Author manuscript, published in "Signal Processing 84, 11 (2004) 2165 - 2173"

DOI : 10.1016/j.sigpro.2004.07.009

Discrete analytical Ridgelet Transform

Philippe CARRÉ, Eric ANDRES

Laboratoire IRCOM-SIC,
bât. SP2MI, av. Marie et Pierre Curie
BP 30179 - 86960 Chasseneuil-Futuroscope Cédex - FRANCE

Abstract
In this paper, we present a new implementation of the Ridgelet transform based on
discrete analytical 2-D lines: the Discrete Analytical Ridgelet Transform (DART).
This transform uses the Fourier strategy (the projection-slice formula) for the com-
putation of the associated discrete Radon transform. The innovative step of the
DART is the construction of discrete analytical lines in the Fourier domain. These
discrete analytical lines have a parameter called arithmetical thickness, allowing us
hal-00354722, version 1 - 21 Jan 2009

to dene a DART adapted to a specic application. Indeed, the DART representa-

tion is not orthogonal it is associated with a exible redundancy factor (depending
on the arithmetical thickness). The DART has a very simple forward/inverse algo-
rithm that provides an exact reconstruction. We discuss the choice of the wavelet
transform applied to the Radon projections and illustrate the extension of the DART
to a Local-DART (with smooth windowing) and Curvelet-DART (with undecimated
quincunx scheme). We apply the DART and its extensions to the denoising and the
partial reconstruction of some images. These experimental results show that the
simple thresholding of the DART coecients is competitive or more eective than
the classical denoising techniques.

Key words: Ridgelet Transform, Discrete Analytical Lines, Denoising, Partial

Reconstruction.

1 Introduction

Image analysis is traditionally aimed at understanding digital signals obtained

by sensors (in our case cameras). Digital information is considered as sampled
continuous information and the theoretical background for it is signal theory.
This is sometimes referred to as digital geometry in opposition to discrete
geometry for computer graphics. These last ten years, since J-P. Reveillès has
introduced it [16], discrete analytical geometry has made an important progress
in dening and studying classes of discrete objects and transformations. This

Preprint submitted to Elsevier Preprint 18 March 2002

 greatly enhanced our understanding of the links between the discrete world
Zn and the continuous world Rn. In the same time, a new discrete signal de-
composition has been developed in image analysis: the wavelet representation.
This new representation has many applications such as denoising, compres-
sion, analysis, etc. One of the aims of this paper is to apply this new insight in
discrete geometry to image analysis and more specically to a particular 2-D
wavelet transform: the Ridgelet transform (Candès and Donoho, 1998 [5]).
Wavelets are very good at representing point singularities ; however they are
signicantly less ecient when it comes to linear singularities. Because edges
are an extremely common phenomena in natural images, an ecient multires-
olution representation of images with edges would be quite advantageous in a
number of applications. A team of Stanford has recently developed an alterna-
tive system of multiresolution analysis, called Ridgelet, specically designed to
eciently represent edges in images [5]. The Ridgelet transform can be com-
puted by performing a wavelet analysis in the Radon domain. However, most
of the work done with Ridgelets has been theorical in nature and discussed in
the context of continuous functions. The important bridge to discrete imple-
mentation is tenuous at best. To our knowledge, we can nd in the literature
only two solutions for the discrete Ridgelet decomposition [9,18]. These two
hal-00354722, version 1 - 21 Jan 2009

strategies are recalled in section 2. This paper presents a new approach that
aims at representing linear singularities with a discrete Ridgelet transform
based on discrete analytical 2-D lines: the Discrete Analytical Ridgelet Trans-
form (DART). The idea behind the associated discrete Radon transform is
to dene each Radon projection by a discrete analytical line in the Fourier
domain:
L![p;q] = (x1; x2) 2 Z2 jqx1 px2j  !2
  




with [p; q] 2 Z2 the direction of the Radon projection and !, a function of

(p; q), the arithmetical thickness.
There are several advantages at using discrete analytical lines: they oer a
theoretical framework for the denition of the discrete Radon projections.
The parameter ! allows us to dene a DART adapted to specic application
(control of the redundancy factor of the transform). We have used, in this
paper, three types of discrete analytical lines: the closed naive lines (DART
redundancy factor  2), the pythagorean lines ( DART redundancy factor
 2:3) and supercover lines (DART redundancy factor  3). This compares
to the transform redundancies of a factor 4 for the Stanford team [18] and a
factor 1 (orthogonal transform) for the Lausanne team [9].
The forward DART is based on the extraction of the 2-D Fourier coecients
belonging to the discrete analytical line L![p;q]. The inverse DART is simply

2
 performed by putting the Fourier coecients back on the exact same place
in the Fourier lattice. The DART has therefore a very simple and rapid for-
ward/inverse algorithm. This simple straightforward approach ensures an ex-
act reconstruction without interpolation nor iterative process (that might be
sensible to noise).
In order to compare the performances of the DART, we have applied the DART
and its extensions (local-DART and Curvelet-DART) to the denoising and the
partial reconstruction of some images. We have compared our approach to ex-
isting ones. The recent developments have shown that the Ridgelet decompo-
sition has very attractive results for the denoising problem [18]. For this, the
Stanford team proposed to apply a thresholding on Ridgelet coecients. Our
experimental results show that the simple thresholding of the DART coe-
cients is competitive or more eective than the wavelet transform. Moreover
the DART denoising results seems to be similar (visually and PSNR measure)
than Stanford's strategy but without any iterative reconstruction algorithm
and with a more exible denition. The analysis reported here show how re-
sults of discrete analytical geometry can be successfully used in image analysis.
In section 2, we recall the continuous theory of the Ridgelet transform and de-
hal-00354722, version 1 - 21 Jan 2009

scribe the existing discrete Ridgelet transforms [9,18]. In section 3, we present

the Discrete Analytical Radon Transform with some details on the theoret-
ical framework of discrete analytical lines. The Discrete Analytical Ridgelet
Transform (DART) is introduced at the end of the section 3. Several illustra-
tions of the DART are proposed in the section 4. In this section, we discuss
the choice of the discrete line type and 1-D wavelet transform associated with
the DART. We compare also the DART with other Ridgelet transform imple-
mentations for the denoising and the partial reconstruction problematics. We
present several extensions of the DART to a local-DART and to the Curvelet
formalism. We end this section with some thoughts on the extension of the
DART to 3-D. We conclude in section 5.

