No subject

a bunch of objects (e.g. tracks & pads) and infer what they are.
However, it is easier to have a bunch of target patterns (e.g. known
footprints), and then do a 2D autocorrelation of the target pattern on
the unknown image (i.e. the Gerber image).  The places where you get a
peak in the autocorrelation correspond to places where the target
exactly lined up with the desired feature on the board.

Therefore, in this scheme you would have a bunch of known, common
footprints stored in some bitmap format, and then you'd convert the
Gerber layer to a bitmap format.  You'd loop over the footprints, and
would then do the 2D autocorrelation of the board bitmap with each
footprint bitmap.  The places where you get a autocorrelation peak
correspond to a pattern match, i.e. the presence of the target
footprint.

By the way, each target footprint would have a known, calibrated
centroid, so the position of the autocorrelation peak could give
exactly the centroid position.

In this scheme you'd have to do some further analysis to
e.g. distinguish the signature of autocorrelating a DIP-14 target
against a DIP-16 on the board.  That's where things would get
interesting, and might require human supervision.

Just a thought.  I'm not promising to do any of it.....

Stuart