You are on page 1of 6

2000 Conference on Information Sciences and Systems, Princeton University, March 15-17, 2000

Wavelet-Based Statistical Image Processing Using Hidden Markov Tree Model


Department of Electrical and Computer Engineering University of Delaware, Newark, DE 19716 E-mail: glfan,xxia@ece.udel.edu

Guoliang Fan and Xiang-Gen Xia1

Abstract | Wavelet-domain Hidden Markov Tree HMT model was recently proposed and applied to image processing, e.g., image denoising and segmentation. In this paper, we further study the applications of the HMT model to image processing. For image denoising, we develop a new initialization scheme for more e cient HMT model training and the Classied HMT model for more accurate image modeling. The image denoising performance can be improved at reasonable computational cost. For texture segmentation, we develop three new context structures which apply to homogeneous regions and texture boundaries. Then, we propose a Hybrid Contextual Labeling Tree by integrating the three context structures together to accomplish the Bayesian interscale fusion, so that we can improve the accuracies of both texture classi cation and boundary localization.
I. Introduction In 1 , a new statistical model in the wavelet-domain, Hidden Markov Tree HMT model, was proposed by Crouse et al. which can model the joint statistics of wavelet coe cients, i.e., interscale dependencies of wavelet coe cients. Romberg et al. and Choi et al. have applied the HMT model to image denoising and segmentation in 2, 3, 4 , respectively. In this work, we want to improve the performance the HMT model for image processing, i.e., denoising and segmentation. In 2 , the HMT model training using the Expectation Maximization EM algorithm was considered as an obstacle to the practical implementation of image denoising using the HMT model, and a reduced-parameter HMT model, called universal HMT model uHMT, was developed which does not need model training and still allows e ective image denoising. In this paper, we can facilitate the HMT model training by developing a new initialization scheme which allows more e cient and accurate computation of the HMT model parameters. In order to obtain more accurate image modeling, We propose the Classi ed HMT CHMT model which has spatial variations and is able to characterize more intrascale dependencies of wavelet coe cients. Comparing to 2 , the performance of image denoising can be improved by 0.4-0.9dB for two real images at reasonable computation cost. In 3, 4 , a new multiscale image segmentation algorithm, called HMTseg, was proposed which uses the direct computation of the HMT model likelihood and the context-based interscale fusion. The interscale fusion is critical to the segmentation performance and speci ed by a context structure.
1 This work was partially supported by the 1998 O ce of Naval Research ONR Young Investigator Program YIP under Grant N00014-98-1-0644 and the Air Force O ce of Scienti c Research AFOSR under Grant No. F49620-98-1-0352.

In order to achieve higher accuracies of both texture classi cation and boundary localization, we develop three new context structures which apply to homogeneous regions and texture boundaries. Then, we develop a Hybrid Contextual Labeling Tree HCLT where the three context structures are integrated together and the EM training is performed three times based the three contexts, respectively and sequentially. Via the HCLT, the segmentation result can be re ned stepby-step. The new algorithm outperforms the original HMTseg algorithm by improving the accuracies of both texture classication and boundary localization. This paper is organized as follows. Some background of the wavelet-domain HMT model and its applications are brie y reviewed in Section II. We discuss our improvements on image denoising and texture segmentation in Section III and Section IV. Simulation results are shown in Section V where we demonstrate the advantages of our improved algorithms and compare with some state-of-the-art methods.
II. Background In the following, we brie y present the background of our work. We refer the reader to 1, 2, 3, 4 for more details.

A Wavelet-Domain HMT Model

The discrete wavelet transform DWT represents an image with both the spatial and frequency characterizations as shown in Fig. 1a. From Fig. 1a, we can see the quad-tree structure of the wavelet coe cient subtree in each subband B , B = LH; HL; HH , as shown in Fig. 1b, where we can nd a nesting structure of the wavelet subtrees. Based on the quadtree structure shown in Fig. 1, given a J -scale N  N wavelet B coe cient set, w, wj;i;k means the i; kth coe cient in the j th scale of subband B , i; k = 0; :::; Nj , 1 and Nj = N=2j . B B The parent of wj;i;k is wj+1;bk=2c;bi=2c where the operation bxc takes the integer part of x. In practice, the probability density function pdf, fW w, of the wavelet coe cients w can be described by a peak centered at w = 0 and heavy-tailed non-Gaussian density, where W stands for the random variable of w. It was shown in 5 that Gaussian mixture models can well approximate this non-Gaussian density. Therefore, we associate each wavelet coe cient w with a set of discrete hidden states, S = 0; 1; :::; M , 1, which have probability mass functions pmf, pS m. Thus, we can parameterize the Gaussian mix2 ture model by  = fpS m; m ; m jm = 0; 1; :::; M , 1g, and the overall pdf of w is determined by

