Image Processing Reference
In-Depth Information
Fig. 6.1
Results of fusion for
the variational approach for
the urban image data from the
Hyperion.
a
Grayscale fused
image, and
b
RGB fused
image
(a)
(b)
First we consider the urban hyperspectral image which depicts some urban regions
of Palo Alto, CA in the form of contiguous 242 bands. The dimension of each band
is
pixels. For fusing the entire hyperspectral image into a grayscale
output, the entire set of input bands is processed in a single step. The local variance
of the data needs to be computed only once, which remains constant throughout
the fusion process. The weight requires re-computation at every iteration as it is
dependent on the newly formed (intermediate) fused image. Figure
6.1
a depicts the
result of fusion using the variational solution. This image represents fusion over the
spectral bandwidth of nearly 2200 nm. The procedure to generate the RGB fused
image remains same as the one employed for the other two methods presented in the
earlier chapters. We partition the input hyperspectral data into 3 subsets of contiguous
bands, each nearly of the same cardinality, i.e., each subset contains nearly
1
(
512
×
256
)
3-rd of
the hyperspectral data. These three subsets are then fused independently using the
variational technique to yield three corresponding fused images. The RGB composite
image is formed by assigning these fused images to the red, green, and blue channels
of the standard display. The corresponding RGB image is shown in Fig.
6.1
b. Both
images in Fig.
6.1
appear visually smooth, and do not produce any visible artifacts.
However, at certain places, the edges and boundaries of various objects in the data
appear less sharper. These features represent the high frequency components in the
image which get attenuated as the regularization term [Eq. (
6.9
)] penalizes sharp
changes or high frequency components in the output.
We have used themoffett
2
hyperspectral image from the AVIRIS as the second test
data. The grayscale output image obtained by the variational method for the moffett
2
data is shown in Fig.
6.2
a. This image is the result of combining 224 hyperspectral
/
