The program works by generating a low resolution fractal (400 points) for a
particular range. Then the fractal is divided into quarters. The each quarter
is given a score based on the number of colours active in it. This method is
fairly effective at keeping the program focused on an active area of the
fractal, although there is much improvement to be done. There is also an
option (on by default) to select a randomly choosen quadrant than the one with
the highest score if the randomly choosen quadrant has a score that is at least
80% of the highest score. This process of selecting a quadrant is repeated a
fixed number of times (by default, 10 times). At this point, a high resolution
(in terms of floating point precision), image is created and saved to disk.
The cutoff value is very high, which is required for deep_zooms. To help avoid
endless cylces a binary tree of the hashes of all previous iterations is stored
and checked against the current iteration.

WWW: http://members.rogers.com/rowley-c/deep_zoom/