Advances in Digital Image Compression by Adaptive Thinning

Laurent Demaret (demaret@ma.tum.de) and Armin Iske (iske@ma.tum.de)

Zentrum Mathematik, Technische Universitat Munchen,

D-85747 Garching, Germany
Abstract
This paper proposes a novel concept for digital image compression. The resulting compression scheme relies on adaptive thinning algorithms, which are recent multiresolution
methods from scattered data approximation. Adaptive thinning algorithms are recursive point removal schemes, which
are combined with piecewise linear interpolation over decremental Delaunay triangulations. This paper shows the utility of adaptive thinning algorithms in digital image compression. To this end, specific adaptive pixel removal criteria are designed for multiresolution modelling of digital images. This is combined with a previous customized
coding scheme for scattered data. The good performance
of our compression scheme is finally shown in comparison with the well-established wavelet-based compression
method SPIHT.
1. INTRODUCTION
Over the past few years, digital image compression has become a major challenge in information technology. For
the efficient transmission of digital image data it is crucial to reduce the amount of information by using a sparse
representation of the image. To this end, effective coding
schemes are used in order to maintain a good approximation quality at low bitrates. For a comprehensive introduction to various relevant aspects of digital image compression, we recommend the textbook [8].
The performance of any image compression scheme depends on its ability to capture characteristic features of the
image, such as sharp edges and fine textures, while reducing the number of parameters used for its modelling. Most
of the well-established compression schemes are using the
bivariate Discrete Wavelet Transform (DWT), see the survey [2] on wavelet-based image coding. At high compression rates, wavelet-based methods provide much better image quality in comparison with the JPEG standard, which
relies on the Discrete Cosine Transform (DCT). The good
results obtained from DWT are due to sophisticated techniques which essentially take advantage of the statistical
structure of the image data.
For instance, the well-established method SPIHT [6]
(Set Partitioning Into Hierarchical Trees) uses judicious clusters of non-significant coefficients (collected in zerotrees)
in order to efficiently encode the most significant coefficients. The new compression standard JPEG2000 (see [8]),
based on EBCOT [7], uses contextual encoding, which models the Markovian structures in the pyramidal wavelet decomposition of the image. At very low bit rates, however, the oscillatory behaviour of wavelet bases typically

leads to undesirable artefacts along sharp edges. Wavelets

require too many non-vanishing coefficients to represent
sharp edges accurately, because they introduce artificial oscillations in their modelling.
This paper proposes an alternative concept for digital
image compression, which is particularly well-suited for
the modelling of sharp edges and related features in digital
images. To this end, piecewise linear functions over adaptive triangular meshes are used. This leads to a reduction of
low-pass filtering effects, which are often due to overquantization of the high-pass coefficients. The modelling relies
on recursive point removal schemes, termed adaptive thinning algorithms, and a customized coding scheme for scattered pixels.
The outline of the paper is as follows. The application
of adaptive thinning algorithms to digital image modelling
is explained in Section 2. Then, in Section 3, the abovementioned coding scheme is briefly discussed. Finally, selected numerical examples are shown in Section 4, where
the performance of our compression scheme is compared
with the wavelet-based compression method SPIHT.
2. ADAPTIVE THINNING IN IMAGE
MODELLING
In order to keep this paper widely self-contained, this section first introduces basic concepts and ingredients of adaptive thinning algorithms, before their application to digital
image modelling is discussed.
In many classical image compression methods, such as
for the aforementioned DCT and DWT, the modelling is
carried out by decomposing the image over a non-adaptive
orthogonal basis of functions. The corresponding coefficients of the basis functions are then quantized, according
to a specific quantization step, which usually depends on a
target compression rate. The performance of the resulting
compression scheme depends on the approximation quality
which results from the non-vanishing coefficients. We remark that such approaches do not allow any adaptivity in
the choice of the representation functions. The high quantization steps required for very high compression rates lead
to undesirable low-pass filter artefacts, also called ringing
artefacts. Visually, they correspond to oscillations around
sharp edges. Examples of such artefacts are shown in Figure 1, where the well-known test image Goldhill is
used.
An alternative modelling concept represents the image
by piecewise linear functions over triangular meshes. The
resulting approaches based on hierarchies of regular (i.e.
non-adaptive) triangular meshes, however, lead to low-pass

Figure 2: Delaunay Triangulation of a point set.

