Image Processing Reference
In-Depth Information
images [150]. PCA transforms a set of intercorrelated variables into a set of new
(potentially) uncorrelated linear combinations of new variables. The computation of
principal components (PCs) of a 2-D image involves calculation of the eigenvectors
of its covariance matrix. The PCA fusion procedure integrates the disparate natures
of multisensor image data into one single image. If we assume the basis vectors,
or equivalently the transform coefficients to be statistically independent, then we
can apply the independent component analysis (ICA) to handle the fusion problem.
In [116], the authors apply a DT-CWT technique of fusion in the bases constructed
using ICA. Some methods of PCA-based hyperspectral image fusion are discussed
in the next section.
Till now, we have discussed a number of fusion techniques mainly related to
remote sensing applications. However, the spatial resolution of all input images
undergoing fusion has been exactly the same. We explore in brief an important field
of fusion that deals with images of different spatial resolution. In remote sensing,
some of the satellites provide two types of images: a single band panchromatic
(pan) image that represents the scene at a high spatial resolution, and a low resolu-
tion multispectral image containing a set of few bands. For example, the QuickBird
imaging sensor provides a single band pan image of spatial resolution of 0.6m, and
a multispectral data containing 4-bands with a spatial resolution of 2.4m. Simi-
larly, the IKONOS captures panchromatic images of spatial resolution of 1m, and a
4-band multispectral image with 4m spatial resolution. The process of combining a
set of low spatial resolution multispectral image with a co-georegistered high spa-
tial resolution panchromatic image enhances the quality of multispectral image by
increasing its spatial resolution. Fusion of pan and multispectral data produces a
high-resolution pseudo-color image which preserves most of the attributes of both
the image types, i.e., the finer spatial features of the pan image, and the spectral signa-
tures from the multispectral bands. The fused image is, thus, spatially and spectrally
enhanced, and hence visually appealing. This process which refers to the sharpen-
ing of a multispectral image using a panchromatic image is known as panchromatic
sharpening, or simply pan-sharpening.
The intensity-hue-saturation (IHS) color space is considered to be quite useful
for describing perception of colors to humans. The intensity represents the amount
of brightness, the hue component refers to the color, and the saturation describes its
purity. The appearance of the pan image is close to the intensity band of an IHS repre-
sentation of the scene. Therefore, during fusion (pan-sharpening), the multispectral
image is projected onto the IHS color space, and the intensity band is replaced by
the pan image. The fusion output can then be obtained by the reverse IHS transform.
A principal component analysis (PCA) transforms the intercorrelated data into a set
of uncorrelated components. The first principal component (PC) can also be consid-
ered to be close to the pan data. The reverse transformation after replacing the first PC
with pan data yields the sharpened multispectral image. The IHS- and PCA-based
pan-sharpening techniques have been proposed in the early 1990s (cf. [33]). The
contents of the pan and multispectral images are often dissimilar to a certain extent.
Furthermore, these data differ in terms of radiometry (local mean) [61]. When these
data are fused using the IHS-based methodology, it introduces a significant color
