Performance Assessment of Fusion Techniques - Hyperspectral Image Fusion - page 146

large number of small rectangular objects, which were almost not seen in Fig. 9.5 a.
A resultant image from the fusion of 150 bands is shown in Fig. 9.5 c, which provides
a still sharper and clearer understanding of the scene contents. It can also be observed
that the subsequent image generated from fusion of the entire dataset as shown in
Fig. 9.5 d is not too different from Fig. 9.5 c. This indicates saturation in the visualiza-
tion process for the moffett 2 dataset for a given technique. The corresponding plots
of the asymptotic reference measures given in Fig. 9.2 are also in agreement with the
process of saturation.
It is hoped that the results of these and similar experiments can assist an observer in
quantifying the efficacy of these quality measures, and in analyzing the performance
of various fusion techniques. The discussed quality measures help in capturing the
complementary characteristics of a given technique and, thus, instead of relying on
the use of a single quality index, one should use a subset of these measures for the
selection of the most suitable technique.
9.5 Summary
This chapter defines the notion of consistency of a fusion technique that refers to
the asymptotic behavior of the technique with respect to a specific performance
measure. The analysis of the consistency can be useful to test the effectiveness and
applicability of the given technique for the fusion of a large number of images. Several
measures related to the consistency can be used to determine its suitability for fusion
of hyperspectral data. For the fusion techniques studied in this chapter, we found
that the band selection-based methods, including those of PCA-based methods, are
not consistent. However, the other methods such as the MRA-based and the bilateral
filtering-based methods are consistent.
We have also presented a detailed analysis of several performance measures as
regards the fusion technique. Certain modifications in some of these measures have
been discussed for a better analysis of the fusion technique, particularly extending
their applicability to the fusion of hyperspectral images where a very large number
of image bands are fused. Again, among the techniques analyzed here, the band
selection-based techniques appear to perform poorly in terms of the measures of the
resultant image compared to the MRA- and bilateral filtering-based techniques.