2) discarding (setting to zero) the coefficients with relatively small or insignificant magnitudes. By discarding small coefficients one actually discard wavelet basis functions which have coefficients below a certain threshold. The denoised signal is obtained via inverse wavelet transform of the kept coefficients. One global threshold derived by Donoho [5,6] Clearly, wavelet thresholding can be viewed a special case of signal/data estimation from noisy samples, which can be addressed within the framework of VC-theory. Consider the following structure on a set of all discrete wavelet basis METHODOLOGY

In many applications, image denoising is used to produce good estimates of the original image from noisy observations. The restored image should contain less noise than the observations while still keep sharp transitions (i.e. edges). Wavelet transform, due to its excellent localization property, has rapidly become an indispensable signal and image processing tool for a variety of applications, including compression [7,8] and denoising [1,2,4,5,6,9]. Wavelet thresholding (first proposed by Donoho [4,5,6]) is a signal estimation technique that exploits the capabilities of wavelet transform for signal denoising and has recently received extensive research attentions. It removes noise by killing coefficients that are insignificant relative to some threshold, and turns out to be simple and effective. Wavelet thresholding solution given by Donoho has also proven to be asymptotically optimal in a minimax MSE (mean squared error) sense over a variety of smoothness spaces [2,6]. It should be pointed out, however, all the proofs were conducted under additive gaussian noise assumptions. In this paper, we interpret image denoising as a special case of signal estimation problem and propose a model selection based denoising method under the framework of VC theory, which was developed for estimating data dependencies from finite samples. under gaussian noise assumption is T = 2 log( n ) , where n is the number of samples and the noise standard deviation.

functions: Each element (of the structure) S k has exactly k wavelet basis functions. Note that once k basis functions in S k are specified, minimizing the empirical risk is trivial due to orthogonality of wavelets and amounts to estimation of the wavelet coefficients via discrete wavelet decomposition. In summary, application of SRM to wavelet thresholding for image denoising involve the following steps: 1) Define a structure by appropriate importance ordering of all wavelet basis functions. Each element S k of a structure consists of the first k basis functions. The original wavelet thresholding technique is equivalent to specifying a structure that use only a magnitude ordering of the wavelet coefficients. Obviously, this is not the best way of ordering the coefficients. A better tree structure is presented in this paper. 2) Estimate the prediction risk for each set of wavelet functions formed in the structure. Since each S k is a set of linear models, VC-bound of the prediction risk (2) is easy to compute. 2.3 Level dependent thresholding and importance ordering Level-dependent thresholding has been proposed to improve the performance of wavelet thresholding method. Instead of using a global threshold, level-dependent thresholding uses a group of thresholds, one for each scale level. One popular level-dependent thresholding scheme [4] is to set the threshold as:

each wavelet coefficient using a tree structure. This ordering scheme include following steps: log (max i , j {|WY ( i , j )|}) 1) Set initial threshold t = 2 2 ( denotes the closest smaller integer), final threshold t f (usually 1) and set the initial ordered coefficient list to an empty list; 2) Scan all the coefficients in an order from low scale to high scale. Within each scale, choose (in certain order) those selected (due to space limit, we refer readers to [11] for details on what coefficients are selected) coefficients that are equal to or larger than the threshold t and append them to the list; 3) Set those coefficients selected in step 2) to N/A (not available next iteration) and halve the threshold t; 4) If t t f , then repeat step 2) and step 3); otherwise, append all the rest coefficients to the list in certain scanning order. 3. EMPIRICAL RESULTS

t j ,n =

2 log( n ) 2 (

j J ) / 2

, j = 0, , J

(3)

