672 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 11, NO. 6, JUNE 2002
. In contrast, in analyzing objects with edges, wavelets
give an
-term squared approximation error only of size ,
and wavelet thresholding gives a corresponding MSE only of
size
and no better.
D. This Paper
So according to theory for a certain continuous-space model,
discrete ridgelet transforms and discrete curvelet transforms
provide near-ideal sparsity of representation of both smooth
objects and of objects with edges. In a certain continuous-space
statistical theory, this implies that simple thresholding of noisy
coefficients in these expansions is a near-optimal method of
denoising in the presence of white Gaussian noise.
In this paper we provide an initial test of these ideas in a dig-
ital image processing setting, where images are available on an
-by- grid. We first review some basic ideas about ridgelet and
curvelet representations in the continuum. We next use these
to develop a series of digital ridgelet and digital curvelet trans-
forms taking digital input data on a Cartesian grid. Next we con-
sider a model denoising problem where we embed some stan-
dard images in white noise and apply thresholding in the digital
curvelet transform domain. Finally we discuss interpretations
and possibilities for future work.
Not surprisingly, other researchers have undertaken efforts to
implement ridgelet and curvelet transforms, and develop appli-
cations. In addition to work mentioned in the body of the article,
we would like to point out the work of Do and Vetterli [13],
Donoho and Duncan [18]. We would also like to mention the
articles of Sahiner and Yagle [21]–[23], Olson and DeStefano
[20], Zhao et al. [30], and Zuidwijk [31], [32] although these
references are less directly related.
II. C
ONTINUOUS RIDGELET TRANSFORM
The 2-D continuous ridgelet transform in can be defined
as follows [2]. We pick a smooth univariate function
with sufficient decay and satisfying the admissibility condition
(1)
which holds if, say,
has a vanishing mean .We
will suppose that
is normalized so that .
For each
, each and each , we define
the bivariate ridgelet
by
(2)
this function is constant along lines
const
.
Transverse to these ridges it is a wavelet. Given an integrable
bivariate function
, we define its ridgelet coefficients by
We have the exact reconstruction formula
(3)
valid a.e. for functions which are both integrable and square in-
tegrable. Furthermore, this formula is stable as one has a Par-
seval relation
(4)
Hence, much like the wavelet or Fourier transforms, the identity
(3) expresses the fact that one can represent any arbitrary func-
tion as a continuous superposition of ridgelets. Discrete analogs
of (3) and (4) exist; see [2] or [17] for a slightly different ap-
proach.
A. Radon Transform
A basic tool for calculating ridgelet coefficients is to view
ridgelet analysis as a form of wavelet analysis in the Radon do-
main. We recall that the Radon transform of an object
is the
collection of line integrals indexed by
given by
(5)
where
is the Dirac distribution. The ridgelet coefficients
of an object are given by analysis of the Radon
transform via
Hence, the ridgelet transform is precisely the application of a
one-dimensional (1-D) wavelet transform to the slices of the
Radon transform where the angular variable
is constant and
is varying.
B. Ridgelet Pyramids
Let
denote a dyadic square
and let be the collection of all such
dyadic squares. We write
for the collection of all dyadic
squares of scale
. Associated to the squares we con-
struct a partition of energy as follows. With
a nice smooth
window obeying
, we dilate
and transport
to all squares at scale , producing a collec-
tion of windows
such that the s, , make up a
partition of unity. We also let
denote the transport operator
acting on functions
via
With these notations, it is not hard to see that
and, therefore, summing the above equality across squares at a
given scale gives
(6)