2 The Ridgelet transform

2.1 Continuous theory of Ridgelet transform

A substantial foundation for Ridgelet analysis is documented in the Ph.D.

thesis of Candès [5]. The continuous Ridgelet transform of s 2 L2 (R2) is
dened by :
Z

r(a; b; ) = a;b; (x)s(x)dx

R2

3
 with x = (x1; x2) 2 R2 and a;b; (x) the Ridgelet 2-D function dened from a
wavelet 1-D function as:
x1 cos  + x2 sin  b
!
1=2
a;b; (x) = a
a

b 2 R is the translation parameter, a 2 R is the dilatation parameter and

 2 [0; 2[ is the direction parameter.
The function a;b; is oriented at the angle  and is constant along lines
x1 cos  + x2 sin  = cst
Transverse to these ridges it is a wavelet function. In comparison, the con-
tinuous 2-D wavelet function are tensor products of 1-D wavelets a;b dened
by:
1=2 x b1
a;b(x) = a1;b1 (x1) a2 ;b2 (x2) with a1;b1 (x) = a1 ( a )
1
hal-00354722, version 1 - 21 Jan 2009

With the Ridgelet transform, the translation parameters (b1; b2) of the 2-D
wavelet transform are replaced by the line parameters (b; ).
A basic tool for calculating Ridgelet coecients is to view Ridgelet analysis as
a wavelet analysis in the Radon domain: in 2-D, points and lines are related
via the Radon transform, thus the wavelet and Ridgelet transforms are linked
via the Radon transform.
The Radon transform of s is dened as:
Z

Rs(; t) = s(x)(x1 cos  + x2 sin  t)dx1dx2

R2
where  is the Dirac distribution. The Ridgelet coecients rs of s are given by
the 1-D wavelet transform to the projections of the Radon transform where
the direction  is constant and t is varying:
Z

rs (; a; b) = a;b(t)Rs(; t)dt

R

The Radon transform can be obtained by applying the 1-D inverse Fourier
transform to the 2-D Fourier transform restricted to radial lines going through
the origin (this is exactly what we are going to do in the discrete Fourier
domain with help of discrete analytical lines):
Z

s(! cos ; ! sin ) = e

b
j!t Rs(; t)dt
R

4
 with s the 2-D Fourier transform of s.
b

This is the projection-slice formula that is used in image reconstruction from

projection methods. All these relations are resumed in gure 1.

Fig. 1. Relation between transforms

2.2 Strategies for Discrete Ridgelet Transform

As we have seen, a basic strategy for calculating the continuous Ridgelet trans-
form is rst to compute the Radon transform Rs and secondly, to apply a 1-D
wavelet transform to the slices Rs(; :). The implementation of the discrete
Ridgelet transform can use the same principle.
hal-00354722, version 1 - 21 Jan 2009

The discrete wavelet decomposition, associated to a lter bank [15], is easy

to implement, is stable and invertible, and can be associated to a discrete
orthogonal representation. The discretization of the Radon transform is more
dicult to achieve. The majority of methods proposed in the literature have
been devised to approximate the continuous formula. None of them, however,
were specically designed to be invertible transforms for discrete images and
can therefore not be used for the discrete Ridgelet transform.
The discrete Radon transform can be computed with one the two following
strategies:
 Spatial strategy for digital Radon transform. The Radon transform is de-
ned as summations of image pixels over a certain set of lines, that are
dened in a nite geometry.
Rs([p; q] ; b) = s(x1; x2)
X

(x1 ;x2 )2L[p;q];b

with L[p;q];b = f(x1; x2) 2 [0; N 1]  [0; N 1] : qx1 px2 b = 0g and

(p; q) the normal vector.
 Fourier strategy for digital Radon transform. The projection-slice formula
suggests that approximate Radon transforms for digital data can be based
on discrete Fast Fourier transforms (FFT). The Fourier-domain computa-
tion of an approximate digital Radon transform is dened as:
(1) Compute the 2-D FFT of s

5
 (2) Extract Fourier coecients along the lines L going through the origin.
(3) Compute the 1-D inverse FFT on each line L (dened for each value of
the angular parameter ).
This approach can be problematic since step 2 is not naturally dened on
discrete data.
Recently, some articles studied the implementation of the discrete Ridgelet
transform. Do and Vetterli proposed in [9] an implementation method of
Ridgelet transform based on the use of the nite Radon transform for Zp2,
the cartesian product of two sets of integers mod p, where p is a prime. This
method achieves both invertibility (the inverse transform is stable) and non-
redundancy (the associated Ridgelet transform is orthogonal). The obtained
denoising results presented in [9] illustrate that this strategy is more eective
than the Wavelet transform in recovering straight edges. This transform is
not geometrically faithful (the nite Radon transform has an important wrap-
around eect) and is only dened for image s such that s 2 l[02 :::p 1][0:::p 1]
where p is a prime number. Moreover this Radon transform integrates over
lines that can be rather arbitrarily spread out over the spatial domain and
discrete lines in the Fourier domain associated with this decomposition are
not closed. This approach is not based on a geometrically faithful notion of
hal-00354722, version 1 - 21 Jan 2009

