Image Processing Reference
In-Depth Information
Table 10.5
Performance measures for various techniques for visualization of the geological data
Fusion technique
Variance Entropy Avg gradient Relative Fusion
Fusion
bias
b
σ
2
H
g
¯
factor
FF
symmetry
FS
Bilateral filtering technique
417.27
6.69
5.98
0.21
1.58
0.19
Bayesian technique
427.74
6.17
5.57
0.10
2.05
0.38
Variational technique
509.15
6.52
5.45
0.11
2.00
0.21
Optimization-based technique
523.46
6.87
6.71
0.19
2.85
0.33
Three band selection
491.07
6.70
6.54
0.09
1.32
0.44
Piecewise linear function
325.52
6.23
5.61
0.24
1.71
0.38
Color matching function
282.37
5.28
4.68
0.23
1.75
0.34
the scene. Figure
10.5
a, c shows the results of fusion of the coral data using the
bilateral filtering-based technique and the variational technique, respectively. We are
able to observe certain features in the top right corner of the scene in both these
images. The optimization-based fusion brings out these details, but they are not very
prominent due to a relatively low contrast in color. In other results, these features
although present do not appear very prominent. The results of the PLF and the CMF
techniques are quite similar although the PLF provides a better discriminability of
features. However, the combination of bands chosen by the band selection technique
has picked up some noisy data which is visible in the form of horizontal strips. The
objective assessment of the coral data from Table
10.6
indicates high amounts of
contrast and sharpness in the result of the band selection technique.
This example indicates susceptibility of the no-reference measures to the noise.
This technique produces good result only when a large number of features exist in
the particular set of three chosen bands. This technique, however, may introduce a
huge shift in the mean value of the output from that of the input data, as indicated
by high values of the relative bias
b
for most of the datasets used in this chapter. The
Bayesian, variational, and optimization-based techniques exhibit high values of the
fusion symmetry (FS) that corresponds to a non-uniform participation from input
hyperspectral bands towards fusion. For the Bayesian technique, the participation of
a band is less if the values of the corresponding sensor selectivity factor
are small.
Finally, let us discuss the result of fusion of the urban hyperspectral data which we
have used for the primary illustrations of all the techniques in the respective chapters
of the monograph. In Fig.
10.6
, we have provided the same results along with the
results from other comparative techniques. These data contain a large number of
smaller objects which are easily discriminable in the fusion result of the bilateral
filtering-based technique as can be seen from Fig.
10.6
a. This technique has also
proved to be quite superior in the case of the Moffett Field datasets which contain
similar agglomeration of smaller objects. This superiority is due to defining the fusion
weights from the locally dominant features present at each pixel in each band of the
data. A Bayesian solution, too, examines every individual pixel in the data, and picks
up those which possess visually important information. The output of this technique
provided in Fig.
10.6
b is also similar to the one from the former technique. The
β
