Você está na página 1de 9

Discrete Shearlet Transform: faithful digitization concept and

its applications
Wang-Q Lim
Institute of Mathematics, University of Osnabrueck, 49069 Osnabrueck, Germany
ABSTRACT
Over the past years, various representation systems which sparsely approximate functions governed by anisotropic
features such as edges in images have been proposed. Alongside the theoretical development of these systems,
algorithmic realizations of the associated transforms were provided. However, one of the most common short-
comings of these frameworks is the lack of providing a unied treatment of the continuum and digital world,
i.e., allowing a digital theory to be a natural digitization of the continuum theory. Shearlets were introduced
as means to sparsely encode anisotropic singularities of multivariate data while providing a unied treatment of
the continuous and digital realm. In this paper, we introduce a discrete framework which allows a faithful digi-
tization of the continuum domain shearlet transform based on compactly supported shearlets. Finally, we show
numerical experiments demonstrating the potential of the discrete shearlet transform in several image processing
applications.
Keywords: shearlets, wavelets, directional representation, sparse approximation, image denoising, image in-
painting, image separation.
1. INTRODUCTION
One of the most important properties of wavelets is the connection between the wavelet transform in the contin-
uum domain and lter banks in the digital domain. The connection was set up by the multiresolution analysis.
Such a connection allows the wavelet transform, initially dened in the continuum domain, to be computed by
a faithful digitization of the continuum theory, which provides a theoretical foundation for the discrete wavelet
transform. However, multivariate functions are typically governed by anisotropic phenomena. For instance,
edges in digital images are typically located along smooth curves owing to smooth boundaries of physical ob-
jects. Although wavelets are good at isolating the discontinuities at edge points, isotropic support of wavelets
does not t into eciently capturing anisotropic features of edges. In addition, wavelets can detect only limited
directional information.
In order to overcome these major drawbacks of wavelets, several representation systems have been introduced.
Among those systems, the shearlet system is in fact the only one which allows a unied treatment of the continuum
and digital world like the discrete wavelet transform while achieving optimally sparse approximations of piecewise
smooth functions.
1
In this paper, we introduce the newly designed discrete shearlet transform based on such a unied treatment
and compactly supported shearlets generated by non-separable shearlet generator. The paper is organized as
follows. In Section 2, we briey overview the compactly supported shearlet system. In Section 3, we describe
the discrete shearlet transforms based on a faithful digitization of the continuum domain shearlet transforms.
In Section 4, numerical experiments are presented and discussed to show the potential of our proposed shearlet
transform in various applications.
Further author information: (Send correspondence to Wang-Q Lim)
Wang-Q Lim: E-mail: wlim@uos.de, Telephone: +49 541 969 3806
2. COMPACTLY SUPPORTED SHEARLETS
In this section, we overview constructions of compactly supported shearlets. In particular, we show a large class
of functions including compactly supported functions such that any function in such a class can generate the
shearlet system which forms a frame for L
2
(R
2
).
Let us rst dene the shearlet system in 2D.
Definition 2.1. Dene
A
2
j =

2
j
0
0 2
j/2

,

A
2
j =

2
j/2
0
0 2
j

and S
k
=

1 k
0 1

.
Also, for some sampling vector c = (c
1
, c
2
) (R
+
)
2
, let M
c
= diag(c
1
, c
2
) and

M
c
= diag(c
2
, c
1
). Then, the
(cone-adapted) regular discrete shearlet system oH(c; , ,

) generated by a scaling function L
2
(R
2
) and
shearlets ,

L
2
(R
2
) is dened by
oH(c; , ,

) = (c
1
, ) (c, )

(c,

),
where
(c
1
, ) =
m
= ( c
1
m) : m Z
2
,
(c, ) =
j,k,m
= 2
3j/4
(S
k
A
2
j M
c
m) : j 0, [k[ 2
j/2
|, m Z
2
,
and

(c,

) =

j,k,m
= 2
3j/4

(S
T
k

A
2
j

M
c
m) : j 0, [k[ 2
j/2
|, m Z
2
.
One key property of the shearlet system oH(c; , ,