Ridgelets and suers from artifacts (for example in denoising application: g-
ure 10).
In [18], Starck et al. proposed to use an interpolation scheme that substitutes
the sampled values of the Fourier transform obtained on the square lattice with
sampled values of s on a polar lattice. They use a pseudopolar Fourier domain
b

that oers a notion of polar Fourier domain better adapted to digital data (the
digital Fourier domain is viewed as a sequence of squares, not circles). The
discrete pseudopolar Fourier transform of a digital image s 2 l[02 :::n 1][0:::n 1]
is dened by sampling the 2-D Fourier transform at the collection of pseu-
dopolar grid points illustrated by gure 2 (from [3]). The paper [18] follows
the strategy based on the pseudopolar grid. However, it uses a simple nearest-
neighbor interpolation scheme to evaluate pseudopolar grid points in terms of
nearby cartesian grid points. More recently Donoho et al. proposed in [3] a
fast pseudopolar Fourier transform based on a chirp-Z transform (for the exact
evaluation of the 2-D Fourier transform at these non cartesian points). The
associated Radon transform (called Slant) is algebrically exact and geometri-
cally faithful. However in order to be invertible this transform requires a xed
factor four of redundancy. Moreover, the inverse transform is ill-conditioned
in the presence of noise and requires an iterative approximation algorithm.
In this paper, we propose to use the Fourier strategy for the digital Radon
transform. Our lines L are dened with help of the discrete analytical ge-
ometry theory in the Fourier domain [2,16]. This solution allows us to have
dierent Ridgelet decompositions according to the arithmetical thickness of

6
 Fig. 2. (a) The pseudopolar grid (b) The conversion from cartesian grid to pseu-
dopolar grid (from [3])
the discrete lines. This approach presents a limited wrap-around eect. As
for the Donoho's transform [3], our representation is redundant however the
degree of redundancy can be adapted by our thickness parameter. Our Radon
backprojection is very simply and permits an exact reconstruction (analytical
Radon transform followed by backprojection analytical Radon transform is
a one-to-one transform). The objects dened in discrete analytical geometry
(3D lines, planes, hyperplanes, ...) allows extensions to 3D and even higher
hal-00354722, version 1 - 21 Jan 2009

dimensions (see section 4.4).

2.3 Multiscale Ridgelet: Curvelet

The Curvelet transform proposed by Donoho [13] allows us to obtain a multi-

scale Ridgelet analysis with interscale orthogonality. The Curvelet transform
is obtained by ltering and then applying a windowed Ridgelet transform on
each bandpass image. The motivation behind the Curvelet transform is that by
smooth windowing, segments of smooth curves would look straight in subim-
ages, hence they can be well captured by a local Ridgelet transform. Subband
decomposition is used to keep the number of Ridgelets at multiple scales under
control by the fact that Ridgelets of a given scale exist in a certain subband.
The window's size and subband frequency are linked such that Curvelets have
support verifying the key anisotropy scaling relation for curves [13]:
width  length2

Starck et al. propose a Curvelet transform that uses an undecimated lter

bank [18] for the subband decomposition. They applied their digital Curvelet
transform to the denoising of some standard images embedded in white noise.
These empirical results reported in [18] and based on simple thresholding of
the Curvelet coecients show that this strategy is very competitive with state
of the art techniques based on wavelets.

7
 The study of the DART generalization to the Curvelet transform is not the
aim of our article but some examples will illustrate how our Ridgelet tranform
can be extended to the multiscale approach.

3 Discrete Analytical Ridgelet Transform : DART

The idea behind our associated discrete Radon transform is to represent each
direction by a discrete analytical straight line. For this we need a discrete
straight line that has a central symmetry and that forms a good approxi-
mation of the corresponding Euclidean straight line (i.e. direction). Without
central symmetry (i.e. if (x; y) and ( x; y) both belongs to the line), the in-
verse Fourier transform leeds to imaginery values. This excludes the classical
Bresenham discrete 2-D line [4]. Instead we chose discrete analytical 2D lines.
It denes not a unique line but a family of lines with a thickness parameter,
called arithmetical thickness. The arithmetical thickness provides a control
over the transform redundancy factor and properties such as the connectivity
of the straight line. This thickness allows us to choose the discrete straight
hal-00354722, version 1 - 21 Jan 2009

line denition that ts best a given application. The discrete line is analyti-
cally dened meaning that the discrete line is dened by equations that allow
a precise study of the properties and immediate extensions to higher dimen-
sions. An important body of theory, called discrete analytical geometry, is now
available since J-P. Reveillès rst proposed such an approach [16].

3.1 Denition and properties of the closed discrete analytical lines

The discrete analytical line we use for our transform are dened as follows [2]:
L![p;q] = (x1; x2) 2 Z2 jqx1 px2j  !2
  




 
with [p; q] 2 Z2 the direction of the Radon projection (we have  = arctan pq )
and !, a function of (p; q), the arithmetical thickness. J-P. Reveillès introduced
the discrete analytical lines in 1991 [16], dened as 0  qx py + r < !. In
this paper, since we need central symmetry, we chose a variant of the closed
discrete analytical lines, dened as 0  qx py  !; studied in [2].
It is easy to see that the closed discrete analytical lines L![p;q] have a central
symmetry regardless of the value of !. Moreover, the discrete analytical line
can easily be extended to higher dimensions as discrete analytical hyperplanes
dened by j ni=1 qixij  !2 [1].
P

8
 The arithmetical thickness ! is an important parameter that controls, among
other things, the connectivity of the discrete lines: let's consider the closed
discrete analytical line L![p;q] and its Euclidean counterpart L[p;q] : qx1 px2 =
0, then:
 For ! < max(jpj ; jqj), L![p;q] is not connected;
 For ! = max (jpj ; jqj), L![p;q] is 8-connected. This is called the closed naive
line. It is directly related to the distance d1 since

jpj;jqj)


2 Z2 d

1   

Lmax(
[p;q] = M 1 M; L[p;q] 
 

2  
A B
A
with d1 (A; B ) = x1 x1 + x2 x2
 B






 For !  p
max (jpj ; jqj), L![p;q] is 8-connected.
 For ! = p2 + q2, L![p;q] is 8-connected. This is called the Pythagorean line.
This type of line is directly related to the distance d2 since
pp2+q2
hal-00354722, version 1 - 21 Jan 2009

2 Z2 d

 1  

L[p;q] = M 2 M; L[p;q] 
q

2
2 2
with d2 (A; B ) = (xA1 x1 ) + (x2 x2 )
B A B

 For ! = jpj + jqj, L![p;q] is 4-connected. This is called the supercover line and
has a theoretical importance [2]. This type of line is directly related to the
distance d1 since

 12
  

L[jp;q
pj+jqj =  

] M 1 M; L[p;q]
2 Z2 d 
   
with d1 (A; B ) = max xA1 xB1 ; xA2 xB2









.
 For !  jpj + jqj, L![p;q] is 4-connected.
These results are direct consequence of a well known result in discrete ana-
lytical geometry and more recent studies on distances [16,2]. The fact that
these lines can be dened with help of distances makes a direct link with
mathematical morphology [17].
In this paper we studied the DART with the closed naive lines, the pythagorean
lines and the supercover lines.