Figure 1: Goldhill. Coding by using SPIHT at a very

high compression rate of 0.125 bpp (bits per pixel) leads to
ringing artefacts.

the target function f is approximated by the unique continuous function L(f, Y ), whose restriction on any triangle
in the Delaunay triangulation DY is a linear function and
which satisfies the interpolation conditions
L(f, Y )(y) = f (y),

artefacts, such as in the case of wavelets. In contrast, irregular adaptive triangulations offer much more flexibility, and they support the appropriate concept of adaptivity
for representing natural features of the image. In view of
the required compression, however, this enhanced flexibility may lead to high coding costs required for coding the
node coordinates, and the topology coding (connectivity
between nodes) of the corresponding mesh.
In order to entirely avoid the required costs for the connectivity coding, adaptive thinning algorithms work with
Delaunay triangulations. Recall that a Delaunay triangulation of a discrete planar point set X R2 is a triangulation, such that the circumcircle of each of its triangles does
not contain any point from X in its interior. An example
of such a triangulation is shown in Figure 2. For further
details concerning triangulation methods, we refer to the
textbook [5].
Now we associate, with any finite set X of points its
unique Delaunay triangulation DX (in the case of co-circular points in X there may be ambiguities which we exclude in the following discussion for the sake of simplicity). Thus, at the decoder, the set of points X can directly
be used in order to uniquely reconstruct the triangulation
DX .
The adaptive thinning algorithm, first developed in [4],
is concerned with the approximation of a bivariate function
f from a finite set of scattered data points X R2 and
sample values {f (x)}xX . To this end, a data hierarchy
X = XN XN 1 . . . Xn

(1)

of nested subsets of X = {x1 , . . . , xN } R2 is constructed. This is done by recursively removing points from
X. At any removal step, one point is removed from the current subset Xp X in (1), so that Xp is of size |Xp | = p,
n p N.
The multiresolution method, associated with adaptive
thinning, works with decremental Delaunay triangulations
over the subsets in (1). To this end, for any subset Y X,

for all y Y .

In the subsequent of this text, we say that L(f, Y ) is the

piecewise linear interpolant of f over DY .
Now the aim of adaptive thinning is to remove at each
step a point xp = Xp \ Xp1 such that L(f, Xp1 ) is close
to the original function f . In order to design a suitable point
removal criterion, we employ a specific norm k k used to
evaluate the approximation error
(Y ) (Y, f ) = kL(f, Y ) f k.
The main application addressed in [4] is digital terrain
modelling. The criterion used to remove a point from a set
is based in [4] on the L -norm, measuring the maximal deviation between the original function and the reconstructed
function. In this case, we have
(Y ) = kL(f, Y ) f k = max |L(f, Y )(x) f (x)|.
xX

Our purpose in this article is to apply adaptive thinning

to digital images. In this case, the initial set of points X
is given by the set of pixels in the original image. Moreover, the function values of f are the luminance values at
the pixels. In view of digital image compression by adaptive thinning, a suitable removal criterion should be related
to the mean square error (MSE). Indeed, as supported by
numerical examples, this helps to improve the quality of the
reduced images, according to the human visual perception.
In this case, we prefer to work with the discrete L2 error (Y ), with respect to the set X given by
X
2 (Y ) = kL(f, Y ) f k22 =
|L(f, Y )(x) f (x)|2 ,
xX

rather than with the above error measure . In order to

design a suitable removal criterion for adaptive thinning,
we let the significance of any point y Y be given by
(y) = 2 (Y \ y) 2 (Y ).

