adaptive 2-d wavelet transform based on the_lifting scheme with preserved vanishing moments_
TRANSCRIPT
-
8/7/2019 Adaptive 2-D Wavelet Transform Based on the_Lifting Scheme with Preserved Vanishing Moments_
1/17Copyright (c) 2010 IEEE. Personal use is permitted. For any other purposes, Permission must be obtained from the IEEE by emailing [email protected].
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON IMAGE PROCESSING, MIROSLAV VRANKIC, DAMIR SERSIC, AND VICTOR SUCIC 1
Adaptive 2-D Wavelet Transform Based on the
Lifting Scheme with Preserved Vanishing MomentsMiroslav Vrankic*, Member, IEEE, Damir Sersic, Member, IEEE, and Victor Sucic
AbstractIn this paper, we propose novel adaptive waveletfilter bank structures based on the lifting scheme. The filterbanks are nonseparable, based on quincunx sampling, withtheir properties being pixel-wise adapted according to the localimage features. Despite being adaptive, the filter banks retaina desirable number of primal and dual vanishing moments.The adaptation is introduced in the predict stage of the filterbank with an adaptation region chosen independently for eachpixel, based on the intersection of confidence intervals (ICI) rule.The image denoising results are presented for both syntheticand real-world images. It is shown that the obtained waveletdecompositions perform well, especially for synthetic imagesthat contain periodic patterns, for which the proposed method
outperforms the state of the art in image denoising.
Index TermsWavelets, second generation wavelets, adaptivelifting scheme, quincunx sampling, interpolating filters, intersec-tion of confidence intervals, image denoising.
I. INTRODUCTION
In many applications, such as image denoising or compres-
sion, transforms are used to obtain a compact representation
of the analyzed image. The wavelet transform relies on a set
of functions that are translates and dilates of a single mother
function, and provides sparse representation of a large class
of real-world signals and images. Wavelet thresholding has
been shown to be a very efficient method for image denoising[1]. The choice of the threshold determines the compromise
between the noise reduction and the degradation of the original
features in the reconstructed image.
In order to obtain a wavelet transform that gives more
compact representation of an image, we propose a pixel-wise
adaptation of the analysis functions according to its local
properties. A locally adaptive wavelet transform that is tuned
to the image features results in transform coefficients which
predominantly contain the noise, i.e. the content that is not
inherent to the original image. In that case, application of a
rather high threshold reduces the noise efficiently, with lower
degradation of the original image.
The properties of the wavelet decomposition depend on thesubsampling scheme used in the corresponding filter bank
[2]. To obtain a more isotropic 2D wavelet transform than
the usual separable approach, we used the quincunx scheme,
M. Vrankic and V. Sucic are with the Faculty of Engineering, Uni-versity of Rijeka, Vukovarska 58, HR-51000 Rijeka, Croatia (e-mail:[email protected]; [email protected])
D. Sersic is with the University of Zagreb, Faculty of Engineering andComputing, Unska 3, HR-10000 Zagreb, Croatia (e-mail: [email protected])
Manuscript received xxxx, 2008; revised xxxxxxEDICS category: TEC-MRS (Multiresolution Processing of Images &
Video - Wavelets; Filter banks)
which is the simplest nonseparable sampling [3]. A compre-
hensive overview of two- and multi-dimensional FIR perfect
reconstruction filter banks (FBs) for arbitrary sampling lattices
can be found in [4] and [5], while multidimensional wavelet
FBs design methods are presented in [6], [7].
The adaptation of wavelet transforms can generally be
performed in two ways: globally or locally. An example of
a global adaptation approach is the best basis algorithm [8],
in which, depending on the criterion (e.g. signal entropy),
the optimal basis for the whole image is obtained. Local
adaptation, on the other hand, is based on the local features
of the analyzed image, and it generally outperforms the global
adaptation methods.
A number of authors have used the lifting scheme in order to
build locally adaptive wavelet FBs. Claypoole et al. proposed
the so-called space-adaptive transform (SpAT) based on the
lifting scheme, which changes the predictor order for every
signal sample to minimize detail coefficients [9]. Therefore,
near the edges present in the image, shorter prediction filters
are used to avoid ringing effects. The update step is per-
formed first, followed by the adaptive prediction. This allows
the maintaining of multiresolution properties across different
decomposition levels in spite of the nonlinearity introduced
through the adaptive predict step.A locally adaptive filter bank structure for image coding has
been proposed by Gerek and Cetin [10]. Their FB is based
on the lifting scheme with adaptation of the prediction filter.
The least mean squares (LMS) adaptation algorithm is used to
minimize the signal at the high-pass channel output. However,
the proposed adaptive structures do not retain polynomial
annihilation properties inherent to wavelets.
An adaptive prediction stage is also employed in [11] to
improve the image decorrelation efficiently for the purpose
of lossless image compression. Similarly to the previous
approach, these FBs lack some crucial wavelet properties, such
as polynomial annihilation.
The approaches presented in [12] and [13], on the otherhand, use the update first structure with adaptation of the
update step. The update filter is changed based on the local
gradient information such that sharp variations in the signal
get less smoothened than the more homogenous regions.
Moreover, the choices of update filters can be automatically re-
produced on the synthesis side, so no bookkeeping is required
for a perfect reconstruction.
In another approach presented in [14], to obtain a multidi-
mensional nonseparable wavelet transform, the prediction step
is designed to minimize the variance of the signal, while the
update step minimizes the reconstruction error.
-
8/7/2019 Adaptive 2-D Wavelet Transform Based on the_Lifting Scheme with Preserved Vanishing Moments_
2/17
Copyright (c) 2010 IEEE. Personal use is permitted. For any other purposes, Permission must be obtained from the IEEE by emailing [email protected].
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON IMAGE PROCESSING, MIROSLAV VRANKIC, DAMIR SERSIC, AND VICTOR SUCIC 2
Recently, many novel wavelet decomposition schemes that
use directional information from images have been introduced,
either adaptive or non-adaptive: curvelets, ridgelets, bandelets,
directional lifting-based wavelets, directionlets etc. [15][17].
Some authors have developed very useful directional algo-
rithms for compression, showing advantages over traditional
separable non-directional wavelets. Nevertheless, there is a
lack of papers that deal with wavelet adaptation in the presence
of noise. In this paper, rather than using directional information
from images, we use more isotropic quincunx 2D wavelets,
and pay attention to reducing the influence of noise on the
adaptation.
A milestone denoising method was introduced by Portilla
[18]. Pointwise shape-adaptive DCT [19] is a novel adap-
tive image decomposition with applications in denoising that
outperforms in many cases the Portillas method. The SA-
DCT changes the shape of the transforms support at each
pixel of the image, while in our work we change the support
of the adaptation region, and, according to the result, the
base functions. Both approaches rely on the intersection of
confidence intervals (ICI) rule. We used SA-DCT, as one ofthe state-of-the-art image denoising methods, for comparison.
A. The Lifting Scheme
The lifting scheme provides an effective tool to build
biorthogonal wavelet FBs [20]. Moreover, due to its inherent
perfect reconstruction properties, it is a great tool to build
second-generation wavelets [21], i.e. wavelets that are not
necessarily translated and dilated copies of a single function.
An example of the lifting scheme structure based on quin-
cunx sampling is shown in Fig. 1. It is built from three
main parts: polyphase decomposition, predict and update steps.