9
 3.2 Denition of the discrete analytical Radon transform

We use the Fourier domain for the computation of discrete Radon transform:
Fourier coecients of s are extracted along the discrete analytical line L![p;q]
b

s(f1k ; f2k ) such that qf1k pf2k  !2

 
P[!p;q]s =
[
 
b  
k2Z+

and we take the 1-D inverse FFT of P[!p;q]s on each value of the direction [p; q].
Formally, our discrete analytical Radon transform is dened by:
KX1
R! s([p; q] ; b) = P[!p;q]s(k):e2j Kk b with K length of L![p;q]
k=0

We must dene the set of discrete directions [p; q] in order to provide a com-
plete representation. The set of line segments must cover all the square lattice
in the Fourier domain. For this, we dene the directions [p; q] according to
pairs of symmetric points from the boundary of the 2-D discrete Fourier spec-
hal-00354722, version 1 - 21 Jan 2009

tra. Figure 3a illustrates this choice of angles with the covering of the Fourier
domain by the associated Euclidean lines. Notice that this set of angles is not
equispaced.
Proposition 1 Let a square lattice be dened as
2N = [ N ; N ]  [ N ; N ].
Let us consider the set of directions (pm ; qm ) with 0  m  2N , (pm ; qm) =
(N; m N ) and for 2N + 1  m  4N 1, (pm ; qm) = (3N m; N ). The
set of all discrete analytical lines dened by jqmf1 pmf2 j  !2m with !m 
sup (jpmj ; jqmj) provides a complete cover of the lattice
2N .
The proof of this proposition is obvious since we are dealing here with discrete
lines that are at least 8-connected as stated in the previous section. As soon
as we take ! = sup (jpm j ; jqmj) 1 the lattice is not covered anymore (because
of Lsup( jpm j;jqm j) 1=0 and L0 ). More complex formulas for ! can probably be
[0;1] [1;0]
propose that provide a complete cover and a lower redundancy. This is still an
open and somewhat dicult question. In our applications we chose to work
only with connected discrete lines.
Figure 3 illustrates the cover of the Fourier lattice (on the rst octant) for
two dierent types of discrete lines. The grey value of the pixel represents
the redundancy in the projection (number of times a pixel belongs to a dis-
crete line). One isolated line is drawn to illustrate shape of the discrete lines
depending on its arithmetical thickness.
 In gure 3b, we show the redundancy for the closed naive lines. They provide
a relatively small DART redundancy factor of  2:05. These lines are well

10
 Fig. 3. (a) Cover of the Fourier domain with the Euclidean line L[p;q]. (b) Redundancy
hal-00354722, version 1 - 21 Jan 2009

on the cover of the Fourier lattice by closed naive lines, (c) by supercover lines
suited for applications such as partial reconstruction where a relatively small
redundancy is an asset.
 in Figure 3c, we show the redundancy for the supercover lines. They provide
a more important DART redundancy factor of  3:05. These lines are well
suited for applications such as denoising.
 The pythagorean lines provide an intermediate DART redundancy factor of
 2:35.
We now give two examples of the discrete analytical Radon transform in ac-
tion. We propose a comparison with the dierents solutions proposed in the lit-
erature. The papers of Stanford [18] reference a Matlab Toolbox developed by
Stanford researchers for performing Ridgelet and Curvelet analysis. Although
this code is actually not open to the public. In place, we use the Radon trans-
form developped by Mathworks that is based on a geometrically faithful no-
tion of Radon and that does not presents wrap-around eect. The result of the
Mathworks transform is visually similar than the Stanford transform. The code
of Do and Vetterli [9] is open to the public (http://lcavwww.ep.ch/~minhdo).
In gure 4, we show the result of the discrete Radon decomposition of an array
containing a single nonzero entry. We observe that :
 In this case, the Radon transform follows a broken line as for the Mathwork
transform.
 The nite Radon transform of Do and Vetterli has an important wrap-
around eect and is dicult to interpret.

11
  Our transform obtains a broken line without wrap-around eect.
Figure 4d illustrates the inuence of the increasing of the redundancy of the
decomposition (in this case, supercover lines are used) : the Radon is an in-
terpolated broken line.
hal-00354722, version 1 - 21 Jan 2009

Fig. 4. The Radon transform of a point : (a) with a geometrical strategy (b) with
the Do and Vetterli approach. (c) with our strategy using the closed naïve lines, (d)
the supercover lines
We want now to nd for a given coecient ([p; q] ; b) of the Radon transform,
which pixels have contributed [3]. For this, we apply the Radon backprojection.
Our analytical reconstruction procedure works as follows:
(1) Compute the 1-D FFT transform for each set R! s([p; q] ; :)
(2) Substitue the sampled value of s on the lattice where the points fall on
b

lines L![p;q] with the sampled value of s on the square lattice.

b

(3) Apply the 2-D IFFT transform.

The previous procedure allows us to obtain an exact reconstruction if the set

of directions of lines provide a complete cover of the square lattice. To make a
comparison, we use the IRadon procedure of Mathworks that uses the ltered
backprojection algorithm (based on an interpolation in the Fourier domain)
to perform the inverse Radon transform. Notice that the results obtained with
this backprojection are visually similar to Stanford's results. We use also the
code of Do and Vetterli. All these backprojections are presented in gure 5.

12
 We can see that our strategy does not exhibit geometrical distortions and a
very limited wrap-around artifacts (that could be removed by zero-padding
in the Fourier domain). The results are similar for the dierent types discrete
analytical lines.

