Re: [Gimp-print-devel] Latest results

[Thomas T. <tt...@bi...>]

Robert L Krawitz wrote:

> Thomas, could you play around with your new matrix some more?  It's
> not in, because having only one matrix caused problems, but it does
> look promising.  It really bloats the executable, though.

I have a number of things I'm trying.

The simplest is to change the ditherpoint call that does the black
decision to use not x and row but x+50 and/or row+50. Given that the
matrix is not very correlated with itself, this should give okay
results.

The second is that I'm generting a new matrix now. It has somewhat
reduced noise, but more interesting may be the change in selected
initial points. The current matrix has starting points selected for
minimum autocorrelation when shifted 1/2 H or V, the new matrix will
shoot for 1/3 or 2/3 H and V. This gives 9 uses for a single matrix:
enough for cCmMyYkK and yet another independent use.

I could also try making another, smaller matrix, for the decision to
'print black': maybe something as small as 31 on a side or as large as
151.

The last option would be to generate a number of matrices to be used for
a single purpose each. I'm afraid this will give huge bloat.

I think I'll try the options in the above order. How do people feel
about these large matrices appearing in CVS? If no one objects, I'll put
the new matrix in when it is ready and produces useful results.

Robert, do you think 199 x 199 is big enough for a release version or
would you prefer 251 x 251 or 257 x 257 ? That would make full of the 16
bits internal resolution at least.

As for the sharpening that is inherent in Floyd-Steinberg: I didn't
think about that myself, but read about it in one of Ulichney's papers.
But I think I understand, and maybe I can even explain:

Error diffusion does the following:

Input -> Approximation -> Simulation -> Delay -> Subtract from next
input values.

Essentially, this is what error diffusion does: except FS does things a
little more abstract and talks about 'error': instead of simulating and
then subtracting from other input values, FS takes a shortcut and
subtracts a value immediately from _coming_ input values if a dot is
printed.

In any case, the approximated version (which is much like the original)
is delayed and blurred (by diffusing into other cells) as it is
subtracted from the input. And what happens if a blurred version of X is
subtracted from the original X: X is sharpened by making the low
frequency components weaker. The process is also called 'unsharp mask'.


Thomas