Moreover, a point y Y is said to be removable from

Y , if and only if y is least significant among all points in
Y by satisfying
(y ) = min (y).
yY

In order to further reduce the resulting computational

costs, we restrict our computations for computing the significance of any point y Y to its local cell C(y) in DY .
Recall that the cell C(y) of the vertex y is given by the triangles in DY which contain y as a vertex. This leads us to
the simpler significance
2
2
(y) = C(y)
(Y \ y) C(y)
(Y ),

where C(y) (Y ) denotes the L2 -error over the cell C(y),

i.e.,
X
2
C(y)
(Y ) =
|L(f, Y )(x) f (x)|2 .
xC(y)X

Having constructed a subset Xn of n most significant

pixels, n N , by adaptive thinning, we are in a position
to compute a reconstruction of the original image, such as
for Lena in Figure 3 (a), at the decoder as follows. The
n most significant pixels, shown in Figure 3 (b) (where
n = 2205) are used in order to create the corresponding
Delaunay triangulation DXn , shown in Figure 3 (c). This
in turn yields, by evaluating the piecewise linear interpolant
L(f, Xn ), the reconstructed image (Figure 3 (d)).

basis function y , y Xn , is the unique piecewise linear

function over DXn satisfying
(
1
for x = y Xn ,
y (x) =
0
for x Xn \ y.
Due to the properties of the least squares approximation
scheme, this yields optimal luminance values (f (y))yXn
at the pixels in Xn , so that the corresponding best approximation
X
L (f, Xn ) =
f (y)y ,
yXn

attains the least squares error by satisfying

kL (f, Xn ) f k22 = min kg f k22 ,
gVXn

where kk2 is the discrete L2 -norm with respect to Xn . For

a comprehensive treatment of least squares approximation
methods, we recommend the textbook [1].
The use of the criterion based on the L2 -norm, in combination with least squares approximation, yields a significant improvement of the image quality, as shown in Figure 4 for the test image Barbara of size 256-by-256.

(a)
(a)

(b)

(c)

(d)

Figure 3: Lena. (a) Original image of size 128-by-128;

(b) subset Xn of n = 2205 most significant pixels; (c) the
Delaunay triangulation DXn ; (d) reconstructed image.
We remark that the reconstruction quality of the so obtained image, such as in Figure 3 (d), can be further improved. This is done as follows. We consider using least
squares approximation (LSA) with respect to the approximation space VXn = span(y )yXn , where the cardinal

(b)
Figure 4: Barbara. Reconstruction from n = 7336 most
significant pixels (a) with L -norm and without LSA; (b)
with L2 -norm and with LSA.

3. CODING SCHEME
This section briefly explains the coding of the most significant pixels, which is also subject of the previous paper [3]. First note that any coding scheme requires a nonambiguous decoding rule which enables the receiver to uniquely reconstruct the image. As already discussed in the
previous section, this can be accomplished by using any set
Xn of n most significant pixels output by adaptive thinning.
The subset Xn can be considered as a set of tridimen(1)
(2)
(1)
sional points (xi , xi , zi ), 1 i n, where xi and
(2)
xi are the integer coordinates of the point xi and where
(1)
(2)
(1)
(2)
zi = f (xi , xi ) = Q(f (xi , xi )) is a quantized
(1)
(2)
value, where the luminances f (xi , xi ) are output by
least squares approximation. We use a uniform quantization step q, so that Q(z) = dz/qe. As shown in Section 2,
the use of Delaunay triangulations avoids the coding of
any connectivity information. Only the tridimensional locations of the points in the subset Xn are required at the decoder. Furthermore, the ordering of the nodes is not needed
for the reconstruction.
W