Fig. 5. The Radon backtransform of a point : (a) with a geometrical strategy (b)
with the Do and Vetterli approach. (c) with our strategy using the closed naive lines,
(d) and the supercover lines
hal-00354722, version 1 - 21 Jan 2009

3.3 Denition of the DART

Now, to obtain the Ridgelet transform, we must take a 1-D wavelet trans-
form along the radial variable in Radon space. The choice of discrete one-
dimensional wavelet transform is discussed by Starck et al. in [18]. They indi-
cate that experience has shown that compactly-supported wavelets can lead
to many visual artifacts when used with nonlinear processing. This is due to
the lack of localization of such compactly-supported wavelets in the frequency
domain. The rst Stanford implementations have made the choice of bandlim-
ited wavelets, whose support is compact in the Fourier domain. For this, they
use the periodic discrete Meyer wavelet that consists of a system of division
in the frequential domain. The discrete Meyer wavelet transform is studied
in Kolaczyk's thesis [14]. After that, Starck et al. chose a specic overcom-
plete system: they dene the scaling function in the frequency domain as a
renormalized B3 spline and the wavelet function as the dierence between
two consecutive resolutions. With this choice, each subband is sampled above
the Nyquist rate, hence avoiding aliasing (this phenomenon is present when
a nonlinear processing is applied on orthogonal wavelet transforms). Do and
Vetterli use classical decimated Symlet Wavelet for the denoising problematic.
In this article, we do not propose a denitive solution associated with the
dierent discrete analytical lines. The choice of the type of 1-D wavelet trans-
form depends of the goal of the transform. This wavelet transform can be
decimated or undecimated and the wavelet base can be adapted according to
the application, as for the classical wavelet decomposition. In the following sec-

13
 tion, we compare the use of Meyer wavelets (whose support is compact in the
Fourier domain) and the use of decimated/undecimated compactly-supported
wavelets for two applications : denoising and partial reconstruction.

4 Illustration of the DART

4.1 naive lines vs pythagorean lines vs supercover lines

A. Denoising
The denoising procedure by Ridgelet transform consists simply in threshold-
ing the Ridgelet coecients and computing the inverse Ridgelet transform.
The thresholding is performed with help of an undecimated method devel-
oped for the wavelet decomposition [8,7]. The redundancy of the wavelet de-
composition, associated with this method, reduces artifacts that appear after
thresholding. Let rs! be the noisy undecimated Ridgelet coecients, we use
hal-00354722, version 1 - 21 Jan 2009

the following hard-thresholding:

8
>
< rs! ([p; q] ; a; b) if rs! ([p; q] ; a; b)  
rs! ([p; q] ; a; b) = >
: 0 otherwise

q
 can be dened as  = 2 log(N ) [10]. We can use two strategies for the
estimation of the variance of the noisy Ridgelet coecients :
 We consider that the DART is not norm-preserving and the variance de-
pends thus on the projection index. In this case, the individual variance is es-
timated using the absolute median of the wavelet decomposition's rst scale
of each radial projection (if the wavelet decomposition is norm-preserving).
 By evaluating the DART (dened with Daubechies D20 wavelets) of a few
standard white noise images, we have observed that the variance of noisy
coecients remains constant. The variance can then be estimated before
the Ridgelet transform and used for all the Ridgelet transform.
Our experiments have shown that the rst strategy is better for denoising with
the DART (the SNR measure of reconstructed image is higher). We use thus
this rst method.
In order to illustrate precisely the result of the denoising algorithm with dif-
ferent types of discrete analytical lines we have generated an articial image

14
 (Figure 6a) and added important white noise (Figure 6b). The SNR 1 of the
noisy image is equal to 15 dB. Figures 6c, d and e are the results obtained
with the three denitions of discrete analytical lines (naive, pythagorean and
supercover) and with the Daubechies D20 wavelets. For a more redundant
decomposition (based on supercover discrete lines, gure 6e) the denoising re-
sult is better than for a less redundant decomposition (gure 6c): the edge is
reconstructed more precisely and the uniform areas are smoother. As for the
wavelet decomposition, overcompletness provides advantages for denoising.
For comparison the result of a denoising carried out by classical decimated
wavelet coecient threshold (gure 6f) is also given. The result is obtained
with a decimated decomposition by Daubechies D20 wavelets and a hard
thresholding. The threshold is based on the noise variance, as Donoho et al.
introduced in [10]. Noise variance is estimated using the absolute median of
the rst decomposition scale. The wavelet method obtains the best SNR mea-
sure, but it exhibits numerous blemishes, that are a result of the nonlinear
processing with critical sampling. This experiment illustrates the limits of the
SNR measure since it does not quantify well this type of artifacts.
hal-00354722, version 1 - 21 Jan 2009

Fig. 6. (a) original image (b) noisy image (to show more precisely the eect of
the noise we plot a vertical line of each image). (c) denoising with naive lines (d)
denoising with pythagoricean lines (e) denoising with supercover lines (f) denoising
with decimated wavelet transform
B. Partial reconstruction
1 All the SNR measures are done with the Stanford Matlab function.

15
 Contrary to the denoising problematic, for a partial reconstruction, redun-
dancy is of course not interesting. Figure 7 compares partial reconstruction
of an articial image by using the 30% largest naive-DART coecients with
reconstruction by using the 30% largest supercover-DART coecients. This
illustrates how the arithmetical thickness of the discrete lines employed in our
Ridgelet transform inuences the quality of the compressed image. The lower
redundancy representation (naive discrete lines) preserves all the features of
the original image after a simple thresholding (Figure 7 b). On the other hand,
with the higher redundancy representation (supercover lines) we loose features
and the image is globally of lower quality.
hal-00354722, version 1 - 21 Jan 2009