fW w =

M ,1 m=0

X pS mfW jS wjS = m:

1

LL3 LH3

HL3

HL2
HH3

HL1
LH2 HH2
j=2

j=3

LH1

HH1
j=1

a

b

c

d

Figure 1: a Tiling of the spatial-frequency plane by the 2-D orthogonal DWT of 3-scale with four subbands LL, HL, LH and HH . The quad-tree structure in subband HH is also shown by the link of the dashed lines. b The wavelet subtree has a quad-tree structure rooted in one wavelet coe cient w in a subband. c 2-D HMT model where the white node represents the hidden mixture state variable S . The black node denotes a continuous wavelet coe cient W which is a random variable. d The pyramid structure of the dyadic image blocks.
In 1 , a new statistical model, called Hidden Markov Tree HMT model, was proposed to characterize the joint pmf of wavelet coe cients, as shown in Fig. 1c. 1 The HMT model is a multidimensional Gaussian mixture model which applies Markovian dependencies between the hidden states of two neighboring scales and parameterized by, i pSJ m: the pmf of the root node SJ , ii m;n = pSj jSj+1 Sj;i;k = mmjSj+1;bk=2c;bi=2c = n: the j;j +1 transition probability; 2 iii j;m and j;m : Gaussian mixture model parameters. with j = 1; 2; :::; J and m; n = 0; 1; :::; M , 1. These parameters can be grouped into a vector as 2 = fpSJ m; m;n ; j;m ; j;m jj = 1; :::; J ; n; m = 0; :::; M ,1g j;j +1 2 The determination of can be approached by the Expectation Maximization EM algorithm in 1 .

C HMTseg Algorithm

B Wavelet-Based Signal Denoising using the HMT Model

In 1 , an empirical" Bayesian approach was proposed for 1D signal denoising and characterizes the wavelet coe cients by two component Gaussian mixture, i.e., M = 2 with j;1 = j;2 = 0. We can obtain the conditional mean estimate of the noise-free coe cient by the chain rule as

E Yj;i jw; =

M ,1 m=0

X pSj;i = mjw;  

2 2 j;m , 2 j;m

wj;i ; 3

2 where j;m is the Gaussian mixture variance of the noise-free 2 coe cients yj;i given state m, j;m the mixture variance of the 2 noisy coe cients wj;i , the variance of the additive white Gaussian noise AWGN, and pSj;ijw;  can be computed during the EM algorithm. Then the last denoised signal is achieved by the inverse DWT of these estimates of the wavelet coe cients. The framework above was applied to image denoising in 2 where a uHMT model characterized by a set of xed parameters was proposed to avoid the computational expensive model training. 1 Since the HMT model is used in each subband respectively, we will omit the notation B unless the complete index is required. In the following, Wj;i;k and Sj;i;k , denote the random variable and the state variable of wj;i;k , respectively. Sj is the state variable associated with j th scale.

