TYPE DESIGN INFORMATION PAGE last updated on Mon Mar 25 16:21:32 EDT 2024

SEARCH THIS SITE:

IMAGE SEARCH:

FONT RECOGNITION VIA FONT MOOSE

LUC DEVROYE


ABOUT







Kerning: A survey [Sebastian Kosch]

Sebastian Kosch surveys existing kerning methods in 2019 as part of his research for his own Google-supported method called YinYangFit. A verbatim excerpt:

  • Fixed-distance methods: A family of approaches that insert pre-defined distances between letter pairs. In their simplest incarnation, these heuristics are equivalent to simply adding sidebearings to every letter, without any kerns. Kernagic, inspired by Frank Blokland's research, uses heuristics to identify stems or stem-equivalents (such as the round sides of an o) in every letter shape, and then aligns them. This works reasonably well with very regular type (think blackletter), but manual adjustments are usually required. Less well known is Barry Schwartz's anchor point implementation of what amounts to basically the same idea. Adrian Frutiger, Walter Tracy and Miguel Sousa have devised similar systems, described in Fernando Mello's MATD thesis. The legendary Hz-Program is also included in this category, as its patent application reveals that letter pair distances were simply stored in a hardcoded table.
  • Gap area quadrature: A family of algorithms that attempt to quantify and equalize the perceived area of the inter-letter gap. The crux, of course, lies in deciding where the gap ends. HT Letterspacer, the crudest one of these tools, considers everything between baseline and x-height (modulo some minor refinements). Simon Cozens's CounterSpace uses blurs and convex hulls to more effectively exclude regions that arguably don't belong to the gap (such as the counter of c). My own Electric Bubble model measures Euclidean instead of horizontal distances, but imposes geometric constraints that produce similar results to CounterSpace. CounterSpace currently wins in terms of performance-complexity ratio but it, too, struggles to fit certain letter pairs.

  • Other shape-based methods: These include more exotic approaches, such as stonecarver David Kindersley's wedge method from the 1960s, which operated on letter area moments of inertia (and didn't really work), and iKern, which produces great results but, just like Adobe's Optical Kerning feature, remains unpublished. Last but not least, the TypeFacet Autokern tool identifies parts of letter outlines that jut out horizontally, and adds kerning to compensate, based on a few parameters.
  • Neural nets: Yes, we can train convolutional nets to recognize images of well-fitted and poorly-fitted type. Simon Cozens has built several versions of his kerncritic model (formerly AtoKern), and the recent ones perform surprisingly well on many (if not all) pairs. While neural nets are fascinating, they tend to be black boxes: we can only make guesses at how they work, and we cannot tune their behaviour to suit our taste. This problem holds not just for convolutional nets, but for statistical function approximators in general.
  • Honorable mention: Bubble Kerning is a proposal that type designers draw a bubble around every letter, such that software can automatically find pair distances by simply abutting the bubbles. While this isn't technically a letterfitting heuristic at all, it’s still worth mentioning as a neat idea that could perhaps save designers some time. Toshi Omagari has built a Glyphs plugin.

EXTERNAL LINKS
Kerning: A survey
MyFonts search
Monotype search
Fontspring search
Google search

INTERNAL LINKS
Font software ⦿ Typography ⦿













Luc Devroye ⦿ School of Computer Science ⦿ McGill University Montreal, Canada H3A 2K6 ⦿ lucdevroye@gmail.com ⦿ http://luc.devroye.org ⦿ http://luc.devroye.org/fonts.html