) is that directionality is achieved by shear matrix S
k
which not only provides directionality like rotation but also leaves the integer grid Z
2
invariant. In fact, this
invariance property leads to a unied digitization concept for the shearlet transform between continuum and
digital world. The following theorem shows sucient conditions for the shearlet generators and

to generate
the shearlet system.
2
Theorem 2.2. Let , L
2
(R
2
) be such that

(
1
,
2
) C
1
min1, [
1
[

min1, [
2
[

and
[

(
1
,
2
)[ C
2
min1, [
1
[

min1, [
1
[

min1, [
2
[

, (1)
for some positive constants C
1
, C
2
< and > > 3. Dene

(x
1
, x
2
) = (x
2
, x
1
). We also assume that
there exists a positive constant A > 0 such that
ess inf

()[
2
+

j0

|k|2
j/2

(S
T
k
A
2
j )[
2
+

j0

|k|2
j/2

(S
k

A
2
j )[
2
> A.
Then there exists a sampling parameter c = (c
1
, c
2
) (R
+
)
2
such that the shearlet system oH(c; , ,

) forms
a frame for L
2
(R
2
).
It can be easily veried that the conditions imposed on and by Theorem 2.2 are satised by many suitably
chosen functions which can be compactly supported in the spatial domain.
2
3. DISCRETE SHEARLET TRANSFORM
We will now discuss implementation strategy for computing shearlet coecients associated with the discrete
shearlet system based on compactly supported shearlets, as introduced in the previous section. Again, one
main focus will be on deriving a digitization which is faithful to the continuum setting. We will only consider
computing shearlet coecients f,
j,k,m
) associated with shearlets
j,k,m
belonging to (c, ). Notice that
shearlet coecients associated with shearlets

j,k,m


(c,

) are computed in the same way except for switching
the order of variables.
Let us start with the following simple equation:

j,k,m
() = 2
3j/4
(A
2
j (S
2
j/2
k
()) M
c
m).
From this, we see that each shearlet element is obtained by applying shearing S
2
j/2
k
to wavelet
j,0,m
associated
with anisotropic scaling matrix A
2
j . Our approach will be based on the following digitization strategies:
Digitization of wavelets
j,0,m
: A multiresolution analysis associated with anisotropic scaling A
2
j can be
applied.
Digitization of shearing S
2
j/2
k
: A faithful digitization of shear operator S
2
j/2
k
has to be achieved with
particular care.
We now consider a shearlet generator as follow. Let
1
L
2
(R) be a compactly supported 1D scaling
function satisfying

1
(x
1
) =

n1Z
h(n
1
)

2
1
(2x
1
n
1
) (2)
for some appropriately chosen lter h. An associated compactly supported 1D wavelet
1
L
2
(R) can then be
dened by

1
(x
1
) =

n1Z
g(n
1
)

2
1
(2x
1
n
1
), (3)
where again g is an appropriately chosen lter. The selected shearlet generator is then dened to be
(x
1
, x
2
) =
1
(x
1
)
1
(x
2
), (4)
and the scaling function by (x
1
, x
2
) =
1
(x
1
)
1
(x
2
). Here, we choose the lter coecients h and g in (2)(3)
so that the shearlet generator in (4) satises decay condition (1). For the signal f L
2
(R
2
) to be analyzed,
we assume that, for J > 0 xed, f is of the form
f(x) =

