Tue, 15 Jun 1999 20:49:22 -0700
> I think no signature function such as you describe in (1) exists,
> except for the domain of one-pixel images.
No need to preserve spatial information. Probably better to ditch it if you
want to catch flips and rotations.
Assuming your 16x16 reduction, here are some 1-d signature functions: (which
may be useless except for argument. I'd need to know more about your
application to be able to say anything useful)
1) are there more 1 or 0 bits?
2) how many more are there?
3) how many bit transitions are there along the principal diagonals/the
principal orthogonals/around circular paths?
One would need a very wide fudge window with all of these, but they should map
similar images into similar values.
You should check the literature. I recall people at a Seybold conference in
94 or 95 with a "find more like this" database product for images; no doubt
the state of the art has advanced somewhat since then.