times by 90 degrees to give 24 images. This resulted in an
imaged target size of the order of 10x10 pixels, target image
intensities being about 170 grey levels. With no compression
the original network precision was of the order of 1:116,000 in
3-D space and about 1/20th of a pixel in image space.
Q factor | Compression 2D RMS Adjustment | Network
ratio image RMS precision
discrepancy image
(pixel) residual
(pixel)
30 532 0.0737 0.082 1:74,820
40 50.0 0.0620 0.075 1:96,100
50 47.9 0.0478 0.071 1:103,360
60 45.6 0.0390 0.069 1:108,170
70 42.4 0.0314 0.066 1:112,110
80 35.3 0.0235 0.065 1:115,130
90 16.4 0.0174 0.063 1:115,810
100 2.6 0.0053 0.063 1:116,000
LZW 4.4 - 0.063 1:116,000
lossles
Table 3 Performance of different Q factors for the DCS420 network
1:120,000 7
60 4 1:110,000 |
1:100,000]
1:90,000 |
Compression ratio
>
o
i L 1
Network precision
1:80,000
0 yyy 1:70,000
20 30 40 50 60 70 80 90 100 LZW
20 30 40 50 60 70 80 90 100 Lzw
Q factors
Q factors
Figure 12. Image compression Figure 13. Network precision with
ratio with different Q factors. different Q factors.
All images were compressed using Q factors ranging from 30 to
100. Target image co-ordinates were then measured as before
using the centre weighted algorithm and downloaded into a
series of identical bundle adjustments. Table 3 and figures 12
and 13 illustrate: the image compression ratio; RMS image
discrepancy by simple 2D comparison with the uncompressed
image and; the RMS image residual computed within the bundle
adjustment.
From these results, it can be seen that the influence of JPEG
compression has to rise above about 1/30th pixel before any
significant influence on photogrammetric adjustment precision
is seen. The strength of this well designed network has allowed
much less redistribution of error into the estimated target co-
ordinates and camera orientation parameters, consequently a
clear trend of Q factor against object space precision can be
seen.
In summary, careful use of JPEG image compression can be
recommended for retro-targeted close range photogrammetry.
The compression ratio can be arranged from 10 to 50 times
according to the qualities of the imaging system,
photogrammetric network and required co-ordinate data
specification. For example, if the system is capable of target
location precision better than 1/20th pixel, the Q factor should
be set between 90 and 100. If the target location precision is
less than 1/10th a pixel, a Q factor can be of 80 or lower can be
used. A particularly useful indicator for an appropriate Q factor
is the discrepancy between target locations measured on the
compressed and uncompressed image.
74
4. ON-LINE DYNAMIC TARGET LOCATION
Both S-VHS and JPEG can store long sequential images in real-
time or near real-time for subsequent processing. However, in
some experimental cases a rapid display of dynamic target
deformation information can be required. A suitable on-line
algorithm has been written to satisfy this requirement.
Practical general purpose algorithms for automatic target image
measurement consist of target search, target recognition and
target location processes. Most of the computational time in this
process is spent on the target search and target recognition
components (Chen, 1995). A new algorithm based on a prior
knowledge of target locations from subsequent images in the
sequence has been written. In this way the time necessary to
search the whole image, recognise any targets and to compute
target matches between any two successive images can be
avoided. A comparison of the computational cost of the target
location algorithm elements, based on a Pentium-90 PC running
Windows 3.1 is shown in table 4. It can be seen that a lot of
time can be spent on unnecessary operations. This is because in
a targeted image the number of useful target image pixels is
very small, typically between 1 and 5%. For example, in a
typical centrifuge experiment image, only 5603 pixels represent
the targets compared with 442368 pixels in total for the 768 x
576 image. In Table 4, it can be seen that 600ms are required to
complete the target measurement process for a 400 target image,
but of that, 556ms is spent on image background scanning in
the target search and target recognition procedures. Only 44ms
is required for the actual computation of all 400 target co-
ordinates. These times do not include the matching and
checking of target numbering between successive images.
Number of targets in 100 200 300 400 500
image
Complete general 330 | 440 | 500 | 600 | 660
algorithm (ms)
Target recognition 119 | 217 | 267 | 356. | 404
section (ms)
Complete prior-
knowledge based 11 23 33 44 56
target location (ms)
Table 4 Some timing performances for target location
calculations on a Pentium-90 PC
Figure 14 On-line soil model analysis computed from centrifuge
image measurement data
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B5. Vienna 1996
In a centi
any two
ordinate
ordinate i
only nece
area surro
location :
image usii
centres fc
algorithm
co-ordinat
onto a sta
ordinates
to provid
successive
operations
Target m:
necessary
point fron
any pair o
as require:
be fitted |
obtain gec
result fron
time of w
one secon
potential t
the progre
The S-VH
6 lead to t
much bett:
secondly
unsuitable
contrast.
The JPEG
use with «
capable o
optimised
However,
retro-targe
compressit
photograrr
the desire
imaged ta
photograrm
experimen
; compresse
location [
precision ¢
JPEG soft:
into a gene
The prior |
suitable fo
where targ
small. T
measurermx
algorithm,
number ar
image me:
assistance