nZ
2
f
J
(n)2
J
(2
J
x
1
n
1
, 2
J
x
2
n
2
). (5)
Our task is now to compute shearlet coecients f,
j,k,m
) for scaling coecients f
J
. In order to digitize shear
operator S
2
j/2
k
for the coecients f
J
dened on the integer grid Z
2
, we observe that
S
2
j/2
k
(2
j/2
Z Z) = 2
j/2
Z Z. (6)
Here, we assume j/2 is integer, otherwise j/2| would need to be taken. From (6), we see that rened integer
grid 2
j/2
ZZ is invariant under S
2
j/2
k
. This leads to discrete shear operator S
d
2
j/2
k
which computes discrete
sample values S
d
2
j/2
k
(f
J
)(n) on sheared integer grid S
2
j/2
k
(Z
2
) with essentially no aliasing eect.
4
To analyze the relation between a shear matrix S
2
j/2
k
and the associated digital shear operator S
d
2
j/2
k
, let
us consider the following simple example: Set f
c
=
{x:x1=0}
. Then digitize f
c
to obtain a function f
d
dened
on Z
2
by setting f
d
(n) = f
c
(n) for all n Z
2
. For xed shear parameter s R, apply the shear transform S
s
to f
c
yielding the sheared function f
c
(S
s
()). Next, digitize also this function by considering f
c
(S
s
())[
Z
2 . The
functions f
d
and f
c
(S
s
())[
Z
2 are illustrated in Fig. 1 (a)-(b) for s = 1/8. Note that the integer lattice is not
invariant under the shear matrix S
1/8
. This prevents the sampling points S
1/8
(n), n Z
2
from lying on the
integer grid, which causes aliasing of the digitized image f
c
(S
1/8
())[
Z
2 as illustrated in Fig. 1 (e). In order
to avoid this aliasing eect, the grid needs to be rened by a factor of 8 along the horizontal axis followed by
computing sample values on this rened grid.
More generally, when the shear parameter is given by s = 2
j/2
k, one can essentially avoid this directional
aliasing eect by rening a grid by a factor of 2
j/2
along the horizontal axis followed by computing interpolated
sample values on this rened grid. We refer to
4
for more details.


(a) (b) (c)
(d) (e) (f)
Figure 1. (a) Original image f
d
(n). (b) Sheared image fc(S
1/8
n) (c) De-aliased image: S
d
1/8
(f
d
)(n). (d) Original image:
DFT of f
d
(n). (e) Aliased image: DFT of fc(S
1/8
(n)). (f) De-aliased image: DFT of S
d
1/8
(f
d
)(n).
Next, dene a linear operator W
j1,j2
:
2
(Z
2
)
2
(Z
2
) by the wavelet transform associated with anisotropic
scaling matrix diag(2
j1
, 2
j2
). We refer to
3, 4
for the precise denitions of S
d
2
j/2
k
and W
j1,j2
. Now, let w
j
(n) be
wavelet lter coecients associated with the anisotropic wavelet transform W
Jj,Jj/2
for each j 0. Then we
have the following exact computation of shearlet coecients f,
j,k,m
):
3
Theorem 3.1. Let f L
2
(R
2
) given as in (5). Retaining the notations and denitions from this subsection, we
obtain
f,
j,k,m
) = (w
j
S
d
2
j/2
k
(f
J
))(2
Jj
c
1
m
1
, 2
Jj/2
c
2
m
2
) for j = 0, . . . , J 1. (7)
There are various functions satisfying the conditions in Theorem 2.2. In particular, one may choose non-
separable shearlet generator
non
rather than separable one as described above. Dene

non
() = P(
1
, 2
2
)

(),
where the trigonometric polynomial P is a 2D fan lter.
5
This denes shearlets
non
j,k,m
generated by non-separable
generator function
non
by setting

non
j,k,m
(x) = 2
3
4
j

non
(S
k
A
2
j x M
cj
m),
where M
cj
is a sampling matrix given by M
cj
= diag(c
j
1
, c
j
2
) and c
j
1
and c
j
2
are sampling constants for translation.
One major advantage of these shearlets
non
j,k,m
is the fact that a fan lter enables renement of the directional
selectivity in frequency domain at each scale. Fig. 2(a)-(b) show the rened essential support of

non
j,k,m
as
compared to shearlets
j,k,m
arising from a separable generator .
(a) (b)
Figure 2. (a) Separable shearlet

in frequency (b) Non-separable shearlet

non
. in frequency
We now apply the same implementation strategy as the discrete shearlet transform associated with separable
shearlet generator to discretize non-separable shearlets
non
j,k,m
. Combining the discrete shear operator S
d
2
j/2
k
with the wavelet lter w
j
, we obtain digitized shearlet
d
j,k
as follows.

d
j,k
(n) = S
d
2
j/2
k

p
Jj/2
w
j

(n),
where p
Jj/2
(n) are the Fourier coecients of the 2D fan lter given by P(2
Jj

