OINTS
artography,
> restitution procedure
s the most boring and
utomation of the most
ally conjugate points,
> restitution procedure
etween the patches of
atching algorithm can
solution very close to
provide the sub-pixel
ion (i.e. less than 2-3
matching technique in
r. Two very important
nd least square image
ling technique on the
| be used in automatic
; also lead to errors in
rrelation process does
- MATCHING
d before from the
ssing of the interlaced
view of the imaged
e single image that is
res. This image can be
ers may be applied on
] to provide useful
mages in a composite
ft epipolar image and
ature is the extremely
ppear when the image
of the imaged object.
lar images brings into
images, providing a
he images even along
that the high textured
o viewing system are
correctly matched.
A new procedure can be introduced to automatically find the
correct stereo viewing in an interlaced image when this image
is getting from the highest to the lowest texture. This is
analogous to the algorithm that is used to automatically focus
an optical device such as a camera or a microscope. However,
in our case, the correct image (in terms of free of parallax
stereo viewing) is supplied when the image is the one with
lowest texture. Thus, defocusing is the correct way to bring into
coincidence conjugate points on the interlaced image.
e A blurred image (I) is generated from the original
composite image (I), through a 3x3 average filter. The
subtraction of the blurred image from the original one
gives a residual image q — I-P. A quality index Q
(summation of all q^ intensity differences) is used for
indication of the quality of the match. A minimum value
of Q indicates high match, whereas big Q values indicate
low match.
e The above step is repeated while the algorithm scrolls
(a) (b)
(c)
(d) (e)
Q-82270 Q-99533
Q=89020 Q=74828
Fig. 1: Auto-focus simulation procedure. Pictures from (a) to (e) have been taken with lens focus 0 to infinite. Every
image is blurred and subtracted from its neighbor. A focus factor Q is calculated for each pair by adding the difference
image values. Image (c) is the best focused since Q gets its maximum values in pairs (b-c) and (c-d).
Many researchers (Liu, 1998, Subbarao and Wei, 1992,
Subbarao et. al. 1993, Xiong and Shafer, 1993) have used this
procedure to determine the depth of view for several optical
devices. One way for successful defocusing is the minimization
of the difference between the original image and its blurred.
The procedure is trying in an iterative way to provide the
correct image and is very quick in its execution since the
creation and processing of the interlaced image is performed
from the graphics subsystem of the DPS. So, very low
processing cost is dedicated from the CPU of the computer for
the algorithm's implementation.
3. IMPLEMENTATION OF THE ALGORITHM
3.1 Procedure presentation
The algorithm has been implemented in an experimental
software application that can easily estimate conjugate points
on the stereo viewing window, using semi-automated
procedure.
In more detail the steps of the algorithm can be described as
follows:
e The approximate values of x-parallax on the epipolar
images are obtained through the existing relative
orientation.
e The user (through a preview window) positions the
cursor on the interlaced image point, and an image patch is
defined with this as a center. The selection of the patch
dimension is obviously related to the application just like
the respective window size used in correlation.
the even rows of the image patch relative to the odd ones
in different steps. The minimum value of Q factor gives
the best approximation for the locus of the matching
points.
e The next step is the refinement of the position to sub
pixel accuracy. For this, the images are zoomed by a user-
defined factor (e.g. 4x), depending on the accuracy
required. The above steps are repeated on those zoomed
images (now the applied steps are fractions of the original
pixel).
Instead, of using Q as the summation of the q’ values the
summation of the |q| may be used. The result would be the
same while the processing time is greatly reduced because the
calculation of a number's square value is more time consuming
than the calculation of its’ absolute value.
3.2 Complexity issues
The algorithm consists of a simple filtering technique and a
simple arithmetic operation between two image patches
(subtracting the blurred P image from the original I). The
algorithm uses less CPU process time than the typical
procedure of correlation and matching since it uses just nxm
multiplications for every possible position of the conjugate
points vs. the nxmx3 multiplications used in the correlation for
every possible position of the template window inside the
search region (where n is the dimension of the square template
window and m is the dimension of the square patch window). It
should also be noted that the arithmetic operations, which are
performed in our case, use integer operands. In the case of
correlation and matching procedure the operands are at least
single float type.
—75—