where n is the total number of signal samples, J is the number of decomposition levels, is the noise standard deviation (to be estimated) and j is the scale level. This scheme uses a larger threshold at finer scale levels. It can be interpreted as: 1) Order the wavelet coefficients with respect to their magnitudes adjusted by scale level as multiplied by 2 j / 2 , where j is the scale level associated with each coefficient. 2) Apply global threshold t n = 2 log( n ) 2 J / 2 . This suggests that the level-dependent thresholding be viewed as a special case of more sophisticated importance ordering in model selection based denoising method. A number of different structures (ordering schemes) can be specified on the same set of basis functions. The choice of a structure can be critical for the success of image denoising. A good ordering should reflect the prior knowledge about the signal/data being estimated. For example, it is not sensible to order a set of polynomial basis functions starting from the highest order term, or order the Fourier basis functions from the highest frequency down (because such orderings contradict the basic assumptions about signal smoothness). Similarly, 2-D image signal estimation with VC approach may require more complicated ordering scheme. Motivated by tree structures used in wavelet-base image compression [7,8], an improved tree-base ordering structure is proposed in this paper. The basic idea is to simultaneously exploit the magnitude, scale and spatial location contribution of

We compared following three denoising methods: 1) WaveThresh: Donohos level dependent thresholding method using (3). The noise standard deviation is calculated using Donohos estimate MAD/0.6745 [4], where MAD is the median of the magnitudes of all the coefficients at the finest decomposition scale. 2) WaveVC: Order the wavelet coefficients using the tree structure proposed in previous section and use VC-bound to choose the optimal number of coefficients (minimizing the bound). 3) Wiener2: Wiener2 in Matlab is a spatial version of Wiener filtering algorithm. Approach 1) and 2) use biorthogonal wavelet filters. The window size, a parameter in Wiener2, is set in our experiments to 3 3. Different image sizes are tested. We mainly compare different methods on two measures: Signal-to-Noise Ratio (SNR) of denoised image and the model complexity of the approximation. SNR is defined as: var( Y ) (4) ( ) SNR = 10 log
10

mse ( Y , Y )

is the denoised where Y is the original clean image and Y image. The model complexity of WaveVC and WaveThresh is just the VC dimension of the model. Wiener2 can be viewed as a local K-mean method doing some local averaging over the noisy image. Its model complexity can be approximated by the VC dimension of K-mean method, which is n/k [3] with n the number of samples and k the size of averaging window. For example, for 512 512 image with 3 3 window size, the model complexity is 512 * 512 / 3 / 3 = 29127.
Due to space limit, we only show results on 8-bit Lenna image in this paper. Fig. 1 and 2 show the comparable denoising results on 512 512 Lenna images corrupted by gaussian white noise ( = 15), using WaveThresh and WaveVC, respectively. Fig. 3 compares the SNR values and the model complexities of the three approaches on 512 512 Lenna images at a variety of different noise levels. The results on 128 128 images and 32 32 images are shown in Fig. 4 and Fig. 5, respectively. In

these results, multiplicative speckle noise is used to show the advantages of our proposed method under non-gaussian settings. We have similar but less dramatic results for gaussian noise settings (which can be found in [11]). Obviously WaveVC performs approximately the same as or better than WaveThresh for 512 512 Lenna images and begins to outperform WaveThresh for smaller (128 128) images. And the relative performance of WaveVC increases further for 32 32 images. The results can be explained as follows: 1) VC theory was designed for finite samples and Donohos threshold was derived under asymptotic assumptions. As the image size gets smaller, the asymptotic assumptions begin to fail. 2) The noise assumption used in Donohos derivation fails when images are not contaminated by additive gaussian noise. In contrast, VC-based approach is more general in this sense. As a global trend, WaveVC tends to use large amount of coefficients for reconstructing the image when the true noise standard deviation is small and use less when is large. And this is true for different image sizes. So when the noise standard deviation is fairly small, meaning the image pretty clean, VC approach tends to keep a large number of coefficients, which makes sense. WaveThresh does not have such clear trends. 4. CONCLUSIONS

Fig. 1 Denoised image by WaveThresh (SNR = 24.99 dB)

