Page 261 - Digital Analysis of Remotely Sensed Imagery

Image Enhancement 223

The first filter is characterized by the same value for all elements
in the same row, but different values in different rows. A common
application of this filter is to remove from Landsat images the drop-
out lines caused by the malfunctioning of one of the detectors. This
removal is accomplished by applying the filter along the dropout
lines. A missing line is replaced by the average of the two scan lines
immediately above and below it. The second kernel places less impor-
tance on the four corner pixels because they are farther away from the
pixel under consideration than the four horizontal and vertical neigh-
boring ones.
No matter how widely the weight varies from one element to
another in the kernel, all three smoothing operants presented above
have one characteristic in common: the sum of all nine elements
divided by the scalar in front of the matrix equals 1. In this way the
image pixel value is not artificially scaled up or down after being
smoothed. The last two kernels are the same as the first one in that
they maintain the input image pixel values unchanged after the con-
volution. This is achieved by having a scalar equivalent of the inverted
sum of all elements.
Image filtering using an operand with differential weights is
called high-pass filtering, during which the difference between adja-
cent pixels is artificially enlarged. Contrary to low-pass filtering,
high-pass filtering attenuates low-frequency features (Gonzalez and
Woods, 1992). As a result, high-frequency features, such as edges
between homogeneous groups of pixels and other sharp details,
stand out. High-frequency filtering produces an effect just opposite
to low-frequency filtering. In high-pass filtered images, large pixel
values become larger and spatial frequency is increased. A high-fre-
quency kernel or high-pass kernel has the effect of enhancing features
of a high spatial frequency. High spatial frequencies are those that
represent frequent grayscale changes in a short distance. For instance,
features that are separated at a large distance are made more visible
on the output image. The net effect of high-pass filtering is the reduc-
tion of slowly varying features and a correspondingly apparent
enhancement of edges and other sharp details (Fig. 6.13c). Unlike
image smoothing, the operant has a strong component of orientation.
Only those edges oriented along a certain direction can be sharpened
during high-pass filtering. It enables edges to be highlighted, but
does not necessarily eliminate other features (Leica, 2006).

− ⎛ 1 −1 − ⎞1
− ⎜ 1 9 − ⎟ 1
⎜ ⎟
− ⎝ 1 −1 − ⎠ 1
Unlike low-pass filtering operants, high-pass operants do not
have a scalar since all the elements in it amount to a sum of zero, as
shown in the example above.