The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
calibrated, the platform pose was corrected only. More details
are not discussed in this article.
Figure 2 Data preprocessing flow
2.2 Preprocessing For fusion
Special filter are designed for the fusion .It is a filter called
neighborhood- cluster vector filter-NCVF (Ma Yanhua, etc,
2006) which base the clustering of a pixel’s neighborhood area
to reduce the scattered minor spectrum. It can keep the spectral
information unchanged when erasing noise and small areas of
odd spectral, at the same time it can sharpen the edge of the
images in some extension. A practicable algorithm that
similar with the vector median filter (VNF) was designed to
realize the filter.
Edge detection is used to segment the High special image for
pure pixel spectral diffusing, and used to detect pure pixel for
hyper-spectral, to be introduced more in this article .The hyper
spectral edge detecting usually adopt a simple method:
principal component transformation and then use ordinary Edge
detection algorithm, such as sobel operator,etc.
2.3 register strategy
To meet different applications, all kinds of image fusion
algorithms are realized ,include fusion algorithms that are now
generally used, such as the PCA, high pass filter , wavelet
transformation algorithm, etc, but most of them are depended
on register precision(figure 3). The difference of GPS of the
images that used in this article is very large, the low special
resolution hyperspectral image has to be blurred, the spectral
are blurred too, so the fusion result must be appended the effect
of image re-sampling, just as showed in figure 4.
In figure 4,(a)shows a hyperspectral image of AMIRS after
register; (b)is the high special image; (c)and(d) is the results of
high pass filter fusion and PCA algorithm. The result shows
that with such a resolution difference, the spectral after register
has been blurred too bad to cite.
Figure 3 general image fusion flows
Figure 4 the low resolution image is blurred after register
At the same time, big resolution difference and short of precise
geometry correction of the experiment images
Depresses the register precision, so a register strategy is put
forward.
Rough registration means getting the spatial relationship of the
pixels in two images, but the hyperspectral image are not
re-sampled according to high special resolution image. General
registration algorithms can be used.
Because the two kind images are captured at the same time and
the two images are calibrated ,and the spectralwave are known
and overlapped ,so the luminance in corresponding bands(the
sum of several bands of hyperspectral image with one band of
RGB image) are high correlation, the registration between the
super pixel and the high special then with the brightness
correlation, you can judge which block(in spatial segment
image ) a spectral belong to by moving the spectral image pixel
in a initialized range around the rough registered position the
RGB image .
By the strategy, the blur of re-sample ,the problem of register
precision are resolved.
3. FUSION
In this article the target of fusion is to get classify information
on spectral and spatial distributing information on
corresponding high spatial images.
3.1 workflow
A workflow to carry out the target is sesigned as below. Showed
in figure 5.
1176