The image at the FB input is split into M = | det(D)|phases, where D represents the subsampling matrix [3]. For
the quincunx case, the image is split in two phases as defined
by the quincunx subsampling matrix:
Dq =
1 1
1 1
. (1)
As shown in Fig. 1, the polyphase decomposition is fol-
lowed by the prediction step. The filter P predicts the pixelvalues of the second (odd) phase, based on the pixels from
the first (even) phase. The prediction error corresponds to
the wavelet detail coefficients, e.g. to the output of the high-
pass channel. The subsequent update step corresponds to the
approximation coefficients, e.g. produces the low-pass output
of the FB.
The reconstruction is readily obtainable. The synthesis FB
is constructed using the same steps but in the reverse order
and with opposite signs. This way, the operations from the
analysis side are reversed and cancelled on the synthesis side.
B. The Proposed Method
An important property of wavelets is their ability to anni-
hilate polynomials. Many real-world signals and images can
be locally viewed as polynomial segments, so the polyno-
mial annihilation is a key to sparse representation in the
transform domain. The sparseness is a desired property for
many applications: feature extraction, denoising, compression,
compressed sensing, blind separation of dependent sources
[22], and many others. Another important property of wavelets
is their multi-resolution nature: numerous families of octave
wavelets (either bases or redundant frames) can be realized
using wavelet filter banks in a numerically efficient way. By
introducing adaptation, we do not want to lose any of the
winning properties of the wavelets!
When building the adaptive wavelet FB, one would like to
retain a number of dual and primal vanishing moments. They
are necessary for convergence of the corresponding wavelet
and scale functions across different decomposition levels,
and they contribute to their smoothness and regularity. The
convergence of decomposition and/or reconstruction functions
across different levels is a key link between wavelets and FBs.
Moreover, the number of vanishing moments is closely related
to the order of the polynomials that the wavelet FB is able to
annihilate [23].
In this paper, we propose two novel adaptive non-separable
2D filter bank structures based on the lifting scheme andquincunx sampling. By using the lifting scheme, the existing
wavelet transform is made adaptive and further improved with
additional lifting steps, while at the same time a desirable
number of vanishing moments is retained. The structures and
adaptation algorithms are improved when compared to our pre-
viously reported work in [24] and [25]. In [24], we proposed
adaptive 1D wavelet FB structures that can be generalized to
2D using a well-known separable approach.
In this work, the adaptation is introduced through an ad-
ditional predict stage with a parameter which is tuned for
each pixel separately, based on the adaptively chosen pixel
neighborhood. Unlike our work presented in [26] and [27],
which is based on changing filter supports for each signalsample, in this paper we do not change filter supports, but
rather change the region on which a FB parameter is being
calculated.
The paper is organized as follows. In Section II, the two
adaptive FB structures, named PPaU and PUPa, are pre-
sented, the first of which was proposed in [25]. The proposed
structures can provide any number of vanishing moments or
adaptive parameters. In this paper, we have chosen the FB
structure PPaU with the shortest possible (i.e. canonical)
symmetric support that provides two primal and two dual
vanishing moments, followed by a single adaptive parameter.
However, the quest for independent adaptation across differ-
ent decomposition levels, with smooth and regular synthesisfunctions, brought us to the structure PUPa. The proofs of
vanishing moment preservation are given in the Appendix.
In Section III the minimization of energy of detail coeffi-
cients is introduced as the adaptation criterion. The adaptation
is performed for every pixel of the image based on the
neighborhood of a free shape. Assuming an additive noise
model, we show that the adaptation is biased. However, we
have found a way to make the adaptation almost insensitive
to noise. The proof is given in Section III-D.
In Section IV we present a method for obtaining the
appropriate adaptation neighborhood for each pixel of the
-
8/7/2019 Adaptive 2-D Wavelet Transform Based on the_Lifting Scheme with Preserved Vanishing Moments_
3/17
Copyright (c) 2010 IEEE. Personal use is permitted. For any other purposes, Permission must be obtained from the IEEE by emailing [email protected].
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON IMAGE PROCESSING, MIROSLAV VRANKIC, DAMIR SERSIC, AND VICTOR SUCIC 3
+
D
D
-
+ -
D
D1 2
( , )X z z
1 2( , )X z z
1 2( , )P z z 1 2( , )P z z1 2( , )U z z1 2( , )U z z
1
1z
1z
eX
oX D
Fig. 1. Quincunx filter bank based on the lifting scheme. Downsampling and upsampling operators are defined with a quincunx subsampling matrix,D
=Dq.
image. The method is based on the intersection of confi-
dence intervals (ICI) rule [28]. The ICI rule provides for the
maximum size of local neighborhoods of similar statistical
properties, thus providing for the best unbiased estimation. To
find the confidence intervals limits, we used the Monte Carlo
approach.
Finally, in Section V the proposed adaptation algorithm is
used within the framework of image denoising by wavelet
thresholding. The image denoising results are presented for
both synthetic and real-world images. Our trade-off between
more vanishing moments and local adaptation using the pro-posed structures has shown to be useful for images containing
periodic components, as expected.
II. THE ADAPTIVE LIFTING SCHEME
A. Quincunx Interpolating Filter Bank
The construction of compactly supported biorthogonal
wavelets and perfect reconstruction filter banks for any lattice,
in any dimension, and with any number of primal and dual
vanishing moments is presented in [29].
The filters used to build the prediction step, called Neville
filters, are based on the polynomial interpolation. Filtering the
polynomial of degree lower than N with the Neville filter P ofdegree N or higher gives the samples of the same polynomialbut sampled on the lattice shifted by . Following the notation
in [29], we define to be a multivariate polynomial, and (Z2)the sequence formed by evaluating the polynomial on the
lattice Z2. Then, applying the Neville filter P to the sampledpolynomial yields
P (Z2) = (Z2 + ), (2)
where R2.The construction of such filters relies on solving the inter-
polation problem. The solution of the interpolation problem in
multiple dimensions is given by de Boor and Ron in [30]. In
[29] the authors give Neville filter coefficients for the quincunx
case obtained by using the de Boor-Ron algorithm. The predict
filter coefficients are given in Table I. Coefficients with equal
values are grouped into rings. The ring numbering scheme is
shown in Fig. 2.
B. Vanishing Moments
An analysis FB has N vanishing moments if its high-passchannel annihilates polynomial sequences of degree lower than
N. Also, a synthesis FB has N vanishing moments if its syn-thesis high-pass channel annihilates polynomial sequences of
degree lower than N. The vanishing moments of the analysis
TABLE IQUINCUNX NEVILLE FILTER COEFFICIENTS [29]. THE COEFFICIENTS ARE
GROUPED INTO RINGS, AS SHOWN IN FIG . 2.
Ring 1 2 3 4 5 6 7
Order
2 1 22
4 10 -1 25
6 174 -27 2 3 29
8 23300 -4470 625 850 -75 9 -80 216
filter bank are called dual, while the vanishing moments ofthe synthesis FB are called primal.
