Image Processing Reference
In-Depth Information
Table 9.1
Performance measures for various techniques for visualization of the moffett
2
data
( c
2011 Elsevier)
Fusion technique
Relative bias
Fusion factor
Fusion symmetry
b
F F
F S
Bilateral filtering technique
0.22
1.62
0.24
Three band selection
0.46
1.01
0.66
Piecewise linear function
0.20
1.64
0.29
Color matching function
0.22
1.64
0.26
MRA technique
0.24
1.18
0.18
details is quite significant for this dataset and this affects the performance. However,
as the number of bands progressively increases, the averaging effect helps reduce
the fluctuations. The spatial domain-based techniques do not suffer much from this
problem. Figure
9.4
focuses on the consistency analysis of the fusion techniques.
The plots of the Bhattacharyya distance and the Jensen-Shannon distance from
Fig.
9.4
a, b, respectively, appear monotonically decreasing, finally converging to
zero. These plots, thus, affirm the consistency of the techniques (except the band
selection technique), and also indicate their suitability for fusion of a very large
number of images. Thus, the techniques proposed in [79, 88] prove to be quite suit-
able for the fusion of hyperspectral images.
The plots for various measures with an asymptotic reference do not provide any
numerical values for comparison of fusion techniques. However, these plots are
quite useful in understanding the impact of subsequent image bands in the process of
fusion. Ideally, we want the fusion process to gradually converge over the entire set of
hyperspectral bands. In this case, the corresponding plot has a nearly uniform slope.
A plot with very high slope on the either end points out negligible contribution of
image bands on the corresponding other end towards the final result. While we have
discussed the ordered hyperspectral data where the contents of the bands change
gradually, an interesting case would be to observe the effect of fusing the bands
in a random order. Through the random permutation, we destroy the inter-band
correlation characteristic of the hyperspectral data. If the image bands are randomly
permuted and then combined, the corresponding plots will tend to saturate quite
quickly as the subsequent bands are no longer highly correlated and the process
starts seeing the redundancy in bands being added later.
The performance of a fusion technique also depends on the dataset. This can be
illustrated easily from these two test datasets. A comparison of Figs.
9.1
-
9.4
suggests
that all these plots tend to saturate quite rapidly for the Hyperion data compared to
the AVIRIS data. In the case of geological Hyperion data, one recovers very little
additional information after fusing around 50-60 bands The MRA and the piecewise
linear function-based techniques were found to saturate at a slower rate compared to
the bilateral filtering-based technique from Figs.
9.1
and
9.2
. For the AVIRIS data,
the process does not saturate even after fusing 120-130 image bands. This brings
out the data dependencies of the performance of any fusion technique.