In 3 , a new multiscale texture segmentation algorithm, called HMTseg, was proposed using the wavelet-domain HMT model. In the HMTseg, an image is recursively divided into four sub-images of same size J times and represented in a pyramid structure of J scales, as shown in Fig. 1d which has the same nesting quad-tree structure as the wavelet subtree in Fig. 1b. For an image of N  N decomposed from a J -scale Haar DWT, we denote a dyadic block as Dj;i;k which means the i; kth block in the j th scale, then its parent block is Dj+1;bi=2c;bk=2c . Given the Haar DWT of a test image, w, and a set of HMT model parameter = f LH ; HL ; HH g which consists of three HMT parameters for three subbands, respectively, Dj;i;k is associated with three wavelet coe cient subtrees in LH HL HH each subband, fTj;i;k ; Tj;i;k ; Tj;i;k g, and each subtree is rooted LH , wHL and wHH , respectively. Thus the computation in wj;i;k j;i;k j;i;k of the likelihood f Dj;i;k j  is a realization of the HMT model and is obtained by LH HL HH f Dj;i;k j  = f Tj;i;k j LH f Tj;i;k j HL f Tj;i;k j HH : 4 The computation algorithm of 4 can be found in 1 . The HMTseg algorithm was implemented by three steps. Firstly, the HMT model training is performed for the Haar DWT of each texture. Then, we compute the Haar DWT of the image to be segmented, and we obtain the Maximum Likelihood ML raw multiscale classi cation results based on the model likelihood computed by 4. Finally, the Bayesian interscale fusions are conducted by computing the maximum a posteriori MAP estimate of the class label of each dyadic block based on a Contextual Labeling Tree CLT. The CLT is speci ed by a context structure which is used for contextual information characterization and critical to the last segmentation result. If the class label of Dj;i;k is denoted as Cj;i;k = c, c = 1; :::; Nc and Nc the class number, and Dj;i;k is associated with a context vector Vj;i;k = v, v = 1; :::; Nv and Nv the number of di erent context values, thus the MAP estimate of Dj;i;k is obtained by ^ Cj;i;k = arg c2fmax g f Dj;i;k jCj;i;k = cpCj = cjVj = v; 1;:::;Nc 5 where we use two notations, Cj and Vj , to denote any Cj;i;k and Vj;i;k in the j th scale, the conditional pmf pCj = cjVj = v is associated with the j th scale, and the likelihood of

f Dj;i;k jCj;i;k = c is given by 4. The CLT is speci ed by two parameters, the pmf of the class labels, pCj c, and the conditional pmf, pVj = vjCj = c which are grouped into one parameter vector as, = fpCj c; pVj = vjCj = cjj = 1; :::; J g: 6 We de ne D, C and V as the collections of all Dj;i;k , Cj;i;k and Vj;i;k . The problem of interscale fusion can be interpreted as, given D and the structure of V, we want to nd and C which satisfy 5 and the solution of this problem can be approached by the EM algorithm proposed in 3 .
III. Image Denoising

Step 4. Set c = c + 1. If it converges, stop, else go to Step 2.

In fact, the within-scale scanning considers the wavelet coe cients are initially independent. The extra computational cost introduced by the within-scale scanning is negligible compared to the computational complexity of the HMT model training.

A.2 Across-Scale Counting

A Initialization for HMT Model Training


The initialization of the EM algorithm is important to the application the HMT model to image processing, since the initialization may e ect the convergence rate and the training result of the EM iterations. In this work, we develop a twostep initialization algorithm for HMT model training which allows more e cient and accurate computation of the HMT model parameters. The new initialization scheme is accomplished via two steps, Within-Scale Scanning and Across-Scale Counting.

This step is to provide the initial transition probabilities m;n j;j +1 between two neighboring scales. Given j and wj;k;i , j = 1; :::; J , k; i = 0; 1; :::; Nj , 1, we can determine the initial hidden state of wj;k;i , Sj;k;i , by Sj;k;i = arg mmax1g pSj;k;i = mjwj;k;i ; j ; 11 2f0; where the conditional pmf is obtained by 8. Then, given the state of each coe cient, we can count di erent state transition frequencies between every two neighboring scales along the quad-tree structured wavelet coe cient set. Therefore, we can set the initial values of the transition probability matrices m;n by the normalized transition frequencies as follows: j;j +1 m;n = Sj;k;i = m and Sj+1;bk=2c;bi=2c = njk; i = 0; ::::; Nj , 1 ; j;j+1 Sj+1;bk=2c;bi=2c = njk; i = 0; ::::; Nj , 1 where A denotes the number of the event A to occur. Using the within-scale scanning and across-scale counting, we can determine the initial value 0 of the HMT model parameter de ned in 2, which roughly characterize the global statistical properties of the whole wavelet coe cient set w. The EM algorithm for the HMT model training can then begin with this starting point, 0 , which allows a faster convergence rate and better training results.
12

