Image Processing Reference
Fig. 6.2 Results of fusion for the variational approach for the moffett 2 image from the AVIRIS.
a Grayscale fused image, and b RGB fused image
bands each of which has a nominal bandwidth of 10 nm. The RGB version of the
fusion output is obtained by independently fusing three nearly equal subsets of the
input hyperspectral data, and then assigning it to the different channels of the display.
As previously stated, we have followed a simple strategy for partitioning the set of
hyperspectral bands, that is, partition the sequentially arranged bands into nearly
equal three subsets. The RGB version of the result for the moffett 2 data is shown in
Fig. 6.2 b. One may notice that the fusion results are quite good, although some loss
of details may be seen at small bright objects due to smoothness constraint imposed
on the fused image. For improved results, one may use a discontinuity-preserving
smoothness criterion as suggested by Geman and Geman . However, this would
make the solution computationally very demanding and hence, is not pursued.
A technique for fusion of hyperspectral images has been discussed in this chapter,
without explicit computation of fusion weights. The weighting function, defined
as a data dependent term, is an implicit part of the fusion process. The weighting
function is derived from the local contrast of the input hyperspectral bands, and is also
based on the concept of balancing the radiometric information in the fused image.
An iterative solution based on the Euler-Lagrange equation refines the fused image
that combines the pixels with high local contrast for a smooth and visually pleasing
output. However, at some places, the output appears oversmooth which smears the
edges and boundaries.
The solution converges to the desired fused image within a few iterations. How-
ever, it requires computation of the implicit weighting function at the end of every
iteration which makes it computationally expensive as compared to the techniques
described in the previous chapters.