It was shown in [29] that in order to have N dual vanishingmoments, the predict filter P of the lifting scheme has to bea Neville filter of order N and shift
= D1t, (3)
where D represents the subsampling matrix and t defines
the shift between the two phases. For the quincunx sampling
matrix from (1), where t = [1 0]T, the shift is = [0.5 0.5]T.As it was also shown in [29], the primal vanishing moments
are defined with the update filter U of the lifting scheme. Inorder to obtain N primal vanishing moments, with P beinga Neville filter of order N and shift = D1t, 2U hasto be a Neville filter of order N and shift . Consequently,2U has to be a Neville filter of order N and shift . Thesimplest choice is to make the update filter the adjoint of the
same-order predict filter divided by two:
U = UN =1
2PN, (4)
where PN is a Neville filter of order N and shift [29].It is important to note that the prediction filter by itself
defines the number of dual vanishing moments, and the update
filter influences only the number of primal vanishing moments
as long as N N [29].For example, the FB with 2 dual vanishing moments and
2 primal vanishing moments will be constructed using theprediction filter P2 and the update filter U2 = 1/2P2 (seeTable I):
P2(z1, z2) =1
4(1 + z11 + z
12 + z
11 z
12 ), (5a)
U2(z1, z2) =1
8(1 + z1 + z2 + z1z2). (5b)
C. The PPaU structure
The adaptive lifting scheme structure, of which a simpler
version has been proposed in [25], is shown in Fig. 3. It
consists of a fixed prediction step PN that guarantees N
-
8/7/2019 Adaptive 2-D Wavelet Transform Based on the_Lifting Scheme with Preserved Vanishing Moments_
4/17
-
8/7/2019 Adaptive 2-D Wavelet Transform Based on the_Lifting Scheme with Preserved Vanishing Moments_
5/17
Copyright (c) 2010 IEEE. Personal use is permitted. For any other purposes, Permission must be obtained from the IEEE by emailing [email protected].
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON IMAGE PROCESSING, MIROSLAV VRANKIC, DAMIR SERSIC, AND VICTOR SUCIC 5
X
D
D
D
-
z1-1
Xe
Xo
UN
A
+
PN~
-
z1
D
D
+
PN~UN
Pa
-
p
+
Pa
p
X
Fig. 4. The PUPa analysis and synthesis adaptive filter bank structure.
take, if the Pa filter is chosen as in (6), where
R N , and S N. (10)
Also, in Appendix B we show that N primal vanishingmoments are preserved as long as
N N. (11)
E. The p parameter
In this paper, we use FBs that have 2 primal and 2 dualvanishing moments fixed, obtained by the following filter
selection:
PN = P2, (12a)
Pa = P4 P2, (12b)
UN = U2. (12c)
This is a canonical (i.e. simplest possible) structure with
symmetric analysis and synthesis 2D functions. As the value
of parameter p is changed, the FB properties will change ac-cordingly. The magnitude frequency responses of the analysis
high-pass filters of the PPaU structure for different values of
parameter p are shown in Fig. 5. As shown in Fig. 5, with
the increase of parameter p the zero ditch appears and growswider in the resulting magnitude response of the analysis high-
pass filters. Therefore, by changing the value of parameter p,it is possible to cancel some frequency components present
in the input image. Similar results are obtained for the PUPa
structure as well.
The limit analysis and synthesis wavelet functions for both
PPaU and PUPa structures for various values of parameter pare shown in Figs. 6-9. The limit functions are obtained by
setting a single value of parameter p at all decomposition levelsand calculating the impulse response of the 7-level iterated FB.
As shown in Figs. 7 and 9, the limit synthesis wavelet
functions are regular for smaller values of parameter p. For
absolute values ofp greater than 5, the limit functions becomeincreasingly irregular. However, for a limited range ofp values,well-behaved analysis and synthesis functions are obtained.
The analysis wavelet functions behave differently for the
two proposed structures. For the PPaU structure, the analysis
wavelet functions change significantly with increasing para-
meter p. Yet, for the PUPa structure, the analysis waveletfunctions are similar for the whole range of p values. Thereason for such behavior lies in the fact that the analysis low-
pass filter of the iterated PUPa filter bank does not depend on
the value ofp, so the iterated analysis high-pass filter changesonly due to the value ofp in the last level of the FB cascade.
III. ADAPTING THE FILTER BANK PARAMETERS
A. Noise Model
An input image corrupted with additive zero-mean white
Gaussian noise w can be modeled as
x(n) = x0(n) + w(n), (13)
where the original image x0 will be treated as deterministic,and the samples of the noise are independent and identically
distributed (IID) random variables, w(n) N(0, 2). Forsimplicity of notation, we will use linear indexing of the image
pixels such that the two-dimensional indexes (j, k) will bereplaced by a single index n.
In deriving the method for adapting the values of parameter
p we will use the undecimated FBs, since it has been shownthat for denoising purposes they outperform decimated filter
banks [31]. The results obtained for the undecimated FB will
be straightforwardly applicable to the decimated FB structure
as well. The adaptation of parameter p will first be derivedfor the PPaU structure and then generalized for the PUPa
structure.
B. The Undecimated Filter Bank and Pixel Numbering
The undecimated analysis PPaU filter bank is shown in Fig.
10(a) and the simplified equivalent of its high-pass channel is
shown in Fig. 10(b). In the figures, the upsampling of the
filters impulse response is denoted by P2(zD), where z =[z1 z2]T. For the case of quincunx subsampling matrix Dgiven in (1) we therefore have P2(z1z
12 , z1z2).
The xp represents the image at the output of the filter P42 =P4 P2 which is quincunx-upsampled, having the impulseresponse:
p42 =1
32
0 1 0 1 0
1 0 2 0 10 2 0 2 0
1 0 2 0 1
0 1 0 1 0
. (14)
The image denoted as xh, resulting after the first predictionstep, is obtained by convolving the input image with the
impulse response of the high-pass filter
h =
0 14
0
14
1 14
0 14
0
. (15)
-
8/7/2019 Adaptive 2-D Wavelet Transform Based on the_Lifting Scheme with Preserved Vanishing Moments_
6/17
Copyright (c) 2010 IEEE. Personal use is permitted. For any other purposes, Permission must be obtained from the IEEE by emailing [email protected].
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON IMAGE PROCESSING, MIROSLAV VRANKIC, DAMIR SERSIC, AND VICTOR SUCIC 6
0
3 0
30
0.5
1
1.5
(a)
0
3 0
30
1
2
(b)
0
3 0
30
1
2
(c)
-3 0 3
-3
0
3
(d)-3 0 3
-3
0
3
(e)-3 0 3
-3
0
3
(f)
Fig. 5. Magnitude frequency responses of the analysis high-pass filters of the PPaU structure with 2 vanishing moments fixed, obtained for different valuesofp (first row), and the appropriate zero locations of the responses (second row). p = 1 for the first column, p = 2 for the second, and p = 5 for the thirdcolumn.
X
P2(zD)
D-
z1-1
P42(zD)
-
+
Xh
A
U2(zD
)
Xh^p
Xp
X
D
P42(zD)
-
p
H(zD)Xh
Xh^
Xp
(a)
(b)
Fig. 10. (a) The first decomposition level of the adaptive PPaU filter bankwhere P42 = P4P2, and (b) the simplified scheme of its high-pass channel.
In order to enumerate pixels from the even quincunx phase
relative to the prediction center n located in the odd phase, we
introduce an alternative notation x(n, i). The sample x(n, i) islocated at position i of the constellation in Fig. 11(a), whichis centered on pixel x(n). Therefore, a pixel xh(n), as shown
5
2
4
1
6
3
9
7
8
10
11
12
13
5
2
4
1 3
(a) (b)
Fig. 11. Filter supports for (a) p42 filter, and (b) h filter are marked with graycircles. The relative pixel numbering starts from the support center (markedin bold).
in Fig. 11(b), is a weighted sum of 5 pixels from the input
image x
xh(n) = x(n, 1) 14
(x(n, 2) + x(n, 3) + x(n, 4) + x(n, 5)).
(16)
The xp(n) and xh(n) from Fig. 10(b) can be expressed interms of the original input image x0(n) and the random noisew(n) as:
xp(n) = (x0(n) + w(n)) p42(n) = x0p(n) + wp(n),
xh(n) = (x0(n) + w(n)) h = x0h(n) + wh(n).(17)
The signals x0p and x0h, which result from filtering theoriginal image, are considered to be deterministic components
ofxp and xh respectively. However, the components wp andwh are stochastic, coming from filtering the input noise.
