Re: [ft-devel] [Fontforge-devel] Peter Wiegel's auto spacing idea

[mskala]

On Thu, 14 Jun 2012, Vernon Adams wrote:
> This is very interesting, reminds me of David Kindersley's work on
> 'optical spacing'. I think that spacing for web and screen fonts could
> be dealt with like this fairly straightforward because of the role that
> the pixel grid plays in determining final rendered spacing. Would be
> interesting to test it. Also thinking this could be built into an app
> like robofont very easily - would just need to add a blur effect etc to
> the glyphs in spacing mode. could that be done in fontforge too? Using
> freetype or pango?

There's code in my Tsukurimashou package that implements an auto-kerning
algorithm serving a similar purpose to these ideas but with a few more
features.  The general idea is that I first calculate a left and right
contour for each glyph; the contours follow the actual shape of the glyph
except that there's a limit on how shallow the contour angles can be.
Then to kern a pair, I compute a distance measure between the contours
equal to the average inverse cube of the distance between the left and
right glyphs' contours.  The glyphs are kerned to make that measure equal
to a constant.

Using the actual shapes of the glyphs, even smoothed, is undesirable
because it can result in one glyph sliding too far under another, for
instance between apostrophe and period; those two can never be pushed
together far enough to make them touch because they don't share any
subset of the vertical extent.  My existing code still needs some tweaking
for this kind of situation, but the slope limit helps a lot.  A glyph like
"." is in effect kerned as if it looked like "/" and "\" on its two sides,
preventing it from sliding too far under any other glyph.

Using "until the contours touch" is undesirable because it places too much
weight on the single point of closest approach, which may not be visually
significant.  This effect is visible in serif fonts with pairs like "AV",
where there may be serifs that should be brought much closer together than
the distance you would generally want for flat parts of glyphs;
auto-kerning algorithms in general tend to put such pairs too far apart
because of putting too much weight on the serifs.  You want to give some
weight to both the nearest-approach points, and other parts of the glyphs
that may be a little further apart; that requires something a little
softer than a pure touch/don't touch test.  Blurring may be intended to
serve a similar purpose, but my own code uses nonlinear weighting on
the hard outlines in order to produce a more controllable effect.  The
exponent of 3 seems to produce the most appealing results in my tests.

As I mentioned, there are still some issues with glyphs that have very
little vertical extent, such as apostrophe.  I've also found that the
adjustable parameters in the algorithm need to be different for Latin,
Japanese, and Korean glyphs, in order to produce results that mix well
with each other.  The spacing readers expect between characters in those
writing systems is simply different.

I don't have a link for just the documentation file describing this, but
both the code and the PDF user manual explaining how it works are in the
Tsukurimashou distribution package available here:
   
http://en.sourceforge.jp/projects/tsukurimashou/downloads/54223/tsukurimashou-0.5.zip/

-- 
Matthew Skala
address@hidden                 People before principles.
http://ansuz.sooke.bc.ca/