The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
112
the actual centre as far as possible; (3) the points detected are to calculate edge intensity and Canny operator to calculate edge
relative with the actual edge points one by one, it is also to say intensity, which can be described as follows 141 :
that the width of edge should be a single pixel.
Canny detector uses the first order gauss differential coefficient
to seek good balance between noise abatement and the edge
detection, and its expression is similar to first derivative of
gauss function. Suppose G(x,y) as the two-dimensional gauss
function, and f(x,y) as the image, the edge intensity and
direction of Canny detector can be obtained as follow
expressions:
Edge intensity :
:C(x,y)=\VG*f(x,y)\
(1)
Edge direction
. n VG*f(x,y)
(2)
\VG*f(x,y)\
However, there is no CFAR (constant false alarm rate) in SAR
image for Canny detector, and the CFAR in high-brightness
area is higher than low-brightness area. Therefore, Canny
detector can’t be used in SAR image edge detection directly 141 .
2.2 Ratio Detector
Touzi and his partners proposed that there is a CFAR
characteristic for Ratio operator in Gamma distribution coherent
speckle 51 ,and using this algorithm to generate SAR image edge
can get a higher precise. The main principle can be described as
follows: in the glide detection window, getting the central pixel
point as the detecting object, calculate the sample averages Ui
and u 2 in non-overlapping region of both sides along a certain
line direction which passes this point (Note: suppose that the
image data are 32 - bit intensity values; if the actual data are the
16 - bit range values, their mean square values should be used).
Calculate the ratio(R) of two averages, and choose a smaller
value larger than one from R and R" 1 as r. By the definition, we
can know that if r tends to one, the average values of two
regions are closer, and the objects of two areas are more likely
same; Otherwise, if r tends to zero, the objects of two areas are
more likely different, and the point to be detected maybe at the
edge of two regions.
Given the edge different orientations, along four directions
shown in Figure 1, each point should be detected, and reserve
the smallest r value as the result. The neighborhood of Ratio
operator shouldn’t too big, because the bigger neighborhood
can only detect the wide edge. As a rule, the neighborhood of
Ratio operator is 5x5. The edge intensity and direction of Ratio
detector can be obtained as follow expressions:
Edge intensity: R[~l /Min(r) (3)
Edge direction: normal direction of corresponding straight line
L
Edge intensity : R L =1 /Min(r)
(4)
n .. VG*f(x,y)
Edge direction : n =
(5)
\VG*f(x,y)\
(b) Vertical direction;
(c) Right-leaning 45 degrees; (d) Left-leaning 45 degrees
Figure 1 Four different detecting directions of Ratio operator
3. ESTABLISHMENT OF “VIRTUAL EPIPOLAR”
FUCTION
In computer vision field, epipolar is called “the polar line”, and
plays an important part in stereo process of remote sensing.
However, due to the difference of imaging mechanism between
SAR and optical image, it is difficult to understand when
epipolar surface and epipolar line are established in SAR
images. However, during making geometry rectification or
generating DEM by SAR images, oblique distance projection
can be transformed to central projection, and then “equivalent
collinear equations” of SAR images can be established. In this
condition, SAR images can build their own “equivalent epipolar
equations” as the optical images to be used in image match. To
distinguish the definition of “epipolar” in optical images,
“virtual epipolar” line is called in this paper.
Many methods can be used to confirm the “epipolar ”, and they
are mainly divided into two kinds: one is based on the
geometric correction of digital image; the other one on coplanar
conditions 161 . If there are some ground control points, the
equivalent elements of orientation can be got by collinear
equation, and then the equivalent “epipolar equations” can be
established. When it is difficult to get elements of interior
orientation or only a part of the image should be dealt with, the
RLT (relative orientation linear transformation) algorithm based
on coplanar condition can be used to establish corresponding
epipolar. Its expression can be described as follows 161 :
2.3 The Integrative Edge Detector q = I% + I^x + I^y + I? 4 x' + I? 5 xx' + I? 6 xy' + I^yx' + I^yy' (6)
In order to make the edge detector with CFAR and a good
direction, the edge direction can be calculated separately.
Whether the edge operator with CFAR or not is determined by
edge intensity. Therefore, the new detector using Ratio operator