Matching is first performed on the top image 
pair in the pyramid. This is the image pair with 
greatest resolution reduction. Interest point 
extraction and point matching are applied at 
this resolution. The initial transformation 
defined between the two images may not be 
very precise, but this should not be a problem 
because interest points extracted at this 
resolution should correspond to major features. 
Moreover any error in the transformation at 
this scale should have little affect on the 
matching procedure 
The resulting transformation defined between 
the two images should be an improvement on 
the initial transformation. This new 
transformation can be passed onto the next 
level of the pyramid. In this manner, the 
pyramid system allows an initial loose 
transformation to be redefined and improved, 
as the procedure passes down the pyramid. 
In practice the image resampling is performed 
in synchronicity with the image smoothing, 
otherwise image aliasing would be a possibility. 
2.5 Region Of Interest Definition 
Within a SPOT image there may be areas which 
are particularly suitable for matching, or 
conversely areas, such as cloud, which will 
detract from the results of image matching. 
The operator can view the two images to be 
registered and highlight areas which will be 
used in the matching, or alternatively can mask 
out those areas that should not be considered in 
the image matching procedure. 
2.6 Feature Extraction 
Feature extraction is the process which 
determines in each image those features which 
have a particular characteristic that make them 
potentially good features to match. They are 
ideally features that exist in both images! The 
features matched in this phase of the PAIRS 
system are points (individual pixels). In the 
next phase of PAIRS matching will also be 
achieved by considering area (polygon) 
features. 
In this process the Forstner operator is used to 
determine ‘interest’ points in an image. The 
Forstner operator in practice requires the 
passing of several filtering kernels over an 
image, followed by several arithmetic 
operations to define two quantities for each 
pixel. These quantities are ‘weight’ and 
'roundness'. The image is thresholded on the 
basis of these quantities. Any remaining pixels 
are non maximally suppressed so that clusters 
of pixels are reduced to a single pixel (interest 
point). 
The ‘weight’ value for each pixel is retained. It 
is used as an initial weighting for a pixel in the 
feature matching process. 
2.7 Feature Matching 
2.7.1 Objective 
The feature matching process takes as its input 
a SPOT image pair, from which interest points 
have been determined. The objective is to 
match interest points in the first image of the 
pair to interest points in the second image of 
the pair. 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996