Image Processing Reference
In-Depth Information
fusion is to combine the information across multiple bands in order to increase the
classification accuracy of the so obtained resultant image. We often come across
images of the same scene with different spatial resolution. The spatial resolution
refers to the size of the smallest discriminable object in the image. Images with
higher spatial resolution are desirable, however they may not be always available
due to high cost involved in such imaging sensor design. The issue of improving
the spatial resolution is often solved through fusion of low-resolution images with
the high-resolution image of the same scene. In remote sensing, this fusion is also
known as the pan-sharpening which explores the sharpening objective of fusion. We
on fusion of hyperspectral data purely for the purpose of visualization intended for
a human observer.
As we traverse along the spectral dimension, different values of pixel intensity can
be observed at the same spatial but different spectral locations.We need to incorporate
these values into the fused image according to their relative importance. Images
with a high value of contrast and sharp features appear visually pleasing. However,
images with a large number of over- and under-saturated pixels do not possess much
visually informative content. As visualization is the focus of this monograph, we
want the fused image to have these desirable characteristics. While satisfying these
requirements, the fusion procedure should not introduce any visible artifacts. The
problem of visualization of hyperspectral data over a standard display device is thus,
quite a challenging problem as hundreds of bands together encompass a large volume
of disparate information.
As the theory of image fusion began to develop, it was highly important to mea-
sure the performance of such fusion systems. This problem becomes challenging
when the reference image or the ground truth is not available. Researchers have
proposed a few measures based on the statistical information such as entropy of the
image. Most of these measures have been defined for a generalized fusion of images
where the number of images being fused is very less, say 2-6, and thus, the calcula-
tions of statistical performance measures are easily implementable. The problem of
evaluation of fusion is more difficult in the case of hyperspectral image fusion due
to a large number of bands, high volume of independent information among them,
and unavailability of any reference. The existing measures of image fusion should
be modified specifically for the fusion of a very large number of bands. We shall
explain several adaptations of existing measures to analyze the fusion techniques for
hyperspectral images in a better manner.
1.3 Tour of the Topic
Thismonograph addresses the problemof fusion of hyperspectral image bands for the
purpose of visualization. We do not assume any knowledge of hyperspectral sensor
parameters and scene contents. Having obtained the fused images, the problem of
performance evaluation of the fusion techniques is quite challenging, particularly