NW NE
SW SE

+NW +NE
SW SE
+

using the Peak Signal to Noise Ratio (PSNR),

2552
PSNR = 10 log10
,
MSE
given in dB, where MSE denotes the Mean Square Error
X
1
j)|2 .
MSE =
|I(i, j) I(i,
N M i,j
In our first example, we decided to use a test image
called Peppers, shown in Figure 6 (a), whose size is 256by-256 pixels. Note that this image contains very few textured areas. Our method provides a PSNR value of 31.13 dB
(corresponding to the image in Figure 6 (b)), whereas SPIHT
yields a better PSNR value of 31.65 dB (Figure 6 (c)).
A second example is shown in Figure 7. The size of the
test image, called Fruits (shown in Figure 7 (a)), is also
256-by-256 pixels. In this test case, our method provides
a PSNR value of 32.13 dB (Figure 7 (b)), whereas SPIHT
yields a PSNR value of 32.77 dB (Figure 7 (c)).
Both examples show that our algorithm achieves, in
contrast to SPIHT, accurate localization of sharp edges, and
so it avoids spurious ringing artefacts. Although our method
is slightly inferior to SPIHT in terms of its PSNR, we believe that it is quite competitive. This is supported by the
good visual quality of the image reconstructions by our
compression method (see Figure 6 (b) and Figure 7 (b)).

SW SE

5. ACKNOWLEDGEMENT
Figure 5: First three splits of the cubic domain .
We code the pixel points by performing a recursive splitting of the domain = [0..N ] [0..M ] [0..P ], where N
and M are the dimensions of the image and P is the num
ber of possible values for fi (typically, P = 28 1 = 255
for unquantized data, but P = 255/q when the quantization step is q).
At each step, we split a non-empty domain , initially
= , into two subdomains 1 and 2 of equal size. If
m denotes the number of most-significant pixels in the
domain , then we have m = m1 + m2 . Thus only one
of the two numbers, say m1 , is added to the bitstream.
At the decoder, the number m2 will be deduced from m
and m1 . Each number is coded by the minimal number
of required bits. For instance, since 0 m1 m , the
number m1 is coded by dlog2 (m +1)e bits. The splitting
of the subdomains is performed recursively until the points
are exactly localized. For the purpose of illustration, the
first three splits of the cubic domain are shown in Figure 5. For further details on this particular coding scheme,
we refer to [3].
4. NUMERICAL RESULTS
We have implemented the compression scheme proposed
in this paper. In this section, numerical examples are used
in order to evaluate the performance of our method. To
this end, we compare our compression scheme with the
wavelet-based compression scheme SPIHT [6]. We evaluate the reconstruction quality of the decoded image I by

The authors were partly supported by the European Union

within the project MINGLE, HPRN-CT-1999-00117.
6. REFERENCES
Bjrck: Numerical Methods for Least Squares
[1] A.
Problems, SIAM, Philadelphia, 1996.
[2] G.M. Davis, A. Nosratinia: Wavelet-based image
coding: an overview, Appl. Comp. Control, Signal &
Circuits, B.N. Datta (ed), Birkhauser, 1999, 205269.
[3] L. Demaret, A. Iske: Scattered data coding in digital
image compression, Curve and Surface Fitting: SaintMalo 2002, A. Cohen, J.-L. Merrien, L.L. Schumaker
(eds), Nashboro Press, Brentwood, 2003, 107-117.
[4] N. Dyn, M.S. Floater, A. Iske: Adaptive thinning for
bivariate scattered data, JCAM 145, 2002, 505517.
[5] F.P. Preparata, M.I. Shamos: Computational Geometry, 2nd edition, Springer, New York, 1988.
[6] A. Said, W.A. Pearlman: A new, fast, and efficient
image codec based on set partitioning in hierarchical
trees, IEEE Transactions on Circuits and Systems for
Video Technology 6, 1996, 243250.
[7] D. Taubman: High performance scalable image compression with EBCOT, IEEE Transactions on Image
Processing, July 2000, 11581170.
[8] D. Taubman, M.W. Marcellin: JPEG2000: Image
Compression Fundamentals, Standards and Practice,
Kluwer, Boston, 2002.

(a)

(a)

(b)

(b)

(c)

(c)

Figure 6: Peppers. (a) Original image; (b) compressed

at 0.44 bpp by our method; (c) at 0.44 bpp by SPIHT.

Figure 7: Fruits. (a) Original image; (b) compressed at

0.57 bpp by our method; (c) at 0.57 bpp by SPIHT.