ı NON-
0 measure
omic force
fa tip with
hape, it is
calibration
ess all the
ige of the
nid
width and
ut the real
0 coin, this
id and be 9
difficulties
axis.
3. PHOTOGRAMMETRIC RESTITUTION
3.1 Parallel Projection
As we have mentioned, we will now consider parallel projection
to be the normal case in microscopic imagery. Several works
have shown, that errors due to the projection model are smaller
than errors in image point measurement at magnifications of
approximately 1000x and more.
The mathematics have to be adapted to the different model. This
will include the bundle block adjustment and any subsequent
section 1n space.
First we will examine how the interior and exterior parameters
of the sensor and the images change.
Since the ray of sights are parallel they will not merge in one
central point, hence there is no more projection centre. The
missing projection centre has effects on both, the sensor and the
image. On the sensor we can no longer construct a
perpendicular on the image plane, known as the principal point.
Also there cannot be a calibrated focal length anymore, i.e. the
distance from the principal point to the projection centre.
Instead the image is described by a global scaling factor. So the
number of unknowns per sensor has dropped from three to one.
However, additional parameters like radial and tangential lens
distortion, or affine transformation parameters may still be
applied (Elghazali, 1984). The global scaling factor, which
substitutes the calibrated focal length is also the mathematical
explanation for the ambiguities described in figure 3, in contrast
to central projection, where each point has an individual scaling
factor, depending on its spatial distance from the projection
centre. Since in case of parallel projection each image point has
the same global scaling factor, the object points distance has no
influence on the images geometry anymore. Hence, the tilting
direction of planar shapes is not distinguishable.
Due to the missing of a projection centre, we cannot clearly
define the position of the image plane in object space. While the
parameters of the rotation remain unchanged, it is sufficient to
give a two dimensional offset along the x-y-plane to represent
the imaging situation exactly. The z-component may be dropped
since any image along the line of sight will give the same result.
So the number of unknowns per image is now five instead of
six. In figure 6 the geometry of parallel projection with its
consequences is illustrated.
A.
»
os.
Figure 6. Parallelprojective geometry
3.2 Parallel Block Adjustment
After making clear why and how the geometry changes, we
would like to describe the differences mathematically.
Previous works often used the samples tilting angle as the only
parameter for the three dimensional object reconstruction. The
forward section was simplified to this assumption and the use of
two images was enough (Hemmleb, 1996; Burkhardt 1981). In
this work we want to extend the mathematical model to full
flexibility for several reasons. First, the parallel block
adjustment provides us with highly accurate results in image
orientation and sensor properties. This way small errors in the
sample tilting process are modelled by the rotation angles as
well. These results will be applicable to any arbitrary image
setup. So this flexibility takes into consideration any future
hardware development, which may allow specimen movement
along more than one axis.
Additionally, any accurate subsequent section in space requires
full knowledge of the image orientations. The formulas which
are introduced later, will show an approach to derive accurate
object points from an arbitrary number of images. So not only
the accuracy is enhanced by taking advantage of observations in
several images, but we are also able to check for blunders, as
will be explained later in this chapter.
As mentioned above we now have to consider 5 unknowns for
each image and only one unknown for each sensor.
In the following, the image coordinates will be given in pixels,
related to the digital images centre. Hence the scale factor has
the unit pixel/um, or pixel/nm respectively.
The collinearity equation for parallel perspective will be as
follows:
x X- X.
0 Z
x = mR; (X = Xo) + Ry, (Y — Y ) + Ry Z]
y 2m|R; (x — Xo)+ Ry (Y — Yp) + Rs, Z]
where | m scaling factor
X, y = image coordinates
X, Y, Z = object coordinates
X0 Yo = projection centre (see Fig. 6)
R = Rotation matrix
The following model will be used as a Rotation matrix:
cop 0 sing) (1 0 0 cosx —sinx 0 (2)
Box 9 1 0 [0 coso -sino||sinx cosx 0
-sing 0 cose) (0 sino coso 0 0 1
When entering the parallel block adjustment, with observed
image points x, y, given control point coordinates X, Y and the
unknowns ©, c, K, X», Yo and m, we have to apply the following
observation equations:
—213—
SIRES.
2