to 
lon 
nel 
nts 
ent 
orf 
nts 
1er 
; 1S 
WS 
els 
nt 
nd 
all 
nts 
he 
Its 
on 
nit 
wi 
2335 
of the convergence radius the starting point is shifted in a spiral manner with increasing radius on the discrete 
lattice around the “true” point location. If the matching algorithm converges and the result is not more than 
0.5 pixels from the correct point location the matching is considered successful and the next starting point is 
taken. If matching is not successful for the first time the smaller radius of the previous match is considered to 
be the convergence radius. In table 3 the convergence radius for some selected sizes of the matching window is 
listed. Included are the multichannel solution and the solutions of the individually matched R,G,B components. 
  
  
window size R G B mc 
11 145 1.31 1419 1.94 
15 2.53 256 1.93 2.85 
23 260 330 314 3.72 
27 400 3.85 361 418 
31 456 4.31 3.93 4.66 
39 553 5.07 463 5.44 
  
  
  
Table 3: Convergence radius for the multichannel (mc) and the single channel solutions (R, G, B). 
The table shows that the convergence radius increases with increasing size of the matching window. Most 
interesting is the relation between the four solutions. For small window sizes the biggest differences are observed. 
The radius for window size 11? of the multichannel solution is about 34 % larger than the radius in red, 48 % 
larger than in green, and 63 % larger than in blue. For the large windows the convergence radius of mc, R and 
G are close together. Only the values of the blue component are always significantly behind the others. 
4. CONCLUSION 
The paper reports on our first investigation into matching colour images. Several possibilities for area based 
colour matching have been taken into account including matching the colour components of an image pair 
separately and matching the intensity components determined by colour transformations. As a new possibility 
a simultaneous multichannel solution is proposed. 
The investigation has shown that with respect to the theoretical precision the multichannel solution has the 
highest potential for precise point transfer. The empirically derived precision did not show so clearly the best 
qualified solution. Preferably matching with the red components of a colour image pair or with all channels 
simultaneously yields the best results. But matching in the green channel leads to a precision not too far from the 
red channel or the multichannel matching accuracy. Because in the RGB representation intensity information 
is included in all three components we suppose that a high correlation of the R, G, and B components causes 
this result. 
With respect to the convergence radius a superiority of the multichannel matching over matching the RGB 
components separately is observed for small window sizes. For large windows the convergence radius of mc, R 
and G are close together. 
Several conclusions may be drawn from the results. First from a matching point of view it is not necessary 
to make a colour transformation from RGB to another colour system. It takes additional time and space to 
transform and store the data without any additional profit for the matching. For the inverse situation this 
is also true. If colour images are represented in YIQ or HSI then the results of matching with the intensity 
components are good enough thus a transformation to RGB for matching purposes is not necessary. 
Colour matching will be useful if local image structure responds in only one channel. For a certain region in an 
aerial image this might be in the green component, for another region in the red component. If now a certain 
point must be transferred to a second image the simultaneous multichannel solution is most flexible because the 
matching exploits directly the information from all involved channels. On the contrary, for example, in DTM 
reconstruction colour will play a minor role. In this case thousands of points are matched and the individual 
point quality is not so important. If a certain point does not match often another point in the neighbourhood 
can be found for which matching is successful. 
In our experiments point selection was done by the operator. He selected (deliberately or not is not clear) points 
in which all colours are represented. It would now be very interesting to use a colour interest operator for point 
selection and check the quality theoretically and experimentally for those points again. We guess that this will 
change the assessment of the usefulness of colour matching. 
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop “From Pixels to Sequences”, Zurich, March 22-24 1995