The color image difference introduced here can be used for fast image
query as well. We propose to make a low resolution (e.g. 
) version of each data-base image. Furthermore, the method will be
not view distance dependent any more. The maximum frequency will be
set so that the most important frequency corresponds to the
rectangles of approximately 1/2 to 1/3 size of the reduced
images. Then, the first 200 rectangles are found, and average L',
u' and v' values are stored in an array. This whole process is
done once, and the L', u', and v' values (200 values for each
image) are stored. Now, the target image is drawn by the user, or
submitted somehow else, and the query begins. The target image is
reduced to the low resolution, 200 rectangles are found (note that
they will correspond to the pretabulated data-base rectangles as the
whole method is deterministic), and the image differences are
computed. The whole data-base does not have to be sorted, we only
want to find for example the top 10 images. When the highest
difference limit value of the top ten club is known, the
current difference evaluation can stop as soon as the sum of
rectangle differences exceeds this top limit. In this way, the
difference computation time will decrease as the top 10 club
will have better and better limits.