The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
147
Finally, the information variability from an image to another is
a very tricky problem for comparing their content: same objects
can produce very different signals depending on capture
parameters and external conditions. For example, different
wavelengths or polarizations provide different results but it is
also the case when using the same Radar parameters and when
external conditions are not the same (after raining, the dielectric
constant of the ground changes and produces a different
response).
Another main difficulty with Radar image interpretation
concerns the speckle, which is a noise that gives the images a
grainy appearance (Rees, 2001). The speckle is a direct
consequence of the wave coherence: the interaction with the
target may shift the wave phase because of the target height
variations and/or physical properties. The result of this
interaction is a multiplicative noise. This effect has been
modeled through a statistical approach (Chitroub et al., 2002)
although it is deterministic. •
Many algorithms have been proposed to reduce this noise,
especially through the creation of filters (Touzi, 2002) that are
supposed to widely eliminate the speckle effect while
preserving most of the information being in the images
(radiometry, contours, texture,) and not generating any artifact.
Specific algorithms have to be designed for extracting
primitives (or landmarks) from images because of speckle. For
example, when using classical tools such as gradient or
Laplacian filters, we obtain a variable rate of false alarms (it
increases in areas where the signal is very intense). Thus, the
main objective when designing a new algorithm for extracting
primitives from Radar images is to create a filter that produces a
constant rate of false alarms; these filters are called CFAR
detectors, CFAR meaning “Constant False Alarm Rate” (Touzi
et al., 1988).
2.2.4 Conclusion on Spaceborne platform Radar images:
Automatic registration of Spaceborne platform Radar images is
still a challenge because of some key difficulties:
We do not have any accurate information on the
image relative positions
These images went through various geometrical
deformations
Radiometry may be modified by speckle but also by
several physical phenomena related to the
backscattered signal.
Concerning this last point (i.e. radiometry modifications):
Speckle: this noise changes the value of the similarity
measure and thus, it reduces the efficiency of
landmark extraction
Physical phenomena: a main question when the two
images come from very different acquisition
conditions is to determine if there is enough common
information to provide an automatic registration.
c) Spaceborne platform Radar image registration
As we discussed in the previous section, registering such Radar
images is still a challenge and thus, many research groups focus
their work on this topic (Wessel et al., 2007). In this section, we
mainly discuss on the matching process itself and on the way
the Radar image generation has an effect on this process.
We remind that we can use either an “area-based” approach or a
“feature-based” approach, the first one using statistical
properties of the pixel neighborhood, and the second one using
primitives extracted from images as landmarks to be paired.
Now, let us see how the Radar imaging specificity makes
difficult such processes.
Neighborhood statistical properties are characteristic of what is
called texture, and thus, “area-based” approaches can be seen as
texture pairing methods. In fact, two textures are superimposed
in Radar images: the texture that is characteristic of the area (we
call it the “scene” texture) and the speckle texture. The main
difficulty is then to find algorithms that describe correlations
between the “scene” textures of the two images without having
any interaction with the speckle texture although these two
textures (“scene” and speckle) are mixed in the image.
Primitive extraction from SAR images is also dependent on
speckle, but in a different way than for “area-based” approaches.
As we have seen before, the SNR (Signal to Noise Ratio) is not
very high and thus there are a lot of false alarms (with a variable
rate): these false alarms result in small variations of the
primitives used as landmarks, which produces a slight
instability or a loss of accuracy.
A question arising when registering two SAR images acquired
with different conditions during data acquisition or with
different sensor features, is the nature of the information they
provide. As already mentioned before, information strongly
depends on these factors; two SAR images gather a common
part of information, while at the same time they carry another
part of unique information due to specific internal and/or
external parameters of acquisition. Matching step in a
registration algorithm is thus a difficult problem in SAR
imagery, as methods should be able to rely on common
information only. Scene texture is dependent on acquisition
parameters in a very sensitive way, so that area-based methods
may be particularly affected. Feature-based are concerned too:
landmarks may simply not occur simultaneously in two SAR
images for example
Finally, local geometrical distortions provide an additional level
of difficulty in the registration problem. These distortions result
in data distribution variations, which has a disturbing effect
when the viewing geometry on a given area is very different
from one image to the other one.
As a conclusion, we can say that classical registration
approaches, whatever they are (area or feature based) have to
face to very tricky problems in SAR imaging because of noise,
distortion and complementary information.
In the next section, we introduce an alternative solution to the
SAR image registration problem that uses the Hough Transform.
Although it is not classically used for solving such kind of
problems, the Hough Transform brings a help to automatically
discriminate the common information from the complementary
one, and thus to have a common support for registering SAR
images that have been acquired using different parameters.
2.4 The Hough Transform
P. Hough introduced his transformation in the early 60s (in
1959) and then, as a US patent, in 1962 (Hough, 1962); in 1972,
R. Duda and P. Hart (Duda and Hart, 1972) showed this
transformation being efficient for the detection of lines and
curves in pictures; then, the Hough Transform has been
generalized to provide the detection of parameterized models
such as transformations, and not only geometrical primitives.