-
8/7/2019 Adaptive 2-D Wavelet Transform Based on the_Lifting Scheme with Preserved Vanishing Moments_
7/17
Copyright (c) 2010 IEEE. Personal use is permitted. For any other purposes, Permission must be obtained from the IEEE by emailing [email protected].
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON IMAGE PROCESSING, MIROSLAV VRANKIC, DAMIR SERSIC, AND VICTOR SUCIC 7
(a) (b) (c)
(d) (e) (f)
Fig. 6. The analysis limit wavelet functions of the PPaU structure obtained for a range ofp values. (a) p = 1, (b) p = 0 , (c) p = 1 , (d) p = 2 , (e) p = 3 ,(f) p = 5 .
(a) (b) (c)
(d) (e) (f)
Fig. 7. The synthesis limit wavelet functions of the PPaU structure obtained for a range ofp values. (a) p = 1, (b) p = 0, (c) p = 1, (d) p = 2, (e)p = 3 , (f) p = 5 .
C. Minimizing the Energy of Detail Coefficients
The image xh represents detail coefficients resulting froma dual wavelet with two vanishing moments. In general, the
filters used in the prediction branch are Neville filters of
degree N. Hence, if the input to the FB is a polynomial ofdegree lower than N, the prediction is perfect, leading to zerodetail coefficients. Our goal is to improve the prediction by
additionally annihilating non-polynomial components, such as
periodic ones. This would lead to smaller detail coefficients
and finally to a more compact image representation. In order
to obtain a sparse image representation we minimize the
detail coefficients, by changing the value of parameter p. Weminimize the energy of detail coefficients, which leads to the
well-known least squares solution.
The prediction error for the n-th pixel is defined as
e(n) = xh(n) xh(n), (18)
where
xh(n) = p(n) xp(n). (19)
The adaptation of parameter p is performed for every pixelof the analyzed image based on an appropriately chosen
neighborhood. A given pixel and its neighborhood form a
-
8/7/2019 Adaptive 2-D Wavelet Transform Based on the_Lifting Scheme with Preserved Vanishing Moments_
8/17
Copyright (c) 2010 IEEE. Personal use is permitted. For any other purposes, Permission must be obtained from the IEEE by emailing [email protected].
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON IMAGE PROCESSING, MIROSLAV VRANKIC, DAMIR SERSIC, AND VICTOR SUCIC 8
(a) (b) (c)
(d) (e) (f)
Fig. 8. The analysis limit wavelet functions of the PUPa structure obtained for a range ofp values. (a) p = 1, (b) p = 0 , (c) p = 1 , (d) p = 2 , (e) p = 3 ,(f) p = 5 .
(a) (b) (c)
(d) (e) (f)
Fig. 9. The synthesis limit wavelet functions of the PUPa structure obtained for a range ofp values. (a) p = 1, (b) p = 0, (c) p = 1, (d) p = 2, (e)p = 3 , (f) p = 5 .
region on which the value of the parameter p is calculated.For a pixel indexed with n, the appropriate region is denotedas Rn. (Determination of regions Rn is dicussed in SectionIV). For every region Rn the adaptation algorithm sets thevalue of the parameter p(n) in order to minimize the sum ofsquared errors. So, the function
F(p(n)) =iRn
e2(i) =iRn
(xh(i) p(n) xp(i))2 (20)
needs to be minimized with respect to p. The least squares (LS)solution minimizes the energy of detail coefficients. Optimal
parameter p(n) is the one for which
F(p(n))
p(n)=iRn
2(xh(i)p(n) xp(i))(xp(i)) = 0, (21)
which gives
p(n) =
iRn
xh(i) xp(i)
iRn
x2p(i). (22)
The numerator and denominator in (22) are functions of
pixels xp(i) and xh(i) belonging to the region Rn. If the
-
8/7/2019 Adaptive 2-D Wavelet Transform Based on the_Lifting Scheme with Preserved Vanishing Moments_
9/17
Copyright (c) 2010 IEEE. Personal use is permitted. For any other purposes, Permission must be obtained from the IEEE by emailing [email protected].
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON IMAGE PROCESSING, MIROSLAV VRANKIC, DAMIR SERSIC, AND VICTOR SUCIC 9
image does not contain the noise, the local adaptation results
in a well-adapted sparse representation. Parameter p is usuallyset to cancel periodic components that have remained after the
fixed polynomial predictor. In the presence of white noise, one
would expect that the estimator (22) is statistically unbiased
and that a larger region should decrease the variance of the
estimation.
The random variable wp(i), which is a part of xp(i), is afunction of a number of independent random variables of the
additive noise present in the input image from the constellation
given in Fig. 11(a). In a similar manner, wh(i), which is apart of xh(i), is a weighted sum of 5 independent randomvariables (see Fig. 11(b)). Therefore, the summations in the
numerator and denominator ofp in (22) are dependent randomvariables. Hence, the statistical properties of the estimated
p are influenced not only by the input noise properties, butalso by the shape of the adaptation region and the value of
the original image, which is generally unknown in image
denoising applications. Unfortunately, the estimator (22) is
biased, and we need some extra steps to cope with the problem.
D. Reducing the Bias of the Estimator for p
The bias ofp is defined as
bias(p(n)) = E[p(n) p(n)]. (23)
If one knew the estimation bias, then an unbiased estimator
ofp could be calculated as
pu(n) = p(n) + bias(p(n)). (24)
Unfortunately, the unbiased estimator is not available, since
there is no closed form expression for the expectation of p[32].
The true value of parameter p is obtained as a function ofthe original image pixels only, i.e.
p(n) =
iRn
xh0(i) xp0(i)iRn
x2p0(i)=
q(n)
r(n). (25)
The estimate ofp is obtained as a function of the image pixelscorrupted with noise
p(n) =
iRn
xh(i) xp(i)
iRn
x2p(i)=
q(n)
r(n). (26)
In Appendix C the closed form expressions for the biases
of the numerator and denominator of p are obtained and it
is shown that the values of bias(q) and bias(r) are neitherdependent on the value of the original signal x0 nor on theshape of the region. The biases depend only on the size of the
region on which the estimation is based and the variance of the
additive zero-mean noise present in the image. It is reasonable
to assume that the variance of the additive noise is known (or
can be easily estimated).
Therefore, we introduce an improved estimator pc as
pc(n) =q(n) + bias(q(n))
r(n) + bias(r(n))=
qc(n)
rc(n), (27)
which is not exactly compensating (24), but it gives acceptable
estimates of parameter p.
Using expressions (75) and (79) for the bias compensation,
the improved estimator of parameter p is obtained as
pc(n) =
iRn
xh(i) xp(i) +N
162w
iRn
x2p(i) 6N
2562w
, (28)
where N represents the number of samples included in theadaptation region Rn. Acceptable results are obtained as longas the denominator is not close to 0. Fortunately, in that case
the adaptation is not necessary at all, since the signal is already
well suppressed by the fixed polynomial part.
The histograms of p(n) and pc(n), calculated for a singlepixel of a test image, are shown in Figs. 12(a) and 12(b)
respectively. The test image used is composed of a single sine
wave pattern with period T = 10 and amplitude of 100. Ascan be seen from the figures, the improved estimator has a
significantly reduced bias.
Similar results are obtained for the PUPa structure as well.
The adaptive PUPa analysis FB is shown in Fig. 13(a), and
the equivalent simplified structure is shown in Fig. 13(b).
