Image Processing Reference

In-Depth Information

Table 10.6
Performance measures for various techniques for visualization of the coral data

Fusion technique

Variance Entropy Avg gradient Relative Fusion

Fusion

bias
b

σ

2

H

g

¯

factor
FF
symmetry
FS

Bilateral filtering technique

1195.82

6.98

6.97

0.21

1.36

0.23

Bayesian technique

1435.03

7.09

7.07

0.26

1.35

0.52

Variational technique

1087.80

6.91

6.73

0.18

1.43

0.51

Optimization-based technique

984.32

6.82

6.87

0.24

1.65

0.61

Three band selection

884.82

6.86

6.71

0.79

0.38

0.14

Piecewise linear function

871.78

6.75

5.83

0.25

1.40

0.42

Color matching function

534.30

6.00

4.70

0.24

1.34

0.19

The details of the quantitative assessment of the results of the urban data are

provided in Table
10.7
. We may also observe high values of the variance and the

average gradient for all the presented techniques as against the corresponding values

for the other techniques described here except possibly for the variational technique.

The smoothness constraint in this case helps producing smooth and thus visually

pleasing images, but at the cost of reduced sharpness. The contributions from the

input bands are also quite high for all the discussed techniques as observed from

the fusion factor FF. Fusion using the PLF and the CMF have captured some of

the subtle features from the data, however, their results appear quite dull, and thus

lack visual appeal and the clarity of details. As per Table
10.7
, all techniques have

performed well in terms of low values of the relative bias
b
with the variational

technique providing the smallest value indicating the least deviation from the input

data with respect to the mean intensity value. Also, low values of fusion symmetry

FS indicate high level of uniformity in the participation from various hyperspectral

bands towards the fusion output for all techniques studied in this chapter.

We now discuss the computational requirements of these fusion techniques before

we conclude this chapter. We have implemented all the algorithms in

®

version R2009a running on a P-IV computer running at 2.66 GHz with 2 GB RAM.

Out of four techniques discussed in the monograph, the bilateral filtering-based one

is a non-iterative single pass procedure. The technique initially produces a bilaterally

filtered output of every band in the data. These outputs are used to compute the fusion

weights while the actual fused image is produced via a normalized sum of bands.

We have used the fast implementation of bilateral filter as described in [126]. The

moffett
2
,moffett
3
, and lunar datasets provided by the AVIRIS contain 224 bands

of dimensions

Matlab

pixels each. For the aforementioned computer system,

the bilateral filtering-based solution took nearly 120 s to generate the fused image

geological, coral, and urban datasets contain 242 bands of the size

(

614

×

512

)

(

×

)

pixels

each, provided by the Hyperion. The proposed solution took nearly 27 s to generate

the fused image with 3 stages of hierarchical processing.

The Bayesian technique can be split into two parts. The first part deals with

the computation of the sensor selectivity factors

512

256

β

over every pixel in the data.