Performance Assessment of Fusion Techniques - Hyperspectral Image Fusion - page 137

9.4 Experimental Evaluation
In this section let us study the performances of some of the representative fusion
techniques. We illustrate the usefulness of the consistency and performance measures
over some of the commonly available fusion techniques and some of the methods
discussed in this monograph. Although we have provided some literature survey in
Chap. 2 , we provide in very brief basic concepts behind some of the other fusion
techniques being studied in this section. For more details on the other techniques,
readers are requested to refer toChap. 2 , or the corresponding references cited therein.
The simplest technique for displaying the contents of a hyperspectral image is to
select three specific image bands from the entire set of image bands in the dataset and
assign them to the red, green, and blue channels to form an RGB composite image
of the scene. Although this technique does not involve actual fusion of the data, this
technique is used to obtain a fast and quick rendering of the data contents (as used
by the image browser for AVIRIS). We include this technique for analysis due to its
technical and computational simplicity. We shall use the 3-band selection scheme as
presented in [49].
We also consider the technique of color matching functions which derives the
fusion weights for each of the bands from the hypothesis of how the synthesized
image would be seen to the human eye if its range of perceived wavelength were
stretched to the wavelength range of the hyperspectral data of interest. The weights
here specify the amount of primary colors to be mixed which will create the sensation
of the same color as that of viewing the original hyperspectral spectrum [78].
Another technique subjected to performance evaluation in this chapter uses a
similar methodology of envelopes as explained for the previous technique of the
color matching functions. These envelopes are the piecewise linear functions over
the spectral range in order to map deep infrared spectra to magenta hues [79].
The multi-resolution analysis (MRA)-based techniques have proved to be very
useful in the field of image fusion. Here, we consider anMRA-based technique where
each of the constituent image bands is decomposed to extract directional gradient
information at each level [189].
We also analyze the edge-preserving technique for fusion of hyperspectral image
described in Chap. 3 . This technique uses an edge-preserving bilateral filter for the
calculation of weights for each of the pixels in each of the image bands [88]. This
bilateral filtering-based technique has been discussed in Chap. 3 of the monograph,
which happens to be the computationally fastest among all the techniques discussed
in this monograph. The choice of the above set of fusion techniques is based on the
fact that these are some of the recent techniques for pixel-level fusion of hyperspectral
data and that together they span almost all different methodologies of image fusion,
such as subspace-based, MRA-based, and local contrast-based techniques; along
with the representative technique of band selection for display. While some of these
techniques are capable of producing RGB outputs, we deal with the grayscale results
of fusion in order to maintain uniformity in the evaluation.