X
P2(zD)
D-
z1-1
P42(zD)
-Xh
A
Xh^p
Xp
+
U2(zD)
X
D
P242(zD)
-
p
H(zD)Xh
Xh^
Xp
(a)
(b)
Fig. 13. (a) The first decomposition level of the adaptive PUPa filter bank,and (b) the simplified scheme of its high-pass channel.
Similarly to the derivation in Section C, the improved
estimation of parameter p can be obtained as
pc(n) =
iRn
xh(i) xp(i) +13N
2562w
iRn
x2p(i) 2533N
1310722w
. (29)
-
8/7/2019 Adaptive 2-D Wavelet Transform Based on the_Lifting Scheme with Preserved Vanishing Moments_
10/17
Copyright (c) 2010 IEEE. Personal use is permitted. For any other purposes, Permission must be obtained from the IEEE by emailing [email protected].
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON IMAGE PROCESSING, MIROSLAV VRANKIC, DAMIR SERSIC, AND VICTOR SUCIC 10
0 0.5 10
100
200
300
(a)
0.5 1 1.50
100
200
300
(b)
Fig. 12. Histograms of (a) p(n) and (b) improved pc(n) for a single pixel of an image with T = 10 sine-wave calculated for a square region of size 5 5.The example is given for the PPaU structure. Input noise is white Gaussian zero-mean with w = 10. The true value of parameter p = 1.106 is shown bythe vertical dash-dot line. The bias of p(n) is 0.2273, while for pc(n) the bias equals to 0.0031.
E. Adaptation in the Iterated Filter Bank
For the iterated PPaU filter bank, the adaptation of parame-
ter pj at the j-th level is based on the corresponding Pj42and Hj filters, which are based on the j-times quincunx-upsampled predict and update filters. Therefore, pj42 is theimpulse response of the j-level FB structure up to the input tothe multiplier pj . Also, hj is the impulse response of the j-level FB structure up to the summation operator at the output
of the pj multiplier, much as in Fig. 10(b).It is important to note that the adaptation at lower levels
affects the Pj42 and Hj filters at all the successive levels. Since
the adaptation is being performed pixel-wise, the impulse
responses pj42(n) and hj(n) are different for every pixel
position n. Therefore, the bias correction at level j will havedifferent values for every pixel of the image, depending on
the adaptation at the previous j 1 levels. The estimation ofparameter p at level j and for the position n is obtained as
pjc(n) =qj(n) + bias(qj(n))
rj
(n) + bias(rj
(n))
. (30)
The numerators bias is obtained as
bias(qj(n)) = iRn
E
wjh(i)wjp(i)
= N 2wk=l
hj(n, k)pj42(n, l),(31)
and the denominators bias is obtained as
bias(rj(n)) = iRn
E
(wjp(i))2
= N 2wk
(pj42(n, k))2,
(32)
where N is the number of pixels included in the region.In contrast, in the iterated PUPa structure the Pj242 and
Hj filters at level j do not depend on the adaptation at thelower decomposition levels. The p parameters at level j arecalculated by using (30), but the numerator and denominator
biases are independent of the pixel position:
bias(qj(n)) = N 2wk=l
hj(k)pj242(l), (33)
bias(rj(n)) = N 2wk
(pj242(k))2. (34)
IV. THE ICI RULE
So far, we have not discussed a way of determining the
neighborhoods of similar statistical properties for every pixel
of the image.
Let us suppose that a compact representation of an analyzed
image can be obtained by setting the appropriate values of
the FB parameter p in distinct and appropriate regions of theimage. In the simplest setting, the adaptation of parameter pcan be performed on a fixed-size region. It is important to
note that wider adaptation regions lead to smaller varianceof p. So, in order to obtain more reliable estimates we wantto have parameter p calculated on a region that is as big aspossible, but still not incorporating parts of the image with
different statistical properties. To obtain such regions, we use
a statistical method based on the intersection of confidence
intervals (ICI) rule [28], [33], which is applied for every pixel
in order to obtain the most appropriate adaptation region for
it.
The ICI rule is based on a number of estimates of some
signal-related parameter. The estimates are based on different
window sizes. We define a set of growing window sizes:
H
= {hk|hk > hk1, k = 1, 2, . . . , K }. (35)Window hk1 is a subset of the next window hk. The Kestimates are calculated for K successive windows. Next, wedefine confidence intervals of the estimate xk as
Ck = [pc,k(n) p, pc,k(n) + p], (36)
where p = std(pc,k(n)), and is an empirically-set constantthat defines the width of the confidence interval.
The ICI algorithm starts with the smallest window size
and calculates the appropriate confidence interval C1. Thenthe second estimate is calculated, and its confidence interval
C2 is obtained. The intersection of confidence intervals for ksuccessive estimates is
Ik = k
i=1Ci. (37)One wants to obtain an estimate that is based on the biggest
possible window size but still statistically related to the previ-
ous estimates. In terms of confidence intervals this means that
the closest-to-optimal scale, which will be denoted as k+, willbe the largest of indices for which Ik = . Therefore, the ICIalgorithm consists of the following steps:
First estimate. Obtain the estimate pc for the smallest window.Calculate L1 and U1 as the lower and the upper bounds of thefirst confidence interval obtained from (36).
Successive estimates. Calculate successive estimates based
on the growing windows. Track the value of the highest
-
8/7/2019 Adaptive 2-D Wavelet Transform Based on the_Lifting Scheme with Preserved Vanishing Moments_
11/17
-
8/7/2019 Adaptive 2-D Wavelet Transform Based on the_Lifting Scheme with Preserved Vanishing Moments_
12/17
Copyright (c) 2010 IEEE. Personal use is permitted. For any other purposes, Permission must be obtained from the IEEE by emailing [email protected].
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON IMAGE PROCESSING, MIROSLAV VRANKIC, DAMIR SERSIC, AND VICTOR SUCIC 12
(a)
pc
=1.099
(b)
pc
=1.12
(c)
Fig. 16. (a) The test SineCircle image. The test image was additionally corrupted with Gaussian white noise with w = 10. (b) Adaptive wedge-shapedregion and (c) the overall region obtained for the PUPa structure with 95% confidence level are shown by black dots. The origin pixel is shown by the whitecircle. The regions are plotted over the noisy test image to outline the true region borders. True values ofp (obtained for the original image without noise)are 1.111 and 1.742 for the inner and outer sine waves respectively.
(a) (b) (c)
Fig. 17. (a) Part of the Barbara image and (b) the corresponding detail coefficients obtained for the first decomposition level of the fixed P2U2 structure,and (c) for the adaptive PUPa structure with 5 5 window adaptation.
As mentioned earlier, for denoising purposes we have used
the undecimated FB. The denoising results are given for 4-
level adaptive wavelet decomposition based on the quincunx
PUPa filter bank with 2 dual and 2 primal vanishing moments.
Therefore, in each decomposition level, fixed P2 and U2 stagesare applied, followed by the adaptive P42 branch in which thevalue of parameter p is calculated based on the 5x5 pixelswide neighborhood of each pixel. Since undecimated FB is
employed, we have used j-times quincunx-upsampled predictand update filters, where j stands for the decomposition level.The synthetic test image used was the 512x512 SineCircle
image. The sine wave inside the circular region has period
T1 = 9 and angular orientation 1 = /8, while for the outersine wave T2 = 5 and 2 = /4.
A part of the noisy SineCircle image, and the images
obtained with the three denoising methods are shown in
Fig. 19. The resulting detail coefficients obtained with the
adapted p parameters are shown in Fig. 19(c). When comparedto the detail coefficients obtained with nonadaptive wavelet
decomposition (Fig. 19(b)), it is obvious that the adaptation
introduced significant improvements, setting the detail coef-
ficients to zero in the whole image except at the circular
boundary of the two sine-wave patterns.
