Image Processing Reference
a high resolution imagette extracted from the original image at 0.7 m into the 5.6 m
smoothed image. Thus apart the encrusted part, this image is similar to the
The examination of the high and low resolution portions in Fig. 11.2 and of their
differences clearly demonstrates the large amount of information which is gained
when the spatial resolution increases from 5.6 to 0.7 m. Contours are blurred in the
low resolution part. Buildings and streets are clearly visible in the original image
while even large streets are hardly visible in the 5.6 m resolution image. The loss
of information between the reference image and the 5.6 m smoothed image is quan-
tified by computing the difference between both images for each pixel of 0.7 m.
The differences are synthesized by a few parameters: bias, difference of variances,
and a correlation coefficient. There is no difference in average values; that is, the
bias is equal to zero. The difference in variances expresses the difference in infor-
mation and is equal to 47% of the mean value. It means that 47% of the information
content is lost when decreasing the resolution from 0.7 to 5.6 m. In other words, the
small size structures lost in the range between 0.7 and 5.6 m account for approxi-
mately half of the information contained in the original image. The correlation
coefficient between both images is only 0.85. This example clearly demonstrates
the interest to have images with the best available spatial resolution and therefore
the need for efficient encrustation procedures.
Encrustation is accomplished using the following approach. First, a high resolution
imagette and a low resolution image of a larger geographical coverage are re-mapped
onto each other or onto a common geographical reference. The resulting re-mapped
image and imagette now have the same pixel size, but different effective spatial resolu-
tions. The encrusted image is then computed from the values of the re-mapped imagette,
when available. Otherwise it can be computed from the values of the re-mapped image.
As Fig. 11.2 shows above, the encrustation creates a clearly defined border which
is created at the periphery of the imagette due to the dramatic change of resolution.
This border is often disturbing in photo-interpretation and image analysis. Also the
discontinuities at the periphery (abrupt changes in derivatives) can prevent digital
processing methods such as edge detection and pattern recognition from performing
well. There is, therefore, a need for improved methods for encrustation.
One common solution is to apply a filter to smoothen the inner periphery (or even
the inner and outer peripheries) of the imagette as shown in Fig. 11.3 below. This
filtering may efficiently attenuate the border but has a drawback since the high reso-
lution content of the periphery of the imagette is replaced by information of lower
accuracy. Likewise, if the smoothing is applied to the outer periphery, it would
result in a degradation of the accuracy for this area.
To solve this problem, we propose a method for encrustation which attenuates
the border but still preserves the information content of the whole imagette.