Image Processing Reference
In-Depth Information
The non-negativity of the weights provides a sufficient condition for the fused
image F
(
,
)
to be also non-negative.
Throughout the fusion procedure, the input hyperspectral data is assumed to be
normalized such that, 0
x
y
. In the present work, we are deal-
ing with the visualization-oriented fusion of hyperspectral images, where a human
analyst will observe the contents of the scene. While almost all of the elements of
image quality are desirable in a single image—this methodology does not focus on
the enhancement of spatial or radiometric resolution. The focus is on other elements
that one would like a fused image to possess. As high values of contrast make visu-
alization appealing and clear, it is one of the desired properties of the fused image.
This is helpful for some of the post-fusion operations such as object detection and
segmentation apart from an enhanced visualization. A high contrast can be obtained
from the pixels that are far apart in terms of their gray values. However, this may lead
to a high amount of saturation (over- or under-exposedness) creating too bright or
too dull regions in the fused image. This leads to loss of information, and deteriorates
the quality.
The fused image is to be obtained from the hyperspectral bands which collectively
encompass a large bandwidth. The reflectance response of the scene, which is a
function of wavelength, also varies significantly over the set of captured bands.
Through fusion algorithm, we want to map pixels from this large dynamic (intensity)
range into a smaller range of the display system. This mapping can be accomplished
using two different strategies:
I k (
x
,
y
)
1
(
x
,
y
)
S-I : Most commercial displays are capable of providing 256 gray levels (which
can be normalized into a range [0, 1]). The images with higher dynamic range
may extend well beyond this range. If the pixel intensities beyond the maximum
possible display gray level are clipped in order to accommodate into the available
8-bit display, several regions in the scene appear over-exposed or over saturated.
The information within these regions gets lost. On the other extreme, when the
pixel values lesser than theminimumdisplay intensity are clipped, the scene depicts
many under-exposed areas. These appear dark and dull, and thus, are not much
useful for observation or any further processing. We can map the pixels from the
input toward the central region (0.50) of the dynamic range of the gray values
around [0, 1]. This mapping helps in bringing the under and over exposed regions
in the scene into the mid-range of the display. The minimization of the distance of
gray values from the middle of dynamic range has, in fact, been used as one of the
objectives for an enhanced visualization of the image [137]. With this mapping,
the dynamic range of the fused image directly gets mapped to the mid-range of
the display device. Thus, the typical requirements of further post-processing such
as non-linear stretching for display are often not necessary. This is preferred in
computer graphics applications where the primary need is to provide a good visual
display.
S-II : The earlier strategy ( S-I ), although suits very well for pure display appli-
cations, it does significantly alter the relationship between the average intensity
of the input hyperspectral image and the output fused image. One may expect
 
Search WWH ::




Custom Search