Fuzz Distance and Transparent Colors -- FIXED

ImageMagick Examples Preface and Index
Known and Fixed Bugs Index
Exploration of the distance algorithm used by -fuzz color matching algorithm. Especially with regards to transparent colors.

The Fuzzy Distance Matching formula should
  1. Oaqpue color distance should reflect the color space (EG: RGB or CMY color cubes, or HSL cones etc). Specifically Black to White color distance should be 100%
  2. All Color with full transparency, should be classed as being identical. That is they have zero color distance.
  3. Two colors with partical transparency should be closer thna the same colors without any transparency (opaque)
  4. Transparent colors should have some distance non-transparent colors.
  5. Fully-Transparent should be equally distance from all fully-opaque colors

The current IM formula (before IM v6.6.6-4) follows all but the last of these recommendation, which causes some problems when transparency is involved.

For example... Here we fuzzy match at various percentage distances between the fully-transparent Black (or 'None' and a color wheel of opaque colors...

  for P in 50 60 80 100 110; do
    convert colorwheel.png -alpha set -channel RGBA \
            -fuzz $P% -fill none -opaque none \
            -fill black -gravity SouthWest -annotate +2+2 "$P%%" \
  done |\
    montage - -tile x1 -background none -geometry +2+2 fuzz_none_color.png

[IM Output]

As you can see any fully-opaque near black color will match the fully-transparent black color 'none' before other fully-opaque colors, between 50 and 60%

In other words the color 'None' is currently closer to black, than other opaque colors.

Stranger still white does not match until just after 110%!

Here is another example that clearly shows that opaque colors are not the same distance from full transparency. This generates a greyscale and transparency gradient, and then replaces 'similar colors' to full-transparency.

  convert -size 100x100 gradient: \( +clone -rotate 90 \) +swap \
          -compose CopyOpacity -composite  gradient.png
  convert gradient.png -channel RGBA \
          -fuzz 25% -fill Red -opaque None  fuzz_gradient_25.png
  convert gradient.png -channel RGBA \
          -fuzz 50% -fill Red -opaque None  fuzz_gradient_50.png
  convert gradient.png -channel RGBA \
          -fuzz 75% -fill Red -opaque None  fuzz_gradient_75.png

[IM Output] ==> [IM Output] [IM Output] [IM Output]

Remember the distance from 'None' to 'White' is just over 110% ! And all fully-transparent colors are 0 distance (regarded as equal).

As of IM v 6.6.6-4, bug has been fixed. Basically the IsMagickColorSimilar() function was set to use the equivelent of the last formula below.

Here is what the LAST image in the above looks like from IM v6.6.6-4 on, where the fuzz factor has been fixed with regard to transparencies.

  convert gradient.png -channel RGBA \
          -fuzz 75% -fill Red -opaque None  fuzz_gradient_75_fixed.png

[IM Output]

As you can see all opaque colors will now be treated as an equal 100% distance from fully-transparent.

Note however that a semi-transparent color, does still have a valid color, and as such opaque colors will not all be equal distance it. However the close a color is to fully-transparent, the more equidistant the opaque colors are from it. That is as it should be.

FX Formulas...

These are expressed as FX formulas where 'u' is the first image and 'v' is the second image, See Using FX, The DIY Image Operator, in the expresion...

  convert xc:color1 xc:color2 -print \
    "%[fx:...expression...]%%" \

RGB Opaque Color Distance...

    "%[fx:(100)*sqrt(( (u.r-v.r)^2 +
                       (u.g-v.g)^2 +
                       (u.b-v.b)^2  )/3 )]%%" \

(no transparency handling)

RMSE Distance?

    "%[fx:(100)*sqrt(( (u.r-v.r)^2 +
                       (u.g-v.g)^2 +
                       (u.b-v.b)^2 +
                       (u.a-v.a)^2  )/4 )]%%" \

(But "compare -metric RMSE" different values!)

Color with Alpha Multiply

    "%[fx:(100)*sqrt(( (u.r*u.a-v.r*v.a)^2 +
                       (u.g*u.a-v.g*v.a)^2 +
                       (u.b*u.a-v.b*v.a)^2  )/3 )]%%"

(This results in 'Black' == 'None' )

Current (buggy) fuzz calculation being used (to IM v6.6.6-3)

    "%[fx:(100)*sqrt(( (u.r*u.a-v.r*v.a)^2 +
                       (u.g*u.a-v.g*v.a)^2 +
                       (u.b*u.a-v.b*v.a)^2 +
                       (u.a-v.a)^2          )/3 )]%%"

(Added transparency difference, but Black still closer to None than white)

The correct calculation should be (implemented IM v6.6.6-4)

    "%[fx:(100)*sqrt( ( (u.r-v.r)^2 +
                        (u.g-v.g)^2 +
                        (u.b-v.b)^2 )*u.a*v.a/3   + (u.a-v.a)^2  )  ]%%"

Note how the 3-D RGB color distances is vastly simplified and handled, almost as a completely separate item to the alpha channel distance.

Also note that if either color is fully transparent, the actual color become irrelevent, and the fuzz factor becomes strictly a simple alpha distance fuzz factor.

In this scheme...

Other color space can also easilly use this formula as the 'color space' distance is just a simple and complete component of the additional alpha channel component.

UPDATE: this color distance metric is now also used for "compare -metric Fuzz"

Created: 7 December 2010
Updated: 8 December 2010
Author: Anthony Thyssen, <A.Thyssen@griffith.edu.au>
Examples Generated with: [version image]
URL: http://www.imagemagick.org/Usage/bugs/fuzz_distance/