Fig. 7. Partial reconstruction of an articial image (a) naive-DART (b) super-

cover-DART

16
 4.2 Inuence of the 1-D wavelet transform choice

We discuss now the choice of 1-D wavelet transform with the DART. For this,
several 1-D wavelet transforms were tested along the radial variable in discrete
analytical Radon space (for these experiments, we use only the naive discrete
line denition, the results are equivalent with the others types of lines):
(1) The discrete Meyer wavelet (whose support is compact in the Fourier
domain) transform.
(2) The Daubechies D20 wavelet (whose support is compact in the time do-
main) transform.
(3) The undecimated Daubechies D20 wavelet transform.
(4) The overcomplete system based on the work of Starck et al. [18] (whose
support is compact in the Fourier domain).
A. Denoising
Figure 8 considers an object used by Do and Vetterli in their paper [9], and
compares the denoising of this image by thresholding of the DART with dif-
hal-00354722, version 1 - 21 Jan 2009

ferent wavelet bases. We use the SNR and the visual analysis to measure the
performance. We observe that :
 The undecimated methods oer superior performances over orthogonal wavelet
decompositions.
 With decimated wavelet decomposition, the compactly-supported wavelets
in the frequency domain obtain better SNR measures than compactly-
supported wavelets in the time domain.
 The Starck-DART reconstruction and undecimated Daubechies D20 DART
reconstruction are very similar.
For denoising with the DART, we can not conclude that the undecimated
compactly-supported wavelets in the frequency domain enjoys superior per-
formance over the undecimated compactly-supported wavelets in the time
domain. However, in this study, we use time lters (D20 lters) that have
very large supports. In this case the lter prole in Fourier domain converges
rapidly to 0 after the cut-o frequency. If the same experiments are done
with too small lters, the reconstructed image will contain some disturbing
artifacts.
More generally, we have generated a set of noisy images from both Do's Object
image and Lenna image. We have then compared the four dierent ltering
procedures. These experimental results have conrmed that the Starck-DART
reconstruction and undecimated Daubechies D20 DART reconstruction are
similar and seem to be the better choice.

17
hal-00354722, version 1 - 21 Jan 2009

Fig. 8. (a) Noisy image (b) denoising with the Meyer-DART transform (c) denoising
with D20-DART transform (d) denoising with undecimated D20-DART transform
(e) denoising with Starck-DART transform
B. Partial reconstruction
We now consider the partial reconstruction of Do's object by the 30% largest
DART coecients (gure 9). Only the decimated wavelet transforms are tested
since the redundancy is irrelevant for partial reconstruction. We observe that,
in the frequency domain, the more compact the wavelet support is, the better
the SNR measure is. These results conrm Stanford's experience. However,
comparing the dierent reconstructed images, one can hardly see the dier-
ence. We have generated a set of partial reconstructions from both Do's object
and Lenna. These experiments conrm previous visual analyzes: partial recon-
structions from D20-DART and Meyer-DART are equivalent.
From these experiments (denoising and partial reconstruction), we propose
to use the D20 wavelets for the DART because the D20 basis can be simply
associated to an orthogonal transform or an undecimated transform (with the
à trous algorithm) and moreover this base is norm-preserving.

18
 Fig. 9. Partial reconstruction of Do's image (a) Meyer-DART (b) D20-DART (c)
D8-DART
4.3 DART vs litterature Ridgelet transform

We briey compare the DART with the other Ridgelet transform implementa-
tions we know about. As we have said the Stanford's code is actually not open
hal-00354722, version 1 - 21 Jan 2009

to the public (a comparison with Stanford's results in PSNR terms is done

in the following section). In place, we use the Radon transform developped
by Mathworks that satises a geometrically faithful notion of Radon (with no
wrap-around eect). This allows us to measure the importance of having or
not a wrap-around eect.
A. Denoising with DART and Local DART
Figure 10 compares the denoising of the very noisy Do's image (SNR 18.73
dB) by thresholding the Ridgelet transform:
 The Ridgelet based on geometrical Radon transform (Mathworks code)
and the Starck overcomplete wavelet system. In this case, the noise variance
is estimated by evaluating the geometrical Ridgelet transform of standard
white noise images (as proposed in [18]).
 The Do and Vetterli orthogonal Ridgelet.
 The DART based on supercover lines and undecimated D20 wavelets.
We illustrate also the results obtain with decimated and undecimated wavelet
transform (implemented by Stanford in the toolbox Wavelab).
We observe that :
 The geometrical Ridgelet and the DART reconstruction does not contain
the many blemishes that one sees in the wavelet reconstructions. The SNR
measures of these two transforms are better.

19
  The decimated wavelet transform displays disturbing artifacts. The use of a
redundant wavelet transform considerably reduces the number of artifacts.
 The Do's Ridgelet transform corresponds to a critical sampling. The recon-
struction associated with a nonlinear processing contains thus many visual
artifacts.
 The DART enjoys superior performance over Geometrical Ridgelet trans-
form 2 .
hal-00354722, version 1 - 21 Jan 2009

Fig. 10. (a) noisy image ; denoising with (b) the decimated wavelet transform (c)
the undecimated wavelet transform (d) the undecimated geometrical Ridgelet (e)
the Do and Vetterli approach (f) The supercover-DART
B. Extension to Local DART and Curvelet DART
The DART can be easily extended to a local transform by a smooth partition-
ing, or more generally to a Curvelet transform by the following steps:
 Decompose the image into a set of wavelet bands.
 Each subband is smoothly windowed into squares.
 Analyse each square by the DART.
Notice that the subbands used in the discrete Curvelet transform of continuum
2 We can obtain better results with the Mathworks Radon transform by using an
other level of threshold. However, because the Radon and the Wavelet transform
associated with this strategy is not norm-preserving, it is dicult to dene a general
level of threshold.

20
 h i
functions have the non-standard form 22l; 22l+2 : Starck et al. propose to use
a specic overcomplete subband ltering: the wavelet function is dened as the
spatial dierence between two consecutive resolutions. We extend our digital
Ridglelet transform with the same principle, but we substitute the Starck's
overcomplete subband ltering with an undecimated quincunux wavelet de-
composition in order to preserve the interscale orthogonality. Moreover this
wavelet transform preserves the variance of the noise. The comparison between
Ridgelet denoising results and Curvelet denoising results is however not the
subject of this paper. These experiments are done precisely by Starck et al.
(http://www-stat.stanford.edu/~jstarck). We illustrate only how the DART
can be extended to the Curvelet formalism.
To illustrate the local DART and the new discrete Curvelet transform based on
discrete analytical lines, we have developed an example of denoising algorithm.
As for the Ridgelet transform, the denoising by Curvelet transform consists
simply in thresholding the Curvelet coecients and computing the inverse
Curvelet transform. For the estimation of the variance  of the noisy Curvelet-
DART coecients, we consider that the Curvelet-DART is norm preserving.
The noise variance is then estimated before the Curvelet transform.
hal-00354722, version 1 - 21 Jan 2009

