Image Processing Reference
In-Depth Information
Fig. 4.7
Plots of Bhattacharyya coefficient between the resultant image from the fusion of entire
data and the resultant images for various values of
κ
for the urban and the moffett
2
data using
the bilateral filtering-based fusion technique over the subsets selected using the input-based band
selection approach (© ACM 2010, Ref: [87])
the resultant image from fusion of the entire dataset is considered as the reference.
Figure
4.7
provides the plots of the Bhattacharyya coefficient (BC) between the
reference image and the resultant images obtained from the fusion of a subset of
image bands obtained for different values of
, selected using the entropy-based
scheme for the urban data as well as the moffett
2
data. The images have been fused
using the bilateral filtering-based technique. The Bhattacharyya coefficient (BC) can
be observed to steadily increase for increasing values of
κ
κ
. This behavior indicates
gradual deviation (in terms of histogram overlap) of the fused images from the
reference image as lesser number of bands get selected for fusion. The entire set of
hyperspectral images was selected for very small values of
, when the BC is zero,
as it can be observed from both plots. It may be observed from Figs.
4.4
,
4.5
and
4.7
that the performance of the band-selection method is very much data dependent, i.e.,
AVIRIS and Hyperion data having different sensor characteristics. However, it may
be observed from the nature of the plots that the band selection scheme performs
equally well on both datasets.
In the case of spectrally ordered hyperspectral data, the band selection scheme
reduces a significant amount of computational time, however it has an overhead of the
calculation of the conditional entropy of each image band. The band selection scheme
is beneficial only when the computation of actual fusion algorithm far exceeds the
computation of the entropy of each of the image band. However, except trivial and
simple techniques like averaging, for most of the existing robust fusion techniques
the time needed for band selection is much smaller than the time taken for fusion
of the entire dataset. Especially when the fusion techniques operate on a per pixel
basis, or they involve some kind of iterative routines, the computation of entropy
is quite negligible as compared to the computational requirements of actual fusion
κ