1
, 2
Jj/2+1

2
). Note that the
lter
d
j,0
=

p
Jj/2
w
j

digitizes wavelets
non
j,0,m
based on the multiresolution analysis and the directional lter

d
j,k
is obtained by applying the discrete shear operator S
d
2
j/2
k
to the lter
d
j,0
. Fig. 3 shows the frequency
response of the digital lters
d
j,0
and
d
j,k
.
(a) (b)
Figure 3. (a) Frequency response of
d
j,0
at the nest scale when j = 4. (b) Frequency response of
d
j,k
at the nest scale
when j = 4, k = 1.
We now dene the discrete non-separable shearlet transform (DNST) associated with the non-separable
shearlet generator
non
by
DNST
j,k
(f
J
)(m) = (f
J

d
j,k
)(2
Jj
c
j
1
m
1
, 2
Jj/2
c
j
2
m
2
), for f
J

2
(Z
2
),
where
d
j,k
(n
1
, n
2
) =
d
j,k
(n
1
, n
2
).
For the inverse of DNST, one may apply the conjugate gradient method which requires computing DNST and
its adjoint in each iteration step. The convergence of this iterative scheme is guaranteed due to frame property
of shearlets
non
j,k,m
. Also, by setting c
j
1
= 2
jJ
and c
j
2
= 2
j/2J
, the DNST simply becomes a 2D convolution.
Thus, in this case, DNST is shift invariant and dual discrete shearlet lters

d
j,k
can be explicitly computed by

d
j,k
() =

d
j,k
()

j,k
[

d
j,k
()[
2
and we obtain the reconstruction formula
f
J
=

j,k
(f
J

d
j,k
)

d
j,k
.
For more details about the DNST, we refer to.
4
4. APPLICATIONS
In previous section, we have described discrete implementation strategy for computing shearlet coecients based
on the multiresolution analysis and discrete shear operator S
d
2
j/2
k
. In this section, we show various applica-
tions of this discrete framework in order to illustrate the potential of the discrete shearlet transform previously
described.
4.1 Image Denoising
We compare the performance of the DNST to the non-subsampled contourlet transform (NSCT)
5
which is
another type of directional transform based on compactly supported directional elements in image denoising.
For this, we perform hard threshold on the subband coecients of the both transforms. We choose the threshold
T = K
j

for each scale. We set K
j
= 4 for the nest scale and K
j
= 3 for the remaining ones. We used ve scales of
decomposition for the NSCT and four scales for the DNST. For the NSCT, we used 4,8,8,16,16 directions in the
scales from coarser to ner while we used 8,8,16,16 directions for the DNST. Fig. 4 shows the PSNR results and
the running time for the both transforms running time of MATLAB implementations for the both transforms
on desktop with 2.53GHz, 1.98GB of RAM. As shown in Fig. 4, the DNST oers a better recovery of noisy
image in terms of visual quality as well as PSNR.
(a) (b)
Figure 4. Image Denoising with the DNST and hard thresholding. The noisy intensity is 30 (18.59dB). (a) Denoised with
NSCT, PSNR = 30.46dB (565.05 sec) (b) Denoised with DNST, PSNR = 30.87dB (4.52 sec).

NSCT toolbox is available from http://www.ifp.illinois.edu/


~
minhdo/software/.
4.2 Image Separation
The task of decomposing a signal into its morphologically dierent components is of great interest in many ap-
plications. In neurobiological imaging, it would, for instance, be desirable to separate spines (pointlike objects)
from dendrites (curvelike objects) in order to analyze them independently aiming to detect characteristics of
Alzheimer disease.
Among various approaches for this data separation problem, Morphological Component Analysis (MCA) has
been recently proposed in.
6
The basic idea of this approach is to choose two frames
1
and
2
adapted to the
two components to be separated in such a way that the two frames provide a sparse representation for each
of the components. Searching for the sparsest representation of the signal in the combined dictionary [
1
[
2
]
would provide desired separation. For the separation of point- and curvelike features, it was in fact proven in
8
that
1
minimization solves this task exploring a combined dictionary of wavelets and curvelets. An associated
algorithmic approach using the undecimated discrete wavelet transform (UDWT) and the fast discrete curvelet
transform (FDCT)

