ackground 
to the back- 
(table 1). 
hing 
m] 
79 
.85 
.90 
of the block 
ze which is 
f 5 times 5 
1t is carried 
of 3.17 um 
' sharp and 
is the 50 96 
the match- 
schnically a 
atching and 
tion outside 
ze 
  
  
  
  
  
  
  
  
rate of suc- 
r of failures 
(but jumps 
signals the 
nd a target 
t of the sig- 
ccuracy ob- 
e observed 
um is 15 9^ 
the manual 
e proposed 
ints at least 
oy a human 
ns. 
3.4 Taking Compression into Account 
The big storage capacity needed to hold the digital images 
of a block on a disc is still a weak point of Digital Pho- 
togrammetric Systems. Therefore, it is quite convenient to 
use compression algorithms such as JPEG. A measure for 
choosing the amount of compression in JPEG is the so- 
called Q-factor. The relation between Q-factor and com- 
pression rate depends on the texture in the images. This 
dependency is shown for this image block in figure 6. 
  
25p—— T | zm 
| i i | 
— N 
Cc © 
T T 
= 
© 
compression rate 1:X 
  
  
  
  
= | p | 
0 100 200 300 400 
q-factor 
Figure 6: Compression rate as a function of the Q-factor 
Of primary interest is the dependency of image compres- 
sion on the measurement precision and on the success 
rate. This is plotted in figures 7 and 8. In this case tem- 
plates t1/t2 are used with a window size of 9 pixels radius. 
  
  
  
0 100 — 200 la "ux —— 7 7400 
q-factor 
Figure 7: Measurement precision vs. Q-factor 
  
      
| 
| 
e0] ni : 
successful matches [%] 
20|- 
  
  
| | 
  
0 1 i i | i 
0 25 50 75 100 125 150 175 200 225 250 275 300 325 350 375 400 
q-factor 
Figure 8: Success rate vs. Q-factor 
A deterioration of the measurement precision can be ob- 
served with values from 2.8 um to 3.7 pm if the Q-factor 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
increases from O to 400. On the first view a loss of accu- 
racy of around 35 % seems to be a relatively small number, 
having in mind that the compressed block reduces the vol- 
ume for storage to 330 Mbyte. But if we look at the number 
of successfully matched signals this rate is below 50 % at 
the high compression factor. Thus it is mainly due to the 
matching procedure that a relatively good accuracy level is 
obtained by eliminating those points which do not fit any 
more under compression. The figures show also that, for 
example with a compression rate of 10 (Q around 100) still 
a reasonable result is obtained with 98 % successfully mea- 
sured signals and without significant loss of accuracy. 
3.5 Comparison with Different Types of Templates 
With the last experiment we want to explore the influence of 
the different types of templates (figure 5) on the measure- 
ment process. For this comparison only three very differ- 
ent compression rates are considered with the Q-factors 0, 
200, 400. A priori it is to be expected that because of the 
usually small resolution of signalized points in images the 
importance of different templates is not as big as it would 
be if a signal is imaged over a image area of e.g. 50? or 
500? pixels. The result of the template combination t1/t2 
was discussed in the section before and is listed here for 
comparison with the results of other template combinations 
(tables 3 and 4). 
Table 3: Matching with other template combinations 
  
  
  
Templ. Q- Matches 
type factor successful failed co 
rel. [%] | abs. | abs. | [pm] 
t1/t2 0 98.7 1692 22 2.78 
  
200 87.9 1506 208 3.18 
400 47.8 819 895 3.74 
t3/t4 0 96.5 1654 60 2.75 
200 79.4 1361 353 3.11 
400 31.6 542 1172 | 4.01 
t5/t6 0 93.4 1600 114 2.78 
200 84.0 1439 275 3.14 
400 27.3 467 1247 | 4.14 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
The number of successfully matched points found with tem- 
plate combination t3/t4 is significantly smaller than the num- 
ber obtained with the t1/t2 combination. And the differ- 
ence becomes bigger with increasing compression rate. 
This means, that unsmoothed structures are more advanta- 
geous in the case of this small signals. The obtained quota 
in matching with the templates of circular structure t5/t6 is 
still some percentage points smaller than that of template 
combination t3/t4. A comparison of the obtained precision 
values between the three template combinations shows the 
familiar result of nearly equal precision at equal Q-levels 
which already was observed from table 1. 
A further concentration on the contours of the signals is the 
reason for selecting the templates t1c to t6c (figure 5). They 
are generated calculating templates of gradient strength of 
t1 to t6. The images in this case are preprocessed in the 
295