A.1 Within-Scale Scanning

It was indicated in 1 that the wavelet coe cients in the same scale are assumed to have the same density. The purpose of this step is to determine the local statistics of the wavelet coe cients in di erent scales. Given w of N  N and J scales, there are Nj  Nj coe cients, wj , in the j th scale of each subband with Nj = N=2j , and the within-scale scanning is performed within the j th scale, and its task is to t a twostate Gaussian mixture model as 2 j = fpSj m; j;m = 0; j;m jm = 0; 1g 7 to wj . In the following, we will develop an EM algorithm 6 for the implementation of within-scale scanning which is performed within j th scale, j = 1; :::; J . Step 1. Set a neutral2setting of the Gaussian mixture model 0 of the j th scale , j , and the iteration counter c = 0. Step 2. E step: Calculate pSj;k;ijwj;k;i ; j  which is the pmf of the hidden state variable Sj;k;i given wj;k;i and j by the following equation,
pSj;k;i = mjwj;k;i ; j  =
 ;0 2 S P1mp=0j pmjgmwj;k;ij;k;i; ; 0j;mj;m ; gw ; 2 S 8

B Classi ed HMT Model

where function g is a Gaussian function. Step 3. M step: Set jc+1 = arg maxj E ln f wj ; Sj  . We c update the entries of j +1 :
1 pSj m = N 2
2 j;m =
2 2

Nj ,1 Nj ,1

P P

j k=0 i=0 Nj ,1 Nj ,1 w2 pS j;k;i = mjwj;k;i ; j  k=0 i=0 j;k;i : 10 Nj2 pSj m

X X pSj;k;i = mjwj;k;i; j 

9

2 The neutral setting means that pSj 0 = pSj 1 = 0:5, j;0 = 2 = 2 2 , 2 , where 2 is the AWGN variance and 2 , and j;1 j j the variance of wj .

The HMT model mainly characterizes the interscale dependencies of wavelet coe cients, and the intrascale dependencies are not well captured. In order to obtain more accurate image modeling using the HMT model, we propose the Classied 2-D HMT model CHMT where the wavelet subtrees are categorized into C di erent groups characterized by di erent HMT models. In practice, a simple but e cient classi cation scheme is to divide the coe cients in each scale of each subband into C blocks of same size. Therefore, the CHMT model can be considered as a model composed of C many 2-D HMT models, and we have C sets of the HMT parameters to characterize the CHMT model, C = f 1 ; 2 ; :::; C g. For simplicity, we only consider the case C = 4p , p = 1; 2; 3 in this work. Since the CHMT model is more speci c for the wavelet coe cient characterization and uses more model parameters, the accuracy of image modeling can be improved. Moreover, since the CHMT model training is performed separately for each HMT model which is associated with a small number of observed data, it is predicated that the model training of the CHMT model should be more e cient. On the other hand, we acknowledge that this consideration may also result in more risk of model over- tting" because of more model parameters introduced, and the lack of the spatial adaptability disables signi cant improvements by this strategy. It is expected that the adaptive spatial classi cation may allow more e ective image modeling. Despite of these drawbacks, the simple CHMT model still enables improved denoising performance for all test images in this work.

Y X

Z Y X Z X

Child block Neighboring block

V2 V1

V3 V2 V1

V2 V1

Block in the finner scale Parent block Neighboring block of the parent block Block in the coaser scale
V2 V1

Vector Context

a b c Figure 2: Three di erent context structures for the ner scale below based on the coarser scale above. a Context-1. b Context-2. c Context-3. The operation is de ned in 13.
Likelihood Computation
Haar DWT

Initialization (1)
arg max (.)

EM Algorithm

(1)

Initialization (2)

EM Algorithm

(2)

Initialization (3)

EM Algorithm

(3)
arg max (.)

arg max (.)

arg max (.)