Results from Figs. 19(d) and 19(e) show that the adaptation
introduced in the FB significantly improves denoising results.
The adaptation transferred the information on the harmonic
components to the filter parameters, and the noise present
in the image was more efficiently suppressed by the wavelet
thresholding operation. Fig. 19(f) shows the result obtained
by using the state-of-the-art pointwise shape-adaptive DCT
method (SA-DCT) [19]. It is clearly visible that our method
more accurately restores the periodic pattern in the image.
Note that the SA-DCT method was used for comparison since
it also uses the ICI rule to obtain appropriate adaptation
regions.
The obtained denoising results are summarized in Table
II. For comparison purposes we have added denosing results
obtained with BLS-GSM denoising method [18] and wavelet-
domain Wiener filtering (WDWF) [35], which uses separable
db2 and sym6 wavelets. For the synthetic SineCircle and
Ripples image, the improvement in terms of peak signal-to-
noise ratio (PSNR) when using adaptive wavelets is signifi-
cant, hence indicating that the proposed method would be of
particular interest in the analysis of images with dominant
periodic components. All the test images are available at
http://www.riteh.hr/mvrankic/images.
-
8/7/2019 Adaptive 2-D Wavelet Transform Based on the_Lifting Scheme with Preserved Vanishing Moments_
13/17
Copyright (c) 2010 IEEE. Personal use is permitted. For any other purposes, Permission must be obtained from the IEEE by emailing [email protected].
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON IMAGE PROCESSING, MIROSLAV VRANKIC, DAMIR SERSIC, AND VICTOR SUCIC 13
(a) (b) (c)
(d) (e) (f)
Fig. 19. (a) Part of the noisy SineCircle image with w = 10 . (b) Corresponding detail coefficients for the first decomposition level of the unadapted P2U2structure, and (c) for the PUPa structure with adaptation using a fixed 5 5 window. Denoising results for (d) 5 5 window-based adaptation, (e) ICI-basedadaptation, and (f) the result of the point-wise SA-DCT method from [19].
VI. CONCLUSIONS
In this paper, we have presented a novel design technique forlocally adaptive wavelet decompositions based on an adaptive
lifting scheme structure. The main purpose of the adaptation
was to obtain a more compact representation of an analyzed
image when compared to the nonadaptive wavelet transform.
The adaptation involves a modification of the predict stage
in the lifting scheme, and two new schemes, named PPaU and
PUPa have been proposed, both of which retain a desirable
number of dual and primal vanishing moments. The main
advantage of the PUPa structure over the PPaU structure
is that the adaptation at one level does not influence the
following levels, leading to its computationally less demanding
implementation.
The adaptation of the filter bank parameters ensured theminimization of the squared prediction error on the neighbor-
hood of a given pixel, which then led to the minimization of
the energy of the wavelet detail coefficients. The influence of
the additive zero-mean Gaussian noise on the estimation of the
adaptive filter parameter p has been also studied. To reducethe bias of p we have proposed the improved estimator pc,whose bias is significantly reduced.
In order to obtain an appropriate adaptation region for every
pixel, we employed a statistical method called the intersection
of confidence intervals (ICI) rule. The ICI method gives re-
gions of similar statistical properties for estimating parameter
p. While it is computationally demanding, the ICI methodsignificantly improves the estimation of parameter p. Wehave shown that, for synthetic images composed of localized
periodic components, the proposed adaptive wavelet FBs give
high-quality image denoising results, even outperforming the
current state-of-the-art pointwise SA-DCT method.
APPENDIX A
PRESERVATION O F VANISHING MOMENTS IN THE PPAU
STRUCTURE
The analysis polyphase matrix for the general lifting scheme
structure from Fig. 1 can be written as:
Hp =
1 U
0 1
1 0
P 1
. (38)
Therefore: H0e H0o
H1e H1o
=
1 U P U
P 1
, (39)
where H0e, H1e, H0o, H1o respectively denote even and oddpolyphase components of the low-pass and high-pass H0 andH1 filters.
The dual vanishing moments (DM) condition requires that
[29]
( D)H1(Z2) = 0, for N, (40)
where N denotes the space of all polynomial sequences oftotal degree strictly less than N. By using the filter polyphase
-
8/7/2019 Adaptive 2-D Wavelet Transform Based on the_Lifting Scheme with Preserved Vanishing Moments_
14/17
Copyright (c) 2010 IEEE. Personal use is permitted. For any other purposes, Permission must be obtained from the IEEE by emailing [email protected].
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON IMAGE PROCESSING, MIROSLAV VRANKIC, DAMIR SERSIC, AND VICTOR SUCIC 14
TABLE IIDENOISING RESULTS IN TERMS OF PSNR. P2U2 STANDS FOR NONADAPTED QUINCUNX WAVELETS WITH 2 DUAL AND 2 PRIMAL VANISHING MOMENTS.
Image name P2U2 PUPa BLS-GSM Pointwise SA-DCT WDWF
SineCircle 5 34.35 41.97 40.40 36.03 36.66
10 28.43 36.46 35.30 31.37 32.15
20 22.74 30.40 30.05 26.75 27.53
40 17.10 21.98 23.78 23.54 20.99
Ripples 5 35.20 43.29 41.27 35.32 38.03
10 29.02 37.36 36.49 30.22 33.12
20 24.12 31.74 31.27 26.04 28.98
40 17.23 25.93 25.74 22.42 25.27
Barbara 5 35.48 35.73 37.60 37.42 37.28
10 30.26 30.98 33.51 33.58 32.87
20 25.87 26.68 29.22 29.88 28.24
40 20.34 23.45 25.41 26.25 24.33
Textile 5 33.01 34.12 33.95 25.34 34.09
10 27.93 28.21 28.54 24.02 28.12
20 22.14 22.57 23.35 21.49 22.56
40 16.05 18.58 18.88 18.42 18.15
Brick wall 5 33.05 34.14 33.98 27.11 34.18
10 27.89 28.25 28.74 25.19 28.30
20 22.08 22.96 23.67 22.24 22.91
40 16.35 19.21 19.60 19.38 18.87
0 5 10 15 20
-2
-1
0
1
2
3
(a)
0 5 10 15 20
0
0.5
1
1.5
2
(b)
Fig. 15. (a) The ICI graph and (b) its magnified view for pc calculatedfor w = 10 and 95% confidence level, resulting in the final wedge-shapedregion in Fig. 16(b). The dashed horizonatal line in (b) represents the truevalue ofp (which is unknown to the ICI algorithm).
representation, the output of the analysis high pass filter can
be obtained with each polyphase component affecting one
(a)
1.8 2 2.2 2.4 2.6 2.80
5
10
15
20
25
(b)
1.8 2 2.2 2.4 2.6 2.80
5
10
15
20
25
(c)
Fig. 18. The histograms of pc parameters for the part of Barbaras scarf,noisy with w = 10, which is shown in (a), for the first decompositionlevel of the undecimated PUPa structure. The histograms are shown for (b)adaptation using a fixed 5 5 window and (c) the ICI-based adaptation with95% confidence level.
sublattice ofZ2, which transforms the DM condition to
H1e(DZ2
) + H1o(DZ2
+ t) = 0. (41)By using (39), the DM condition becomes
P (DZ2) + (DZ2 + t) = 0, (42)
which yields
P (DZ2) = (DZ2 + t). (43)
As given in [29], (43) can be expressed in the downsampled
domain as
P (Z2) = (Z2 +D1t), for N. (44)
Now, introducing the overall prediction filter from Fig. 3,
which consists of the fixed part PN and the adaptive part pPa,
-
8/7/2019 Adaptive 2-D Wavelet Transform Based on the_Lifting Scheme with Preserved Vanishing Moments_
15/17
Copyright (c) 2010 IEEE. Personal use is permitted. For any other purposes, Permission must be obtained from the IEEE by emailing [email protected].
