Image Processing Reference
In-Depth Information
hierarchical implementation. However, the successive bands in the hyperspectral data
exhibit a high degree of similarity. When the randomly arranged bands are grouped
for the first level of hierarchical implementation, the corresponding fusion results
contain integration of highly dissimilar information as compared to the sequential
case.
The advantages of the hierarchical scheme are as follows:
•
This scheme reads only a fraction of the hyperspectral image bands at a given
time to generate an intermediate fusion result. The hierarchical scheme requires
only a small subset of the input to be read in the memory and process it. Thus, the
memory requirement is significantly reduced.
•
It enables the user to scale up the system to efficiently fuse any number of images.
It can easily accommodate any increase in the number of bands without compro-
mising with the performance.
•
The system is computationally efficient, and makes it possible to parallelize the
implementation. All the subsets of the hyperspectral image can be processed in
parallel to produce a set of intermediate fusion results.
•
The resultant images at the intermediate stages facilitate analysis and visualization
of midband reflectance response of the scene. The fused images at the first-stage
represent the response of the scene over a bandwidth that is
M
times that of an
individual hyperspectral band. These and the subsequent intermediate results can
be used to visualize the combined response of the scene over a range of bandwidth
encompassed by the number of bands being fused.
3.6 Implementation
The fusion procedure requires selection of three parameters-
σ
R
, and
C
.The
choice of these parameters is important to achieve a better quality of the output.
The implementation of the bilateral filter is also an important parameter as far as
the computational complexity and timing requirements of the entire procedure are
concerned. Here we use the implementation based on the approximation of bilateral
filter provided in [126]. In order to automate the fusion algorithm without much
degradation in quality, we have adopted the guidelines suggested in [126] to select
the values of first two parameters.
σ
S
,
σ
S
=
C
1
×
min
(
X
,
Y
)
(3.7)
C
2
×
max
))
σ
R
k
=
(
I
k
(
x
,
y
))
−
min
(
I
k
(
x
,
y
∀
k
,
(3.8)
where
C
1
and
C
2
are positive real numbers which can be set to obtain the desired
quality output. The choice of
C
1
is related to the size of spatial details retained during
fusion. We have used
C
1
=
16 in all test experiments. The value of the range kernel
defines the minimum amplitude that can be considered as an
edge
.Wehaveset
C
2
to 0.05 in our experiments.
1
/
