Image Processing Reference
In-Depth Information
Fig. 4.5
Plots of PSNR of the fused images for various values of
κ
on the Hyperion and the AVIRIS
data using the bilateral filtering-based fusion technique over the subsets selected using the input-
based approach. The resultant image from the fusion of the entire dataset is used as a reference in
each case for comparing the PSNR value (© ACM 2010, Ref: [87])
error for the fused image with 0
1 is calculated in terms of PSNR values.
The nature of the PSNR values of the fused images over the subset of bands selected
using the entropy-based technique can be observed in Fig.
4.5
. The fused images have
been obtained using the bilateral filtering-based fusion technique over the subsets
of bands for various values of the threshold parameter
<κ<
. We assume the resultant
image obtained through fusion of the entire dataset using the same fusion technique
as the reference. As all the image bands of the moffett
2
data were chosen for fusion
for the values of
κ
κ
up to 0.15, the fusion results obtained are exactly same as the
reference image. The PSNR values decrease as
increases indicating fewer bands
are being selected. It can be seen that the resultant image with the fusion of 48 bands
(
κ
40) gives visually almost similar results with PSNR being above 40dB. This
result, in fact, has been formed from fusion of merely
1
κ
=
0
.
/
4
-th of the original set of
hyperspectral bands. A similar nature of PSNR plot can be observed for the urban
data in Fig.
4.5
, where a bilateral filtering-based technique has again been used for
fusion. A resultant image obtained from the fusion of a subset of 69 bands chosen
by selecting
κ
to be 0.45 was found to provide a PSNR of nearly 35dB. For the
values of
less than 0.20, the band selection scheme selects all the image bands in
the Hyperion dataset, and produces the same final result of fusion as the reference
image. Thus, for both datasets, the band selection technique proves to be capable of
selecting small, yet efficient subset of hyperspectral bands that can produce visually
almost comparable results even when a small fraction of the original data is selected.
A visual comparison of the fusion results of the entropy-based band selection
can be carried out from Fig.
4.6
where the resultant images obtained from the fusion
of selected subsets of the AVIRIS data using the bilateral filtering-based fusion
technique are shown. The images in Fig.
4.6
a-d are obtained from the subsets of
κ