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON IMAGE PROCESSING, MIROSLAV VRANKIC, DAMIR SERSIC, AND VICTOR SUCIC 15
we obtain(PN + pPa)(Z
2) = (Z2 +D1t),
PN(Z2) + pPa(Z
2) = (Z2 +D1t).(45)
Since PN is a Neville filter of order N and shift = D1t,
we have
PN(Z2) = (Z2 +D1t), for N, (46)
which combined with (45) gives
(Z2 +D1t) + pPa(Z2) = (Z2 +D1t), (47)
which further simplifies the DM condition to
pPa(Z2) = 0, for N. (48)
So, in order to keep N dual vanishing moments introducedwith PN, the filter Pa must cancel polynomials of degree lowerthan N. As stated in Section II-C, we propose filter Pa suchthat
Pa = PR PS , (49)
where PR and PS are Neville filters of shift = D1t and
orders
R N, S N, (50)
respectively. Applying the filter to a polynomial sequence N gives
p(PR PS)(Z2) = 0, (51)
since
PR(Z2) = PS(Z
2) = (Z2 +D1t). (52)
Therefore, with the choice of the Pa filter from (49), the Ndual vanishing moments are guaranteed no matter what values
the p parameter may take.Now, we will show that for the PPaU filter bank the number
of primal vanishing moments N that is fixed with the updatefilter is not influenced by the adaptive prediction branch as
long as N N
. The synthesis polyphase matrix follows
directly from the analysis polyphase matrix as Gp = H1p
Therefore,G0e G0o
G1e G1o
=
1 P
N+
UN 1 UNP
N+
, (53)
where G0 and G1 denote synthesis low-pass and high-passfilters respectively, and PN+ is the overall prediction filter asgiven in (8). The primal vanishing moments (PM) condition
states that in order to have N vanishing moments of thesynthesis FB, the following relation must be satisfied
G1( D) = 0 for N, (54)
where N denotes the space of all polynomial sequences of
total degree strictly less than N. Stated in terms of polyphasecomponents, the PM condition becomes
G1e(DZ2) + G1o(DZ
2 + t) = 0, (55)
which combined with (53) gives
UN(DZ2) + (1 UNP
N+)(DZ2 + t) = 0, (56)
UN(DZ2) + UNP
N+(DZ2 + t) = (DZ2 + t). (57)
IfN N, R N, and S N, thenPN+
(DZ2 + t) = (PN
+ p(PR
PS
))(DZ2 + t)
= PN
(DZ2 + t)
= (DZ2),
(58)
since
p(PR
PS
)(DZ2 + t) = 0. (59)
Therefore, the PM condition from (57) becomes
2UN(DZ2) = (DZ2 + t), for N, (60)
which is always true, since 2UN = PN, as stated in (4).
APPENDIX B
PRESERVATION O F VANISHING MOMENTS IN THE PUPA
STRUCTURE
The polyphase matrix of the analysis PUPa filter bank from
Fig. 4 is
Hp =
1 0
pPa 1
1 UN
0 1
1 0
PN 1
= 1 PNUN UNpPa(1 PNUN) PN pPaUN + 1
,(61)
where filter Pa is given as in (49). Therefore, the polyphasecomponents of the analysis high-pass filter are
H1e = pPa(1 PNUN) PN,
H1o = pPaUN + 1.(62)
The DM condition can be stated in the polyphase form as
H1e(DZ2) + H1o(DZ
2 + t) = 0, for N. (63)
Since, as stated in (48), it holds that
Pa(Z2) = 0, for N, (64)
using (62) the DM condition from (63) becomes
PN(DZ2) = (DZ2 + t), for N, (65)
which is always true, and therefore the N dual vanishingmoments are guaranteed for any value of parameter p.
The synthesis polyphase matrix of the PUPa filter bank from
Fig. 4 can be obtained from the analysis polyphase matrix (61)
as
Gp =
pPaUN + 1 pPa(1 P
NUN) + P
N
UN 1 P
NUN
. (66)
Therefore, the polyphase components of the synthesis high-
pass filter G1 are
G1e = UN,
G1o = 1 P
NUN,
(67)
which are the same as those for the PPaU structure; so,based on the proof in Section A, we conclude that N primalvanishing moments are preserved as long as N N.
APPENDIX C
NUMERATOR AND DENOMINATOR BIAS FOR THE
ESTIMATOR OF p
The bias of the numerator q from (26) is
bias(q(n)) = E[q(n) q(n)] = q(n) E[q(n)] (68)
-
8/7/2019 Adaptive 2-D Wavelet Transform Based on the_Lifting Scheme with Preserved Vanishing Moments_
16/17
Copyright (c) 2010 IEEE. Personal use is permitted. For any other purposes, Permission must be obtained from the IEEE by emailing [email protected].
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON IMAGE PROCESSING, MIROSLAV VRANKIC, DAMIR SERSIC, AND VICTOR SUCIC 16
where
E[q(n)] = E
iRn
xh(i)xp(i)
= E
iRn
(x0h(i) + wh(i))(x0p(i) + wp(i))
= iRn
(x0h(i)x0p(i) + E[wh(i)wp(i)])
= q(n) +iRn
E[wh(i)wp(i)] ,
(69)
since E[wp(i)] = 0 and E[wh(i)] = 0. Combining (68) and(69) the bias can be expressed as
bias(q(n)) = iRn
E[wh(i)wp(i)] . (70)
Knowing that (see Fig. 11)
wp(n) = (w p42)n =2
32
5i=2
w(n, i) +1
32
13i=6
w(n, i), (71)
wh(n) = (w h)n = w(n, 1) 1
4
5n=2
w(n, i), (72)
and by using the fact that
E[w(n, k)w(n, l)] =
0 for k = l
2w for k = l, (73)
we can write
E[wh(i)wp(i)] = E
1
4
2
32
5k=2
w2(i, k)
= 1
4
2
32 42w =
1
162w.
(74)
By combining (70) and (74), the bias of q can be expressed
as
bias(q(n)) = iRn
E[wh(i)wp(i)] = N2w16
, (75)
where N is the number of pixels included in the region Rn.The bias of the denominator r from (26) is
bias(r(n)) = E[r(n) r(n)] = r(n) E[r(n)] (76)
where the expectation of r(n) is:
E[r(n)] = E
iRn
x2p(i)
= E
iRn
(x0p(i) + wp(i))2
= EiRn
x20p(i) + 2x0p(i)wp(i) + w2p(i)
=iRn
(x20p(i) + E[w2p(i)])
= r(n) +iRn
E[w2p(i)],
(77)
since from E[w(i)] = 0 it follows that E[wp(i)] = 0. Next,since wp = w p42,
E[w2p(i)] = 2w
n
p242(n) =6
2562w. (78)
By using (76), (77), and (78) the denominators bias can now
be expressed as
bias(r(n)) = iRn
E[w2p(i)] = N6
2562w, (79)
where again N is the number of pixels included in the regionRn.
REFERENCES
[1] D. L. Donoho and I. M. Johnstone, Adaptating to unknown smoothnessvia wavelet shrinkage, J. Amer. Statist. Assoc., vol. 90, no. 90, pp. 12001224, 1995.