Mosaic Image

HMT Model Dictionary

Raw-ML-Seg

Context-1

MAP-Seg-1

Context-2

MAP-Seg-2

Context-3

MAP-Seg-3

Figure 3: The block diagram of the improved HMTseg algorithm using HCLT. The dashed line represents the association of a context structure in the EM algorithm and the initialization. The operation of arg max : is de ned in 5 where the conditional decision is denoted by the dotted lines. The HMT model dictionary stores Nc many HMT model parameters, f 1 ; :::; Nc g. The Raw-ML-Seg" denotes the raw ML multiscale segmentation results, and MAP-Seg-1", MAP-Seg-2" and MAP-Seg-3" represent the step-by-step segmentation results.
between the neighborhood of the parent block and Dj;i;k . Z represents the relation between Dj;i;k and its neighborhood. From Fig. 2, we can determine the context vectors of Dj;i;k , Vj;i;k , by three di erent ways: 1 2 Context-1: d = 2, Vj;i;k = X , Vj;i;k = Y ; 1 2 3 Context-2: d = 3, Vj;i;k = X , Vj;i;k = Y , Vj;i;k = Z ; 1 2 Context-3: d = 2, Vj;i;k = X , Vj;i;k = Z . A Three Context Structures Three context structures above have di erent e ects on the We rst introduce the neighborhood" of a dyadic block Dj;i;k texture segmentation with respect to the classi cation of hoR which is denoted by Dj;i;k to indicate the region of eight neigh- mogeneous regions and the localization of texture boundaries. R boring blocks of Dj;i;k . Then we de ne the class label of Dj;i;k We expect that Context-1 is more e cient for the classi cation R which can be obtained by of homogeneous regions, Context-2 has some balanced considas Cj;i;k eration between homogeneous regions and texture boundaries, i+1 k+1 and Context-3 is more suitable for the localization of texture R Cj;i;k = arg c2fmax g pCj;m;n = cjD; V boundaries. This statement has been veri ed by extensive 1;:::;Nc m=i,1 n=k,1 experiments in practice. ,pCj;i;k = cjD; V; 13 where the conditional pmf's of eight neighboring blocks of B Hybrid Contextual Labeling Tree R R Dj;i;k but not itself are used to determine Cj;i;k . Cj;i;k is able In practice, neither single context works well, and it is rational to provide some context information regarding to the class to integrate them together. Based on the analysis in Section A label of Dj;i;k or those of its child blocks. major task of texture segmentation, There are two factors related to the vector context V = and theContextual Labeling Tree HCLT for we develop a Hybrid the 1 d fVp ; :::; V g, the dimension d, and the value of each entry fusion which cascades the three context structures in interscale the order V , p = 1; :::; d, which takes value in f1; :::; Nc g, thus we have of Context-1, Context-2 and Context-3. The EM training of d . In Fig. 2, we develop three context structures Nv = Nc  and we assume that the context of Dj;i;k may be related to the HCLT is performed based on three contexts, respectively and sequentially, so is the initialization of the EM algorithm. three variables de ned as follows, The block diagram of the improved HMTseg algorithm using X = Cj+1;x;y 14 the HCLT is shown in Fig. 3. We know that the initialization of the EM algorithm is ofY = CjR ;x;y 15 +1 ten important to the e ciency and e ectiveness of the model R Z = Cj;i;k 16 training. In this work, we use the counting method to dewhere x = bi=2c and y = bk=2c. X characterizes the e ect termine the initial setting of for the EM algorithm which from the parent block Dj+1;x;y , Y considers the dependencies enables the e cient HCLT model training.
IV. Texture Segmentation In the HMTseg algorithm, the context structure in the CLT plays an important role to the result of interscale fusion. In order to achieve more accurate multiscale information characterization for both homogeneous regions and texture boundaries, we develop a Hybrid Contextual Labeling Tree HCLT based on three contexts as follows.

X X

Initialization of the 2D HMT model Training 1


1

Image denoising using the Classified 2D HMT model


The training of the Classified 2D HMT model

35

0.9

0.95

34
0.8
0.9

The normalized HMT model likelihood

0.7

The normalized HMT model likelihood

0.85

The PSNR(dB) of the denoised image.

