Image Processing Reference
In-Depth Information
The first technique is based on an edge-preserving filter known as the bilateral filter
for the calculation of the
α
-matte. The hyperspectral bands contain a large number of
objects and features. Some of the features are weak, and visually apparent over only
a very small number of bands. The use of a bilateral filter enables us to extract these
weak and minor features from the hyperspectral data which might get lost during the
process of fusion otherwise. The fusion weights are calculated from these extracted
features in order to accentuate their representation in the fused image. This technique
uses the edge-preserving property of the bilateral filter in order to define the fusion
weights which are functions of the locally dominant features within the bands. This
technique exploits all available information at all locations in all bands, unlike the
methods that assign the same weight for the entire band. This solution turns out to
be very fast with the availability of faster implementations of the bilateral filter.
We have also explained a hierarchical scheme for an efficient implementation of
the bilateral filtering-based solution. This scheme can accommodate any increase in
the number of bands with a minimal degradation in the quality of the fused image.
This implementation operates over only a fraction of data at any given instant, and
therefore, does not require the entire data to be loaded into the memory. This scheme
allows fusion of bands up to any given spectral bandwidth for the midband visualiza-
tion which can be useful if the features from a certain set of bands need highlighting.
The hierarchical scheme is generic, and therefore can be suitably applied over any
pixel-based fusion techniques.
We have then developed a Bayesian framework for fusion of hyperspectral data.
Here, fusion has been posed as an estimation problem where the fused image has
been regarded as the
true scene
to be estimated. This fusion solution employs a
model of image formation which relates the fused image with the input hyperspectral
data through a first order approximation. We discussed an approach to determine the
parameters of this model which indicate the closeness of the input data from the fused
image to be estimated. When the fused image is meant for visualization, one expects
a higher contribution towards fusion from the pixels that possess a high amount
of visually important information. This fusion technique, thus, utilizes some of the
characteristics of the input hyperspectral bands to determine the parameters of this
model. It considers the well-exposedness and the sharpness of pixels as the quality
measures for the computation of these model parameters. Like the first solution,
this solution also operates on a per pixel basis in order to exploit the entire data.
The fused image is then estimated using the MAP framework. The prior term in the
MAP formulation has been a total variation (TV) norm-based penalty function to
acknowledge the intra-pixel spatial correlation in natural images. The advantage of
the TV norm-based penalty lies in its discontinuity preserving characteristic which is
responsible for producing the resultant image with a high degree of sharpness. Apart
from the output quality, the Bayesian fusion technique provides other advantages
like flexibility in the computation of the corresponding model parameters through
the choice of various image quality measures, and the availability of a large number
of tools from the estimation theory.
The variational technique developed for fusion of hyperspectral images elimi-
nates the process of generating fusion mattes. The fusion weights which act as the
