Image Processing Reference
In-Depth Information
Chapter 4
Band Selection Through Redundancy
Elimination
4.1 Introduction
Pixel-based image fusion techniques compute the fusion weights for each pixel
in every band. Thus, every band in the hyperspectral data is subjected to the
process of computing weights through some kind of saliency measure which is a
time-consuming operation. Let us divide the process of fusion into two steps- first,
computation of fusion weights through saliency measurement, and second, a linear
combination of the bands using these weights. The later step is quite trivial, and
requires a very little computation. The first step, however, is critical to the perfor-
mance of the algorithm, and can be computationally demanding as per the chosen
algorithm. While some attempts towards a quick visualization of the image contents
in the form of an RGB image have been investigated, some of these methods typically
include selection of three image bands satisfying certain criterion [49, 79]. However,
these approaches select only three bands for the display, and they do not involve any
kind of image merging or fusion at any level.
For the data (or pixel) level image fusion, most available techniques evaluate the
importance of a pixel within its spatial neighborhood, and then assign appropriate
weights to the pixels while fusing them over various image bands. For example, in
the previous chapter, the residual image after subtraction of the image band from
its bilateral filtered version provides the necessary fusion weights. Since this step
consumes the major amount of computation, usually on a per pixel basis, the time
taken for fusion is directly proportional to the number of image bands. An observer
has to wait until the completion of calculation of weights for the entire set of image
bands, followed by the successive weighted addition to get the final result of fusion.
Therefore, the fusion techniques tend to be slower due to the large number of image
bands. The process of calculation of weights can be computationally quite demand-
ing and time-consuming when the fusion techniques and /or the saliency detection
techniques are iterative in nature. This often limits us from exploring some sophis-
for hyperspectral image fusion.
