Results and Discussions - Hyperspectral Image Fusion - page 167

Table 10.8 Average computation time requirements for various fusion techniques
Fusion technique AVIRIS dataset ( s ) Hyperion dataset ( s )

(614 × 512 × 224) (512 × 256 × 242)

Bilateral filtering technique 120 27

Bayesian technique 615 208

Variational technique 712 273

Optimization-based technique 740 281

10.4 Remarks

We have provided results of fusion using four different techniques discussed in
the monograph, along with some other recent and commonly used techniques for
visualization of hyperspectral data. We have provided the resultant fused images over
several hyperspectral datasets, and also provided several quantitative performance
measures for the same. From both the visual and the quantitative analysis, it can
be inferred that all the presented techniques yield better quality fused images. The
bilateral filtering-based technique identifies the locally dominant features in the data,
and defines the fusion weights at every pixel location based on those features. The
edge-preserving property of the bilateral filter helps generating the fused image
without any visible artifacts. The fused image possesses a high amount of contrast
and sharpness imparted by the dominant local features. This solution has low values
of mean bias b as the resultant image does not deviate much from the radiometric
mean of the input data. High values of the fusion factor FF, and low values of the
fusion symmetry FSindicate higher and yet uniform contribution from the constituent
spectral bands towards the fusion process.
The Bayesian fusion technique determines the visual quality of each pixel in
the data based on two different quality factors-well-exposedness and sharpness. As
the pixels with higher values of these factors contribute more towards fusion, the
resultant images possess high amount of sharpness and visual clarity. However, this
leads to a somewhat uneven contribution of spectral bands towards the final fused
image, which results in high values of the fusion symmetry FS. The TV-norm based
regularization prevents smoothening of edges and other sharp discontinuities, and
yet maintains the naturalness of the fused image.
The matte-less technique combines the input hyperspectral bands without explic-
itly computing the fusion weights. The variational framework incorporates the
weighting function as a part of the cost function. Through an iterative procedure,
this technique generates a fused image with a very small deviation from intensities
of the input bands. Thus, we observe very small values of the relative bias b for the
variational technique. It also imposes a smoothness constraint on the fused image,
which however leads to a reduction in the sharpness of the fused image, and thus,
the values of the variance and average gradient are not very high in this case.