33

0.6

0.8

0.5

0.75

32

0.4

0.7

31

C=1 C=4 C=16 C=64

0.3

0.65

0.2

0.6

0.1

Using the neutral setting Using the proposed scheme


10 20 30 40 50 60 70 The iteration number of the EM algorithm 80 90 100

0.55

C=1 C=4 C=16 C=64


0 10 20 30 40 50 60 70 The iteration number of the EM algorithm 80 90 100

30

0.5

29 20

21

a b c Figure 4: Simulation results for the Lena image 512  512, 8bpp, = 20. a Comparison of the HMT model training using two di erent initialization schemes. b Comparison of the model training of the CHMT models of di erent C C = 1; 4; 16; 64. c Comparison of the image denoising performances of the CHMT models.

22 23 24 25 26 27 The PSNR(dB) of the noisy image, Lena (8bpp, 512 512).

28

29

a b c d Figure 5: a Noisy image 256  256, 8bpp, = 25, PSNR=20.3dB. b Donoho's Soft thresholding PSNR=25.5dB. c Wiener lter in Matlab PSNR=26.8dB. d CHMT model PSNR=27.6dB.
V. Simulation Results

A Image Denoising

We rst compare the proposed initialization scheme with the neutral initial HMT model setting in Fig. 4a, where we plot the model likelihood computed during the model training as the function of the iteration numbers of the EM algorithm. It is shown that the new initialization scheme makes the model training faster to converge and have larger likelihood, as well as improved denoising performance. Moreover, the performance of image denoising is usually stable after several iterations, so we x the iteration number to be 10 for all cases. The training of the CHMT model with di erent C is shown in Fig. 4b and the PSNR comparison of the image denoising performance is given in Fig. 4c. Two gures show the improvements of the CHMT model on both the e ciency of the model training and the e ectiveness of image denoising. We will use the CHMT model with C = 64 in the following. We show the denoising results of the Lena image using di erent techniques in Fig. 5 where we can see that the new algorithm using the CHMT model can preserve the details very well, and the noise is greatly suppressed without oversmoothing" e ects. We also compare the proposed algorithm with other denoising methods 2, 7, 8, 9, 10 in terms of PSNR gain in Table 1, which shows that the CHMT model using the proposed initialization scheme provides us with the state-ofthe-art denoising performance and outperforms the method reported in 2 by 0.4-0.9dB for two real images.

Table 1: The PSNR dB performance comparison of image denoising results for images 512  512, 8bpp corrupted by AWGN of di erent variances = 10; 15; 20.
Noisy Images Donoho' HT 11 WienerMatlab HMT 2 uHMT 2 CHMT SAWT 7 LAWMAP 9 AHMF 10 SSMMW 8 10 31.6 32.7 33.9 33.6 34.4 34.3 34.5 34.8 Lena 15 29.8 31.3 31.8 31.6 32.4 31.8 32.4 32.5 32.5 20 28.5 30.1 30.4 30.2 31.1 30.5 31.0 31.1 10 28.6 28.4 31.9 28.3 32.7 32.5 32.7 32.4 Barbara 15 26.5 27.4 29.4 26.4 30.3 29.2 30.2 30.3 30.0 20 25.2 26.5 27.8 25.3 28.7 27.6 28.6 28.7 -

B Texture Segmentation

We use two criteria to evaluate the segmentation performance, PAC which is the percentage accuracies of the pixel classi cation, and PAB which is the percentage of points on the localized the boundary that coincide with the true boundary. The simulation of this work was conducted on a number of examples composed of 128  128 texture mosaics taken from Brodatz album 12 . We show the segmentation results on ve mosaics in Fig. 6. We also tabulate the segmentation performances of PAC and PAB in Table 2. From Fig. 6 and Table 2, we can see that the newly proposed Hybrid Contextual Labeling Tree using three context structures does give us good results of texture segmentation in terms of both PAC and PAB . The segmentation result

Figure 6: The segmentation results on 5 mosaics using the HCLT. The rst row shows ve texture mosaics, Mosaic1 wood grass, Mosaic2 leather weave, Mosaic3 ra a grass leather, Mosaic4 weave wood leather cloth and Mosaic5 cloth leather weave wood. The second row presents the last segmentation results using the HCLT.
can be improved step-by-step using the HCLT. The last results of PAC are consistently better than 95 and those of PAB are between 48:7 and 100:0. We also compare our newly proposed segmentation algorithm to some state-of-theart algorithms in Table 3, where we can nd that the new algorithm has the highest PAC and the performance of PAB is comparable to that in 13 where the boundary localization in texture segmentation was investigated.