In our example, a Gaussian noise with a standard deviation equal to 60 was

added to the classical Lenna image (Figure 11a). Several methods were used
to lter the noisy image: thresholding of the undecimated wavelet transform
(Figure 11b), thresholding of supercover DART (Figure 11c), thresholding of
local supercover DART (Figure 11d), thresholding of Curvelet DART based
on supercover discrete lines and undecimated quincunux scheme (Figure 11e).
We observe that the Curvelet DART and Local DART enjoy superior per-
formance over the other transforms. For these decompositions the analysis is
better localized in the spatial domain. This property eliminates the parasitic
lines present in the DART reconstruction and due to the selection of an im-
portant contour in an area of the image. As for the Ridgelet transform, we can
use other denitions of discrete analytical lines for Local DART and Curvelet
DART.
To study the dependency of these methods on the noise level, we generated
a set of noisy images from both Do's Object and Lenna. We then compared
the dierent ltering procedures. This series of experiments is summarized in
Figure 12 for Lenna. These experimental results show that the Local DART
and Curvelet-DART correspond, in general, to the better results.
In order to make a comparison with Stanford's results, we use one of the ex-
ample of denoising proposed in [18]: a Gaussian noise with a standard de-
viation equal to 20 is added to the Lenna image (size: 512 by 512). The
reconstruction by local DART and Curvelet-DART are presented gure 13.
The results of Stanford's algorithm are visible at the website http://www-

21
hal-00354722, version 1 - 21 Jan 2009

Fig. 11. (a) noisy image (b) denoising with the undecimated wavelet transform (c)
denoising with DART based on supercover lines (d) denoising with local DART
based on supercover lines (e) denoising with DART-Curvelet based on undecimated
quinconx wavelet transform and supercover lines
64 (1) 32 (1) 16 (1) Curvelet
DART 31.42 31.38 30.61 31.23
Stanford 30.79 30.97 30.87 31.95
Table 1
PSNR measures of DART and Stanford reconstructions. (1) Size of the window
stat.stanford.edu/~jstarck. From the gure 13 and the Stanford Web site,
one can not see important dierences between the Stanford strategy and the
DART strategy.
In order to make an quantitative measure we used the PSNR. We indicate in
the table 1,
 The PSNR values 3 after ltering with the local DART and Curvelet-DART.
 The PSNR values after ltering with the Stanford algorithm indicated in
[18].
3 The PSNR measures of DART are done with the Stanford Matlab function.

22
 Fig. 12. SNR of reconstructed image versus SNR of noisy image for dierent repre-
sentation. The panel corresponds to Lenna.
hal-00354722, version 1 - 21 Jan 2009

Fig. 13. (a) noisy image with the same level of noise than [18] (b) denoising with
local DART based on supercover lines (c) denoising with DART-Curvelet based on
undecimated quincunux wavelet transform and supercover lines.
These measures conrm the visualy analysis: the Stanford reconstruction and
DART reconstruction are very similar. However, as we have said, the back-
projection DART algorithm is simpler and the DART representation is more
exible (with the variable line thickness).
Notice that Color denoising results obtained with the DART are presented in
[6].
C. Partial reconstruction
We now consider the partial reconstruction of Do's object by the 512 most

23
 important Ridgelet coecients (gure 14). Only the decimated wavelet trans-
forms associated with naive lines are tested. Since our decomposition is re-
dundant, we observe that the DART is not adapted to this problematic con-
trary to the orthogonal wavelet representation and Do's Ridgelet. However in
order to obtain a more sparse representation, we use the strategy called ortho-
Ridgelet by Stanford and introduced very recently in [11]. This transform may
be viewed as a discrete Radon transform followed by an orthogonal 2-D wavelet
transform (we obtain ortho-Ridgelets by taking the wavelet transform along
the angular variable of the Ridgelet projection). The ortho-Ridgelet represen-
tation is more sparse: the transform along the angular variable has compressed
the laterally features of the Ridgelet transform into point-like features. If we
apply this  wavelet transform to the DART representation, we obtain for the
partial reconstruction a result that is close to Do's result. However the ortho-
DART reconstruction has a smaller SNR measure than the two orthogonal
transforms. This result is a rst experimentation and needs to be developed
(choice of the best basis along the angular variable, a local approach ...).
hal-00354722, version 1 - 21 Jan 2009

Fig. 14. Partial reconstruction of Do's image (512 coecients) (a) Wavelet (b) Do's
Ridgelet (c) Naive DART. Bad result because of the redundancy (d) Naïve or-
tho-DART

4.4 DART 3-D ?

As for the 2-D domain, the 3-D Radon transform of an object is related to
its 3-D Fourier transform via the central slice theorem: the 3-D Radon trans-
form can be obtained by applying the 1-D inverse Fourier transform to the
3-D Fourier transform restricted to radial lines going through the origin. Once
again, we propose to extract the Fourier coecients along a 3-D discrete an-
alytical line going through the origin. Once a discrete 3-D line is dened, the
principle of the method is the same as in the 2-D case with the same properties
(exact reconstruction, rapidity, exible denition). This idea works well for the

24
 supercover model as the 3-D supercover discrete analytical line is connected
(see example in gure 15). Thus the set of 3-D supercover lines covers the
3-D Fourier cubic lattice if we dene the line directions according to pairs of
symetric border points of the lattice. The denition of a 3-D supercover line
of direction [p; q; r] is given by [2]:

jqx1 px2j  jpj +2 jqj

jrx px j  jpj + jrj
1 3
2
jrx2 qx j  jqj + jrj
3
2

In the same way, all the supercover of all m-ats in dimension n are well dened
[2]. This allows us to generalize this approach to 3-D planes (see example in
gure 15) and dimension n. For the other models (naive and pythagorean
for example) however the denition and the extension to 3-D lines or higher
dimensions is more problematic. For instance, the 3-D naive discrete analytical
line is not connected in general and does not ensure a complete cover of the
hal-00354722, version 1 - 21 Jan 2009

3-D cubic Fourier lattice. We hope to report on this developpements in a

forthcoming paper.

