Image Processing Reference
intermediate fused image obtained prior to inclusion of this band. The band selection
process now involves computation of the intermediate fused images whenever a new
input band is selected. This output-based band selection turns out to be computa-
tionally expensive. Unlike the first method of band selection, it selects only those
hyperspectral bands that contain significant additional information as compared to
the existing fusion output. For both band selection schemes, there is a very minimal
degradation in the quality of the resultant fused images even when one selects less
than 25% of the original data.
11.3 Future Directions
The focus of work in this monograph has been on fusion of hyperspectral images
and the performance evaluation of the fusion algorithm. We have explained in detail
various fusion techniques and each method has been shown to have its own merits
and demerits. Despite dealing with different useful topics in this monograph, the
research in hyperspectral image visualization is far from being complete. There are
several interesting aspects which can be considered for further research.
The edge preserving solution uses a bilateral filter for calculation of fusionweights.
The bilateral filter has two parameters to control the amount of filtering in the range
and spatial domains, respectively. The current solution assumes the values of these
parameters in a heuristic way. One would like to calculate these values from the
data in an adaptive manner. The values of the bilateral kernel parameters should
be based on the statistical and spatial properties of the input data.
Several improvements to the bilateral filter have been suggested, for example
a trilateral filter . It would be worth exploring whether such improvements
would help in extracting the minor features in the data in a better manner. Further,
anisotropic diffusion has been shown to be closely related to the mathematical
definition of bilateral filtering [11, 12]. Hence, one may explore the usefulness of
anisotropic diffusion instead of a bilateral filter.
The Bayesian solution involves a model for image formation. The computation
of the parameters of this model is based on multiple quality measures of a given
pixel. While our experiments have considered a couple of such measures, it will
be interesting to experiment with various other measures of image quality. The set
of computationally simple measures that can capture the visual information most
efficiently should be identified.
If a high value of performance measure is the key goal, it will also be interesting to
evaluate the performances of various desired quality measures by embedding them
directly into the cost function of the optimization-based solution. Since our fusion
solutions are intended for human visualization, designing of the performance crite-
ria functions that incorporate some the human visual system (HVS)-based quality
measures would be quite useful and relevant.
The Bayesian and the optimization-based techniques explicitly consider the spa-
tial correlation, i.e., the intra-band correlation. The consecutive bands of the