has been implemented in MCALab

.
In this section, we discuss image separation using wavelets and shearlets to separate point- and curvelike
objects from single image. For this, we compare the performance of our separation scheme presented in
9
to
MCALab. In this comparison test, we choose an articial image composed of two subimages P and C, where P
solely contains pointlike structures and C contains curvelike structures. We further add white Gaussian noise
with = 20. For our separation scheme, we applied the same algorithm as MCALab to approximately minimize
the
1
norm of shearlet and wavelet coecients. For a combined dictionary of wavelets and shearlets, we used
the DNST with 8, 8, 16, 16 directions and the UDWT. In Fig. 6, it can be observed that some part of curve
is missed and appears in the pointlike part. This deciency becomes worse with growing curvature. In contrast
to this, compactly supported shearlets oer much better spatial localization than band-limited curvelets, which
positively aects the capturing of localized features of the curve.
Figure 5. Noisy image: points + curves
4.3 Image Inpainting
Image inpainting refers to the lling-in of missing data in digital images based on the information available in
the observed region. Applications of this technique include lm restoration, text or scratch removal, and digital
zooming.
We nally show the numerical test results of image inpainting. For this, we compare our inpainting scheme
using the DNST with 8, 8, 16, 16 directions to MCALab using a combined dictionary of local cosine basis and
curvelets to provide sparse representations for the texture and the cartoon respectively. The both inpainting
schemes are based on MCA for image inpainting
7
and the same algorithm is applied to approximately minimize
the
1
norm of the expansion coecients. Although a single dictionary of shearlets is used in our scheme, it
outperforms MCALab in terms of PSNR as well as visual quality as shown in Fig. 7.

CurveLab (Version 2.1.2) is available from http://www.curvelet.org.

MCALab (Version 120) is available from http://jstarck.free.fr/jstarck/Home.html.


(a) (b)
(c) (d)
Figure 6. Image Separation with UDWT + DNST. (a) MCALab (UDWT+FDCT): pointlike component. (b) MCALab
(UDWT+FDCT): curvelike component. (c) Our scheme: pointlike component. (d) Our scheme: curvelike component.
5. CONCLUSION
In this paper, we have introduced a implementation strategy for explicitly computing shearlet coecients. Our
approach is based on the multiresolution analysis and a faithful digitization of shear operator S
k
allowing a
unied treatment between continuum and digital world. We have shown various applications of the discrete
shearlet transform based on this approach to illustrate the potential of our proposed implementation scheme. A
Matlab shearlet toolbox is freely available for download from www.shearlab.org.
ACKNOWLEDGMENTS
The author acknowledges support from DFG Grant SPP-1324, KU 1446/13.
REFERENCES
[1] G. Kutyniok and W.-Q Lim, Compactly supported shearlets are optimally sparse, J. Approx. Theory, to
appear.
[2] P. Kittipoom, G. Kutyniok, and W.-Q Lim, Construction of compactly supported shearlet frames, Constr.
Approx., to appear.
[3] W.-Q Lim, The discrete shearlet transform: a new directional transform and compactly supported shearlet
frames, IEEE Trans. Image Proc., 19, 1166-1180, (2010).
[4] W.-Q Lim, Non-separable shearlet transforms, preprint.
[5] A. L. Cunha, J. Zhou, and M. N. Do, The nonsubsampled contourlet transform: theory, design, and
applications, IEEE Trans. Image Proc., 15, 3089-3101, (2006).
[6] J.-L Starck, M. Elad, and D. Donoho, Image decomposition via the combination of sparse representation
and a variational approach, IEEE Trans. Image Proc., 14, 1570-1582, (2005).
(a) (b)
(c) (d)
Figure 7. Image Inpainting with the DNST. (a) Original Barbara image. (b) Barbara image with 80% missing pixels
(c) Inpainted image with the local cosine transform + FDCT (26.38dB, 1558.86 sec). (d) Inpainted image with DNST
(27.64dB, 530.90 sec).
[7] M. Elad, J.-L. Starck, D. Donoho and P. Querre, Simultaneous cartoon and texture image inpainting using
morphological component analysis (MCA), Appl. Comput. Harmon. Anal., 19, 340-358, (2005).
[8] D. L. Donoho and G. Kutyniok, Microlocal analysis of the geometric separation problem, preprint.
[9] G. Kutyniok and Wang-Q Lim, Image separation using wavelets and shearlets, Curves and Surfaces
(Avignon, France, 2010), Lecture Notes in Computer Science, Springer, to appear.

Você também pode gostar