Image Processing Reference
hyperspectral data also exhibit a very degree of correlation among them, i.e.,
the spectral or the inter-band correlation. We have not explicitly considered this
factor which may lead to some further improvement in the results.
The variational technique employs the local variance measure for computation of
contrast at a pixel level. One may want to replace it by other measures such as
spatial gradient which may be more effective, and faster in computation.
We have discussed a framework for the consistency analysis of a fusion technique,
and demonstrated the same through experimentation. However, the task of mathe-
matically proving that a given fusion technique is indeed consistent still remains.
Based on the consistency analysis, we would like to develop a fusion technique
that will be consistent with respect to a given performance measure. With a wide
choice of performance measures available, these experiments might provide a
variety of solutions along with some interesting insights towards the concept of
image fusion. It should also be noted that since the result of fusion depends on
the input data, proving mathematically that a fusion technique is consistent for
a given performance measure for all possible datasets (i.e., data independence)
will be very difficult. If an appropriate statistical model of the hyperspectral data
cube is assumed, it may be possible to prove the consistency of a fusion scheme
In hyperspectral imaging, the spectral signatures of various types of materials on
the earth surface are often available. Would such a prior knowledge help in the
fusion process?Our understanding is that it may be possible to use such information
for a better quality image fusion. However, this would require a segmented map
of the region on the earth surface being imaged, when one may have to combine
the task of scene classification and image fusion into a single problem definition.
Since hyperspectral data cube has a very large number of image bands with a high
amount of correlation along the spectral axis, we have discussed two different
ways by which a large number of constituent bands can be eliminated while fusing
the data. However, both these approaches are greedy in nature. The first band is
trivially selected. It is quite possible that, for a given fraction of the total number
of bands selected, a different combination of image bands may provide a better
visualization. It may be worthwhile to explore how one can obtain the solution
that provides the best visualization.
In this monograph several different methods of image fusion have been discussed,
and their performances have been compared. While closing, it is quite natural to
ask which specific method is being prescribed to solve the general problem of
hyperspectral image fusion? It is, indeed, quite difficult to answer this question.
None of the methods yields the best scores for all performance measures defined
in this monograph. However, it does appear from the results given in the previous
chapter that the optimization-based technique tends to yield better scores for a
majority of these performance measures. However, it involves more computations
which may limit its applicability for fast visualization of the hyperspectral data
cube. Fortunately, a large amount of such computations can be done in parallel.
It would be interesting to have this technique, and possibly other techniques too,
implemented on a GPU to speed up the computation.