Image Processing Reference
In-Depth Information
monograph is limited to remote sensing applications, more specifically hyperspec-
tral images. However, we also provide a brief description of fusion methodologies
pertaining to the other application areas.
Remote sensing has been one of the leading image fusion applications with a
large number of dedicated publications. The research in this area has been contin-
uously growing as more precise and sophisticated imaging devices are being used.
In 1999, Pohl and van Genderen have presented an in depth review of the existing
work in multisensor image fusion till then [140]. This work covers a comprehensive
range of fusion techniques with their objectives. It covers basic arithmetic tech-
niques such as addition or ratio images, computationally demanding subspace-based
techniques based on the principal component analysis (PCA), and wavelet-based
multi-resolution techniques as well. This article introduces a number of applications
of fusion including topographic mapping, land usage, flood monitoring, and geology.
Furthermore, some pre-processing techniques and commonly used fusion schemes
have also been reviewed.
Since a number of groups are actively working in the area of image fusion, the
meaning and taxonomy of the terms differ from one to another. The establishment
of common terms of reference helps the scientific community to express their ideas
using the same words to the industry and the other collaborative communities. Wald
presented a report on the work of establishment of such a lexicon for data fusion in
remote sensing carried out by the European Association of Remote Sensing Labora-
tories (EARSeL) and the French Society for Electricity and Electronics (SEE) [180].
According to this definition, data fusion is a
framework
containing means and tools
for alliance or combination of data originating from different sensors. While fusion
has been aimed at obtaining information of greater quality, the quality itself has been
associated with the application. The definition of the word
fusion
has been compared
with those of integration, merging, and combination. It has been suggested that these
terms are more general, and have a much wider scope than fusion. It has also been
argued that the term pixel does not have a correct interpretation as the pixel is merely
a support of the information or measurement. The suggestions include usage of terms
signal or measurement to describe the level of fusion. In this monograph, however,
we use the term
pixel-level
to describe the same nomenclature as it has been followed
by a large community.
Now, let us define a generalized model for basic pixel-based image fusion. Let
I
1
and
I
2
be two images of dimensions
pixels having the same spatial
resolution. Then the resultant fused image
F
is given by Eq. (
2.1
).
(
X
×
Y
)
F
(
x
,
y
)
=
w
1
(
x
,
y
)
I
1
(
x
,
y
)
+
w
2
(
x
,
y
)
I
2
(
x
,
y
)
+
C
.
(2.1)
The quantities
w
1
and
w
2
indicate the relative importance assigned to the corre-
sponding pixel at
of that image, and these are known as the fusion weights,
or simply weights. As we want the fused image to be composed of the constituent
images
I
1
and
I
2
, we are essentially looking for an
additive
combination of these
images. The fusion weights should also be ideally non-negative in nature. Addition-
ally, the weights are normalized, i.e., the sum of all the weights at any given spatial
(
x
,
y
)
