n from the public domain
nd integrated into an in-
measuring system. The
e compression schemes
ftware library (FTP site:
ompression are in image
ntrifuge application the
>s is currently the major
1ave a high information
compression has a very
he LZW lossless method
; times. The influence of
as been tested in a series
oth retro-reflective and
conditions. Experimental
o target location quality,
hich JPEG is optimised.
[S image | Max. image
crepancy | discrepancy
0.083 0.952
0.069 0.643
0.063 0.247
0.056 0.165
0.048 0.133
0.041 0.125
0.032 0.102
0.024 0.096
0.002 0.020
PEG with different Q
was used to provide a
sion analysis. The image
anging from 20 (high
ression). ‘Target image
computed and compared
original image. Table 1
performance is closely
pancy. Even given image
MS image discrepancy is
> 8 illustrates discrepancy
measurements and those
or of 60. When compared
d by compressing the
nna 1996
43 45 45 44 4| 41 42 42 42 48 4B 4! 41 45 44 42 42 43 33 46 4H AN 40 34 di 4&6 47 47 4&6 45 44 44
42 45 49 49 44 42 41 41 44 46 45 4£ 43 41 43 43 44 GE 4& 4& 47 4? Al Ab 43 44 44 43 41 40 A6 46
42 34 41 41 43 SE G6 46 45 45 43 40 42 48 43 43 44 45 45 48 47 47 43 44 45 45 46 43 42 AU 42 4a
43 43 44 AG 45 44 45 A5 45 Ab 45 44 44 38 34 47 — 4? 42 43 4A 4A 45 45 48 47 49 49 48 47 47 44 44
41 42 41 40 4U 44 44 40 42 45 A) 4) AS AR 45 44 19 38 40 41 4i 42 35 38 40 43 45 45 45 45 48 48
4i 4b Ab Ab db 40 42 43 44 44 45 46 A6 45 46 48 42 42 42 AZ 42 42 42 $1 45 46 4& 45 43 42 45 46
48 43 48 44 44 43 39 38 43 46 45 42 43 45 45 43 45 46 46 46 4& AS 42 43 45 AG 4b A5 43 42 AB AB
b) Intensity values of the target after
target before JPEG compression JPEG compression at a Q=60
with figure 2, it can be seen that the discrepancies are even
smaller than the random target location noise between any two
successive directly grabbed images.
The main cause of the degradation in target location precision is
in the JPEG compression quantization procedure, where pixel
intensity values can be changed. This can result in a shift in the
computed target location co-ordinates. Fortunately the effect is
not as great as might be expected since the JPEG algorithm
achieves most of its compression by reducing information in the
low frequency portion of the image. This results in a merging of
the background pixel levels and lower target intensity levels but
has little influence on the target centroid properties. Figures 9a
and 9b show the pixel intensity distribution for a target before
and after compression at a Q factor of 60. It can be seen that the
target images are not smeared and that in fact an even higher
contrast has resulted.
3.3 JPEG within an analogue CCD camera network
For a photogrammetric evaluation of the JPEG procedure, two
testfields were built (Figure 10a and 10b). The first, a black
retro-reflective targeted testfield, consisting of a 250mm x
230mm aluminium base plane with 28 inserted rods of differing
lengths. About 50 round retro-reflective targets, 2mm in
diameter, were placed on top of the rods and to the base of
testfield. The second testfield, representative of the centrifuge
Figure 10 a) Retro target test field b) Conventional target test field
case, consisted of 80 black targets on a simple white board. A
Pulnix TM-6CN camera, with a 16mm Fujinon ‘C’ mount lens,
was used to grab images at each of the four corners of an
imaginary square based pyramid network. The target image size
in each exposure varied from 3 to 5 pixels in diameter. JPEG
compression was carried out on each set of images using Q
factors ranging from 20 to 100.
A free net bundle adjustment was computed for each image
compression set. Camera calibration parameters were computed
for the lossless case, then held fixed for each different image
compression set. In this way results between adjustments could
be directly compared. Table 2 shows the compression ratio over
the same Q factor range, 2D image measurement discrepancies,
and RMS image co-ordinate residual after each
photogrammetric adjustment.
Q- Compression RMS 2D image Adjustment
Factors ratio discrepancy RMS image
(pixel) residual (pixel)
retro | conv. retro conv. retro conv.
20 32.0 37.0 0.054 | 0.101 | 0.098 | 0.124
30 28.1 31.8 0.038 | 0.081 | 0.098 | 0.117
40 25.2 274 0.032 | 0.069 | 0.097 | 0.060
50 22.7 23.1 0.026 | 0.068 | 0.103 | 0.060
60 20.4 19.1 0.020 | 0.060 | 0.096 | 0.053
70 17.2 14.8 0.016 | 0.055 | 0.098 | 0.056
80 13:5 11.0 0.014 | 0.042 | 0.096 | 0.055
90 8.5 6.9 0.009 | 0.027 | 0.099 | 0.053
100 25 23 0.002 | 0.003 | 0.099 | 0.053
LZW 2.5 2:3 - - 0.098 | 0.053
Table 2 Performance of JPEG with different Q factors for two
test images
Despite differences in image content, both retro and
conventional target cases have a very similar compression ratio
at a given Q factor. However, the retro-reflective targets have
provided 2D RMS image residuals which are about two times
better than those attained with conventional targets. This is
because the retro targets can provide a very high contrast target
image, about 220 intensity levels out of the available 256
intensity levels in the 8 bit image. The conventional targets
provide a signal of the order of 150 intensity levels.
With the exception of conventional targets at a Q factor of 30
and less, change in Q factor does not significantly affect the
photogrammetric precision achieved. This is shown clearly in
figure 11, where it can be seen that at Q factors of 40 and over
about 1 part of 11,000 is achieved for all photogrammetric
networks. The slightly better result in the conventional target
case is due to the planar nature of the test field used in this case.
It should be stressed that the four image network combined with
the limited optics and electronics inherent in the analogue CCD
camera used to record the images has only allowed a limited
evaluation of JPEG compression.
® Conventional targets
1:4000 + ® Retro targets
NNetwork Precision
8
8
L
EN
Q factor
Figure 11 Co-ordinate precision for the two
testfields and different JPEG Q factors
3.4 JPEG within a strong digital CCD camera network.
To evaluate JPEG further it was decided to repeat the series of
tests with a strong network of images captured with a state of
the art digital camera. A suitable data set was kindly offered by
Professor Mark Shortis of the University of Melbourne (Shortis
et al. 1996). The data consisted of a convergent image set of a
targeted wall taken from 6 camera stations. At each station a
single DCS 420 camera fitted with a 20mm lens was rotated 4
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B5. Vienna 1996