References

Table 2: The step-by-step segmentation results.


Texture Mosaics Mosaic1 Mosaic2 Mosaic3 Mosaic4 Mosaic5

PAC

Seg-MAP-1 PAB 97.5 33.8 97.7 34.0 97.4 48.2 96.0 40.4 99.8 81.5

PAC

Seg-MAP-2 PAB 98.7 64.0 97.8 40.4 97.7 53.7 96.9 54.1 100 97.3

PAC

Seg-MAP-3 PAB 99.2 77.8 98.2 48.7 98.4 67.1 97.2 59.9 100 100

Table 3: Comparison of segmentation performance.


References Fig. 6 using the HCLT Choi & Baraniuk 3 Yhann & Young 13 Kaplan 14 97.2-100 93.9-99.7 87.7-100 92.7

PAC

48.7-100 24.8-65.8 54.5-100 -

PAB

VI. Conclusions

In this paper, we have discussed the applications of the wavelet-domain HMT model to image denoising and texture segmentation based on 1, 2, 3, 4 . For image denoising, we proposed a simple initialization scheme for more e cient and e ective HMT model training and the Classi ed HMT CHMT model for more accurate image modeling, so that image denoising performance can be improved by 0.4-0.9dB for two real images. For texture segmentation, we developed three new context structures for more accurate multiscale information characterizations, which are integrate together to form the Hybrid Contextual Labeling tree. The new segmentation algorithm improves the accuracies of both texture classi cation and boundary localization.

1 M. S. Crouse, R. D. Nowak, and R. G. Baraniuk, Wavelet-based statistical signal processing using hidden Markov models," IEEE Trans. Signal Processing, vol. 46, no. 4, pp. 886 902, April 1998. 2 J. K. Romberg, H. Choi, and R. Baraniuk, Bayesian treestructured image modelling using wavelet-domain hidden Markov model," in Proc. of SPIE, Denver, CO, July 1999, vol. 3816, pp. 31 44. 3 H. Choi and R. Baraniuk, Image segmentation using waveletdomain classi cation," in Proc. of SPIE, Denver, CO, July 1999, vol. 3816, pp. 306 320. 4 H. Choi and R. G. Baraniuk, Multiscale image segmentation using wavelet-domain hidden Markov models," Preprint. 5 H. Chipman, E. Kolaczyk, and R. McCulloch, Adaptive Bayesian wavelet shrinkage," J. Ameri. Stat. Assoc., vol. 440, no. 92, pp. 1413 1421, Dec. 1997. 6 B. Everitt, Ed., Finite Mixture Distribution, Chapman and Hall, 1981. 7 S. G. Chang, B. Yu, and M. Vetterli, Spatially adaptive wavelet thresholding with context modelling for image denoising," in Proc. of ICIP, Chicago, IL, Oct. 1998. 8 J. Liu and P. Moulin, Image denoising based on scale-space mixture modeling of wavelet coe cients," in Proc. of IEEE Int. Conf. on Image Proc., Kobe, Japan, Oct. 1999. 9 M. K. Mihcak, I. Kozintsev, and K. Ramchandran, Lowcomplexity image denoising based on statistical modeling of wavelet coe cients," IEEE Signal Processing Letters, vol. 6, no. 12, pp. 300 303, Dec. 1999. 10 S. Xiao, I. Kozintsev, and K. Ramchandran, Stochastic waveletbased image modeling using factor graphs and its application to denoising," in Proc. of SPIE, 2000, to appear. 11 D. Donoho and I. Johnstone, Adapting to unknown smoothness via wavelet shrinkage," J. Ameri. Stat. Assoc., vol. 90, pp. 1200 1224, Dec. 1995. 12 P. Brodatz, Textures|A Photographic Album for Artists and Designers, New York: Dover, 1966. 13 S. R. Yhann and T. Y. Young, Boundary localization in texture segmentation," IEEE Trans. on Image Processing, vol. 4, no. 6, pp. 849 855, June 1995. 14 L. M. Kaplan, Extended fractal analysis for texture classi cation and segmentation," IEEE Trans. on Image Processing, vol. 8, no. 11, pp. 1572 1585, November 1999.

You might also like