Image denoising problem can be cast as a 2-D signal estimation problem. In this paper, VC-based model selection method is integrated with a variation of the wavelet thresholding method and performs well on this problem. An importance ordering structure (the tree structure), which reflects the prior knowledge about the data and the basis functions used, turns out to characterize the importance of noisy wavelet coefficients successfully. However, there may exist better ordering scheme for this wavelet-based denoising problem. Wiener filtering is an optimal linear MSE estimator and Donoho has proven his methods to be minimax optimal under certain assumptions. However, both methods are based on white noise model and true only in asymptotic sense. In contrast, model selection based denosing method is more general and does not need any noise assumption. And compared to Wiener filtering, thresholding uses a sparse structure to approximate the original signal so provides a compressed representation of the original signal (only a small number of coefficients need to be kept). Obviously our method has a lot more potential applications. 5. ACKNOWLEDGEMENT

Fig. 2 Denoised image by WaveVC (SNR = 25.26 dB) [2] A. Chambolle, R. A. DeVore, N-Y Lee and B. J. Lucier, Nonlinear wavelet image processing: variational problems, compression and noise removal through wavelet shrinkage, IEEE Trans. Image Processing, vol. 7, pp. 319-335, 1998 [3] V. Cherkassky and F. Mulier, Learning from Data: Concepts, Theory and Methods, Wiley Interscience, 1998 [4] D. L. Donoho, "Wavelet Thresholding and W.V.D.: A 10minute Tour", Int. Conf. on Wavelets and Applications, Toulouse, France, June 1992 [5] D. L. Donoho and I. M. Johnstone, "Ideal spatial adaptation via wavelet thresholding", Biometrika, vol. 81, pp. 425-455, 1994

This work was supported, in part, by a grant from Minnesota Department of Transportation. 6. REFERENCES

[1] S. G. Chang and M. Vetterli, "Spatial Adaptive Wavelet Thresholding for Image Denoising", Proc of IEEE Int. Conf. on Image Processing, 1997

[6] D. L. Donoho, "De-Noising by Soft-Threshholding", IEEE Trans. Information Theory, vol. 41, No. 3, May 1995 [7] A. Said and W. A. Pearlman, A New Fast and Efficient Image Codec Based on Set Partitioning in Hierarchical Trees, IEEE Trans Circ. and Syst. Video Tech., vol. 6, June 1996 [8] J. M. Shapiro, Embedded Image Coding using Zerotrees of Wavelet coefficients, IEEE Trans. Signal Processing, vol. 41, pp. 3445-3462, Dec. 1993
34 32 30 SNR (dB) 28 26 24 22 20

[9] X. Shao and V. Cherkassky, "Model Selection for Waveletbased Signal Estimation", Proc. IEEE Int. Joint Conf. on Neural Networks, Anchoradge, Alaska, 1998 [10] V. Vapnik, The Nature of Statistical Learning Theory, Springer, 1995 [11] S. Zhong and V. Cherkassky, Image Denoising using Wavelet Thresholding and Statistical Learning Theory, submitted to IEEE Trans. Image Processing, Feb. 2000
4 Model complexity (VC-dimension) x 10
4

. - Wiener2 + - WaveThresh o - WaveVC

3.5 3 2.5 2 1.5 1 0.5

. - Wiener2 + - WaveThresh o - WaveVC

10 20 30 Noise standard deviation

40

50

10 20 30 Noise standard deviation

40

50

Fig. 3 Denoising results for multiplicative speckle noise on 512 by 512 Lenna image
28 Model complexity (VC-dimension) . - Wiener2 + - WaveThresh o - WaveVC 6000 5500 5000 4500 4000 3500 3000 2500 2000 1500 16 0 10 20 30 Noise standard deviation 40 50 1000 0 10 20 30 Noise standard deviation 40 50 . - Wiener2 + - WaveThresh o - WaveVC

26 24 SNR (dB)

22

20 18

Fig. 4 Denoising results for multiplicative speckle noise on 128 by 128 Lenna image
24 Model complexity (VC-dimension) . - Wiener2 + - WaveThresh o - WaveVC 500 450 400 350 300 250 200 150 100 12 0 10 20 30 Noise standard deviation 40 50 50 0 10 20 30 Noise standard deviation 40 50 . - Wiener2 + - WaveThresh o - WaveVC

22 20 SNR (dB)

18

16 14

Fig. 5 Denoising results for multiplicative speckle noise on 32 by 32 Lenna image

Вам также может понравиться