Cruz, Santiago 
  
Once the areas have been extracted, by means of two rotations (left and right edges) and a vertical flip (lower edge), the 
left, right and lower edges can be converted into an upper horizontal edge, so only a template (upper horizontal line, in 
figure 2) will be used (figure 4). Once the edges have been extracted and measured they will be returned back to the 
original positions. 
  
Extraction of edge areas Conversion to upper edge 
(upper horizontal line template) 
Flip vertical 
  
(2) Rotation 90? CCW 
  
(3) m——————X | emains unaltered 
  
(4) Rotation 90° CW 
   
  
  
(1) 
  
  
Figure 4. Conversion of lower, right and left edges into upper ones in order to use only one filter (upper horizontal 
template). After detection and measurement of edges, windows are returned back to original positions. 
The final process to extract the edges is summarized in figure 5. It is usual that in 35 and 70 mm films appear a lot of 
marginal data (photo and ISO marks, film type and trademark), holes for film advancement, etc. All these elements can 
be confused with linear elements (in fact, they do). This problem has been solved by an image binarization on the 
extracted and filtered areas. By default a threshold of 80 digital number (DN) has been selected. This threshold has 
shown to be efficient for properly contrasted edges. Once the image is binarized, the line with more white pixels (255 
DN) is searched. Because the edges can be slightly tilted, this line is looked for with a tolerance between 0.5-1°. Then a 
narrower window (30 pixel width) is extracted again around this line, window (d) in figure 5. This way allows working 
with a narrow band around the edge in the filtered image, but the most part of details that can produce anomalous data 
have been suppressed. 
Once edge has been detected, transversal profiles regularly spaced are made and the gradient through those profiles is 
measured. Where gradient shows higher values the edge is expected to be there. But the edges can be a few pixels width 
because of poor image definition and the scanning resolution. So, a routine will measure the three higher gradient 
values in each profile. This implies that for a horizontal line three row values, y-coordinates, correspond to one column 
value, x-coordinate. With all these data (one x-coordinate and three y-coordinates per profile) a fit line is computed. 
User can select the number of profiles. The resolution and the image size limit the number of profiles, but fine results 
can be achieved with 50-200 profiles each edge (for 600-1200 ppi image resolutions). As mentioned, the measurements 
involve the three higher gradient values each profile. That means between 150-600 measurements per profile made 
automatically in a few seconds (with a conventional PC Pentium™ II at 350 MHz and 64 Mb RAM). 
The calculation of the fit regression line is made with high redundant data. Because noise can produce high gradient 
values, there will be a need for blunder detection. Although rigorous treatments, such as robust statistics (Huber, 1981), 
can make efficient blunder detection, a simpler approach has shown to be efficient enough. User can state a maximum 
limit for rejection of observations. For the most cases, a limit of residuals based on 3 times the standard deviation can 
filter the noise. Once blunders have been rejected, the lines are fitted again. For color photographs, the three RGB 
channels are separated and an adjustment is made for each channel, getting the values for the best adjustment. 
  
152 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000. 
  
  
  
2.3 
Once c 
system 
lower « 
accord 
becaus 
digital 
The m 
becaus 
this w: 
severa 
coordi 
access 
(as ma 
is not 
compa 
errors 
values 
Coord 
orient: 
instea 
becau: 
may L 
metho 
frame 
larges 
taken 
found 
differ 
coord