Fig. 15. Discrete analytical 3-D line and plane

5 Conclusion

In this paper, we have proposed a new strategy for implementing the discrete
Ridgelet transform. So far, the developement of the discrete Ridgelet trans-
form has been investigated by two teams (Stanford and EPFL) in previous

25
 works. Our innovative choice is to use the formalism of the discrete analyti-
cal geometry theory in the Fourier domain, in order to dene a new discrete
Ridgelet transform: the Discrete Analytical Ridgelet Transform (DART).
The DART algorithm is easy to implement. It provides an exact reconstruc-
tion property: the DART followed by a reverse DART is a one-to-one trans-
form. Our experiments have shown that our approach presents a limited wrap-
around eect, that does not inuence the denoising or partial reconstruction
results. Moreover, by using the analytical formalisation, we dene a exible
Ridgelet transform: we can dene dierent DART decompositions according
to the arithmetical thickness of the analytical discrete lines. As for Stanfords
transform, our representation is redundant but the transform redundancy fac-
tor depends on the type of the discrete lines used and can be adapted for each
application.
Our DART uses a 1-D wavelet transform based on Daubechies D20 wavelet.
Our experiments have shown that this basis associated with the analytical
Radon transform obtains satisfactory results for the two studied problematics.
Moreover, this basis can be simply associated to an orthogonal transform
for partial reconstruction or to an undecimated transform for denoising, and
hal-00354722, version 1 - 21 Jan 2009

it is norm-preserving. We have seen that if, instead, we use a basis that is

compactly supported in the Fourier domain (orthogonal or redundant), it does
not improve the results of the DART, but the decomposition is less exible.
We have illustrated the performances of the DART for denoising problemat-
ics, even with very noisy images (with gaussian noise). This study indicates
that the Local-DART and the Curvelet-DART thresholding rivals Stanford's
Ridgelet denoising results, that are considered as the reference for Ridgelet's
denoising applications, and more generally, classical denoising techniques that
have been developped over the last decade. Howerver, contrary to the Stan-
ford strategy, the reverse DART does not require an iterative approximation
algorithm and the redundancy factor of the representation is not xed.
This work can be extended in several directions. One of the theoretical ques-
tions in discrete geometry is the problem of dening an arithmetical thickness
function that provides a smaller redundancy and a cover of the Fourier lat-
tice. This is still an open and dicult arithmetical problem. One of the most
important interests of the discrete analytical approach is the possibility to ex-
tend easily our work to 3-D. We are currently investigating the application of
a 3-D DART to the denoising process of animated video images. In this paper
we present some rst thoughts abouth these questions.
We are also considering extending our partial reconstruction algorithm with
more sophisticated representations. The ortho-Ridgelet proposed by Stanford
seems to obtain good results for partial reconstruction [12] and constitutes an

26
 interesting new area of research.

References

[1] E. Andres, R. Acharya, and C. Sibata. Discrete analytical hyperplanes.

Graphical Models and Image Processing, 59(5):302309, 1997.
[2] Eric Andres. Modélisation analytique discrète d'objets géométriques.
Habilitation, Université de Poitiers, 2000.
[3] A. Averbuch, R. Coifman, D. Donoho, and M. Walden. Fast slant stack: a notion
of radon transform for data in a cartesian grid which is rapidly computible,
algebraically exact, geometrically faithful and invertible. Technical report,
Stanford University, 2000.
[4] J. Bresenham. Algorithms for computer control of a digital plotter. IBM Sys.
Journal, 4(1):2530, January 1965.
[5] E. Candès. Ridgelets: Theory and Applications. PhD thesis, Stanford, 1998.
[6] P. Carré and E. Andrès. Ridgelet transform based on reveillès discrete lines.
hal-00354722, version 1 - 21 Jan 2009

In DGCI'02. (Published in Lecture notes in Computer) Springer-Verlag, April

2002. Bordeaux, France.
[7] P. Carré, H. Leman, C. Marque, and C. Fernandez. Denoising the EHG
signal with an undecimated wavelet transform. IEEE Trans. on Biomedical
Engineering, 45(8):11041113, September 1998.
[8] R. Coifman and S. Donoho. Translation-invariant de-noising. In A. Antoniadis
and G. Oppenheim, editors, Wavelets and Statistics, volume 103 of Lecture Notes
in Statistics, pages 125150, New York, 1995. Springer-Verlag.
[9] M. Do and M. Vetterli. Discrete ridgelet transforms for image representation.
Submitted to IEEE Trans. on Image Processing, April 2001.
[10] D. Donoho. Wavelet shrinkage and w.v.d.: A 10-minute tour. Technical report,
Standford University, 1992.
[11] D. Donoho. Orthonormal ridgelets and linear singularities. SIAM Journal on
Mathematical Analysis, 31(5):10621099, 2000.
[12] D. Donoho and A. Flesia. Digital ridgelet transform based on true ridge
functions. Technical report, Stanford University, January 2002.
[13] D. L. Donoho and Emmanuel J. Candès. Curvelets: A surprisingly eective
nonadaptive representation of objects with edges. Technical report, Stanford
University, 1999.
[14] E. Kolaczyk. Wavelet Methods for the Inversion of Certain Homogeneous Linear
Operators in the Presence of Noisy Data. PhD thesis, Stanford University,
October 1994.

27
 [15] S. Mallat. A theory for multiresolution signal decomposition: the wavelet
transform. IEEE Trans. on PAMI, 11(7):674693, 1989.
[16] Jean-Pierre Reveillès. Géométrie discrète, calcul en nombres entiers et
algorithmique. Habilitation, Université Louis Pasteur de Strasbourg, 1991.
[17] J. Serra. Image Analysis and Mathematical Morphology. Acamedic Press,
London, 1992.
[18] J. L. Starck, E. J. Candès, and D. L. Donoho. The curvelet transform for image
denoising. Technical report, Department of Statistics, Stanford, November 2000.
hal-00354722, version 1 - 21 Jan 2009

28