| 2004
ly 30
EDGE-PRESERVING SMOOTHING OF HIGH-RESOLUTION IMAGES WITH A PARTIAL
MULTIFRACTAL RECONSTRUCTION SCHEME
Jacopo Grazzini!, Antonio Turiel*, Hussein Yahia! and Isabelle Herlin!
1 CLIME Project - INRIA Rocquencourt
Domaine de Voluceau BP105
78153 Le Chesnay, France.
{jacopo.grazzini,hussein.yahia,isabelle.herlin } @inria.fr
t Dept. of Geophysics, Institut de Ciences del Mar
Pg. Maritim de la Barceloneta, 37-49
08003 Barcelona, Spain.
turiel@iem.csic.es
KEY WORDS: Remote Sensing, Segmentation, High-Resolution, Compression, Reconstruction, Hierarchical, Texture,
Edges.
ABSTRACT
The new generation of satellites leads to the arrival of very high-resolution images which offer a new quality of detailed
information about the Earth's surface. However, the exploitation of such data becomes more complicated and less efficient
as a consequence of the great heterogeneity of the objects displayed. In this paper, we address the problem of edge-
preserving smoothing of high-resolution satellite images. We introduce a novel approach as a preprocessing step for
feature extraction and/or image segmentation. The method is derived from the multifractal formalism proposed for image
compression. This process consists in smoothing heterogeneous areas while preserving the main edges of the image.
It is performed in 2 steps: 1) a multifractal decomposition scheme allows to extract the most informative subset of the
image, which consists essentially in the edges of the objects; 2) a propagation scheme performed over this subset allows to
reconstruct an approximation of the original image with a more uniform distribution of luminance. The strategy we adopt
is described in the following and some results are presented for Spot acquisitions with spatial resolution of 20 x 20 m?.
1 INTRODUCTION
In the latest years, the improvement of the technology for
observing the Earth from space has led to a new class of
images with very high spatial resolution. New satellite
sensors are acquiring images at finer resolutions, such as,
for instance, resolutions of 2.5m (Spot), 1m (Ikonos) or
even 0.7 m (QuickBird). High resolution (HR) imagery of-
fers a new quality of detailed information about the prop-
erties of the Earth’s surface together with their geograph-
ical relationships. Consequently, this has given rise to a
growing interest on image processing tools and their ap-
plication on this kind of images (Mather, 1995). Smaller
and smaller objects (such as house plots, streets...), as well
as precise contours of larger objects (such as field struc-
tures) are now available, automatic methods for extract-
ing these objects are of great interest. However, due to
the fact that HR images show great heterogeneity, stan-
dard techniques for analyzing, segmenting and classifying
the data become less and less efficient. When the reso-
lution increases, the spectral within-field variability also
increases, which can affect the accuracy of further classifi-
cation or segmentation schemes (Schiewe, 2002): images
contain more complicated and detailed local textures; char-
acteristic objects of the images (fields, buildings, rivers)
are no more homogeneous. Thus, classical approaches
cannot produce satisfactory results because they may in-
duce simultaneously under- and over-segmentation within
a single scene, depending on the heterogeneity of the con-
sidered objects, that confuse the global information and
prevent further analysis. Generally several preprocessing
steps may be required before such methods can be ap-
plied (Schowengerdt, 1997).
This paper introduces a new approach to edge-preserving
1125
smoothing of HR images as a preprocessing step for fea-
ture extraction and/or image segmentation. The problem
can be related with the idea of resolution reduction: the re-
tained technique should enable to preserve the features of
the original image corresponding to the boundaries of the
objects while homogeneizing the other parts (non-edges)
ofthe image. Recently, Laporterie-Déjean et al. (Laporterie-
Djean et al., 2003) proposed a multiscale method for pre-
segmenting HR images. The algorithm consists in per-
forming a non-exact reconstruction: it decomposes and
partially reconstructs images using morphological pyramids
that enable to extract details wich regard their structures
within the image.
Like in (Laporterie-Djean et al., 2003), we address this
problem with a method derived from data compression.
We use the multifractal approach introduced in (Turiel and
Parga, 2000) as a resolution-reducing device. Like many
image processing techniques, it makes use of the image
edge information that is contained in the image gradient.
Such an approach is ideal, as it assumes that objects can
be reconstructed from their boundary information (Turiel
and del Pozo, 2002). Moreover, it constitutes a multiscale
approach which is well recognized to attain the best per-
formance in processing purposes over satellite data, like
image fusion and image merging (Yocki, 1995). Even on
a single scene, different scales of analysis are needed, de-
pending on the homogenity of the objects under consider-
ation and on the desired final application. The process we
finally propose is done in two steps:
e First, meaningful subsets of the original image, which
mainly consists in its boundaries, are extracted using
a multifractal decomposition scheme.
