Image Processing Reference
In-Depth Information
due to a large number of constituent bands, and unavailability of any reference
image. We also address the problem of evaluation of different techniques for fusion
of hyperspectral data. The highlights of the topic are discussed below.
First, we provide an overview of the process of image fusion. We familiarize the
readers with various concepts related to image fusion. Then we discuss several
techniques of generalized image fusion, and specific techniques of hyperspectral
image fusion. We also brief the readers with some of the performance measures
for quality assessment of fusion techniques.
For an effective visualization of the entire hyperspectral data cube, it is natu-
rally expected that the fused image should retain as many features of the data
as possible. One would especially like to preserve the weak features of the data
during the process of fusion for a better visualization of the scene. Most of the
fusion techniques generate the resultant image as a weighted linear combination
of the set of input hyperspectral bands where the choice of weights is the most
critical aspect. We discuss a methodology that assigns appropriate weightage to
the weak textures in the data so that these can be prominently represented in the
fused image. We explain the use of an edge-preserving filter known as bilateral
filter to extract the textural content of the hyperspectral data. We define the fusion
weight for every individual pixel based on the amount of textural content at that
location obtained as the difference between the original pixel and the output of the
bilateral filtering at the same location. Through this weighing scheme, we assign
higher weights to the weak edges and features in the data that exist over a very
few bands as they might get lost in the process of fusion otherwise.
Fusion of hyperspectral data involves processing over nearly 200
+
bands. For
processing all the bands together, one is required to read the entire hyperspectral
data cube into the memory. Due to a high volume of the data, the memory require-
ment often goes beyond a few hundreds of megabytes. Furthermore, when a large
number of bands are being used together to assign weights, some pixels might get
assigned with very small values even comparable to the truncation limits of the
storage system. This leads to a risk of loosing the contribution of some of the pix-
els towards the final result. We discuss a hierarchical scheme of fusion to prevent
the aforementioned problems of hyperspectral image fusion. In this scheme, we
split the hyperspectral data into several subsets where each of the subset contains
nearly 10% of the original number of bands. We fuse these subsets independently
using the bilateral filtering-based solution. Since only a smaller chunk of the data
is required, the problems of memory requirement and smaller weights are circum-
vented in this scheme. Also, the fusion of each subset is independent of other, which
provides a scope for possible parallelization in order to speed up the entire process.
Most pixel-based fusion techniques compute the fusion weight for every pixel in
every band of the input hyperspectral image. The adjacent bands in hyperspectral
image exhibit a very high degree of spatial correlation as they depict the reflectance
response of the scene over contiguous wavelength bands. The successive bands,
thus, contribute a very little additional information towards the fusion process.
Therefore, the addition of such bands brings a nominal information gain to the