[2] J. Kovacevic and M. Vetterli, Nonseparable two- and three-dimensionalwavelets, Signal Processing, IEEE Transactions on [see also Acoustics,Speech, and Signal Processing, IEEE Transactions on], vol. 43, no. 5,pp. 12691273, 1995.
[3] D. E. Dudgeon and R. M. Mersereau, Multidimensional Digital SignalProcessing (Prentice-Hall Signal Processing Series). Prentice Hall,1983.
[4] G. Karlsson and M. Vetterli, Theory of two-dimensional multiratefilter banks, Acoustics, Speech, and Signal Processing [see also IEEETransactions on Signal Processing], IEEE Transactions on, vol. 38,no. 6, pp. 925937, 1990.
[5] E. Viscito and J. Allebach, The analysis and design of multidimensionalFIR perfect reconstruction filter banks for arbitrary sampling lattices,
IEEE Trans. Circuits Syst., vol. 38, no. 1, pp. 2941, Jan 1991.[6] J. Kovacevic and M. Vetterli, Nonseparable multidimensional perfect
reconstruction filter banks and wavelet bases for rn, IEEE Trans. Inf.Theory, vol. 38, no. 2, pp. 533555, Mar 1992.
[7] E. Simoncelli and E. Adelson, Nonseparable extensions of quadraturemirror filters to multiple dimensions, Proceedings of the IEEE, vol. 78,no. 4, pp. 652664, 1990.
[8] R. Coifman and M. Wickerhauser, Entropy-based algorithms for bestbasis selection, IEEE Trans. Inf. Theory, vol. 38, no. 2, pp. 713718,1992.
[9] R. Claypoole, G. Davis, W. Sweldens, and R. Baraniuk, Nonlinearwavelet transforms for image coding via lifting, IEEE Trans. ImageProcess., vol. 12, no. 12, pp. 14491459, 2003.
[10] O. Gerek and A. Cetin, Adaptive polyphase subband decompositionstructures for imagecompression, IEEE Trans. on Image Process.,vol. 9, no. 10, pp. 16491660, 2000.
[11] N. Boulgouris, D. Tzovaras, and M. Strintzis, Lossless image com-
pression based on optimal prediction, adaptive lifting, and conditionalarithmetic coding, IEEE Trans. Image Process., vol. 10, no. 1, pp. 114,2001.
[12] G. Piella, B. Pesquet-Popescu, and H. Heijmans, Adaptive update liftingwith a decision rule based on derivative filters, IEEE Signal Process.Lett., vol. 9, no. 10, pp. 329332, 2002.
[13] G. Piella and H. Heijmans, Adaptive lifting schemes with perfectreconstruction, IEEE Trans. Signal Process., vol. 50, no. 7, pp. 16201630, 2002.
[14] A. Gouze, M. Antonini, M. Barlaud, and B. Macq, Design of signal-adapted multidimensional lifting scheme for lossy coding, IEEE Trans.Image Process., vol. 13, no. 12, pp. 15891603, 2004.
[15] W. Ding, F. Wu, X. Wu, S. Li, and H. Li, Adaptive directional lifting-based wavelet transform for image coding, IEEE Trans. Image Process.,vol. 16, no. 2, pp. 416427, 2007.
[16] V. Velisavljevic, B. Beferull-Lozano, M. Vetterli, and P. Dragotti, Di-rectionlets: Anisotropic multidirectional representation with separable
filtering, IEEE Trans. Image Process., vol. 15, no. 7, pp. 19161933,2006.
[17] O. Gerek and A. Cetin, A 2-D orientation-adaptive prediction filter inlifting structures for image coding, IEEE Trans. Image Process., vol. 15,no. 1, pp. 106111, 2006.
[18] J. Portilla, V. Strela, M. Wainwright, and E. Simoncelli, Image denois-ing using scale mixtures of gaussians in the wavelet domain, IEEETrans. Image Process., vol. 12, no. 11, pp. 13381351, 2003.
[19] A. Foi, V. Katkovnik, and K. Egiazarian, Pointwise shape-adaptiveDCT for high-quality denoising and deblocking of grayscale and colorimages, IEEE Trans. Image Process., vol. 16, no. 5, pp. 13951411,2007.
[20] W. Sweldens, The lifting scheme: A new philosophy in biorthogonalwavelet constructions, in Wavelet Applications in Signal and ImageProcessing III, A. F. Laine and M. Unser, Eds., vol. 2569. SPIE, 1995,pp. 6879.
-
8/7/2019 Adaptive 2-D Wavelet Transform Based on the_Lifting Scheme with Preserved Vanishing Moments_
17/17
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON IMAGE PROCESSING, MIROSLAV VRANKIC, DAMIR SERSIC, AND VICTOR SUCIC 17
[21] , The lifting scheme: A construction of second generationwavelets, SIAM J. Math. Anal., vol. 29, no. 2, pp. 511546, 1997.
[22] I. Kopriva and D. Sersic, Wavelet packets approach to blind separationof statistically dependent sources, Neurocomputing, vol. 71, no. 7-9,pp. 16421655, Mar 2008.
[23] S. Mallat, A Wavelet Tour of Signal Processing. Academic Press,September 1999.
[24] D. Sersic, Integer to integer mapping wavelet filter bank with adaptivenumber of zero moments, in Proc. IEEE International Conference onAcoustics, Speech, and Signal Processing ICASSP 00, vol. 1. IEEE,
2000, pp. 480483.[25] M. Vrankic and D. Sersic, Image denoising based on adaptive quincunx
wavelets, in Proc. IEEE 6th Workshop on Multimedia Signal Process-ing. IEEE, 2004, pp. 251254.
[26] M. Tomic, D. Sersic, and M. Vrankic, Edge-preserving adaptive waveletdenoising using ICI rule, Electronics Letters, vol. 44, no. 11, pp. 698699, 2008.
[27] J. Lerga, M. Vrankic, and V. Sucic, A signal denoising method basedon the improved ICI rule, IEEE Signal Process. Lett., vol. 15, pp. 601604, 2008.
[28] V. Katkovnik, A new method for varying adaptive bandwidth selection,IEEE Transactions on Signal Processing, vol. 47, no. 9, pp. 25672571,1999.
[29] J. Kovacevic and W. Sweldens, Wavelet families of increasing order inarbitrary dimensions, IEEE Trans. on Image Process., vol. 9, no. 3, pp.480496, 2000.
[30] C. de Boor and A. Ron, Computational aspects of polynomial interpo-
lation in several variables, Math. Comp., pp. 58:705727, 1992.[31] R. R. Coifman and D. L. Donoho, Translation-invariant de-noising,
Stanford University, Department of Statistics, Tech. Rep., 1995, waveletsand Statistics, Anestis Antoniadis, ed. Springer-Verlag Lecture Notes.
[32] B. Silverman, Density Estimation for Statistics and Data Analysis.Chapman & Hall/CRC, 1986.
[33] L. Stankovic, Performance analysis of the adaptive algorithm for bias-to-variance tradeoff, IEEE Transactions on Signal Processing, vol. 52,no. 5, pp. 12281234, 2004.
[34] I. M. Johnstone and B. W. Silverman, Wavelet threshold estimators fordata with correlated noise, J Royal Statistical Soc B, vol. 59, no. 2, pp.319351, 1997.
[35] S. Ghael, E. P. Ghael, A. M. Sayeed, and R. G. Baraniuk, Improvedwavelet denoising via empirical wiener filtering, Proceedings of theSPIE, Mathematical Imaging, pp. 389399, 1997.