[magick-users] spreading edge colors (WAS filter image to use in opengl)

Tue Jan 6 04:12:11 PST 2009

Hi Anthony,
Thank you for your answer you wrote a lot of interesting information. I 
found something intersting with gimp.
There is a filter called erode, it seems to do what i would want, so i 
will check what is it exactly to be sure it's what i want.
Anyway with your information i will be able to do something near.

Thank you

Cheers,
Cedric

Anthony Thyssen wrote:
> On Tue, 6 Jan 2009 11:38:21 +1000
> Anthony Thyssen <A.Thyssen at griffith.edu.au> wrote:
>
> | On Mon, 05 Jan 2009 11:49:42 +0100
> | magick-users at imagemagick.org wrote:
> |
> | | Hi,
> | |
> | | I have an image exported by inkscape to use in a 3d program, my problem
> | | is that the border of the image is not really conveniant with the alpha
> | | blending.
> | | What i would like is to process the image to extend the border of the
> | | rgb then the aplha will fade to the good color of border and not a
> | | background flat white.
> | | I thought to a filter that could detect the border of image then extend
> | | it, but i did not success to do it. I am not really experienced in
> | | filter processing, so advise will be welcome.
> | |
> | | here the picture to understand what i am explaining
> | |
> |
> | ... previous response deleted ...
> |
>
> Looking again at your email....
>
> One of the things you mentioned was..
> | What i would like is to process the image to extend the border of the
> | rgb...
>
> This can be tricky, and is something that I have thought about a lot
> over the last couple of years.
>
>
> The following is a running experiment, so please bear with me.
>
>
> So lets start with a test image...
>
>   convert -size 100x100 xc:none \
>           -draw "fill magenta path 'M 30,60  A 1,1 0 0,0 70,40 z'" \
>           -draw "fill cyan    path 'M 30,60  A 1,1 0 1,1 70,40 z'" \
>           -draw "fill blue    circle 50,50 37,37'" \
>           circle_halves.png
>
> Ok the problem here is that any pixel that his NOT fully-transparent
> will have an undefined color.  Generally and mathematically algorithms
> tend to set that undefined color to black.
>
> (Scale the image larger to see it better...
>
>   convert circle_halves.png  -scale 500%  show:
>
> Note that there are properly defined semi-transparent pixels
> around the edge.
>
> For example...
>
>    convert circle_halves.png  -alpha off  show:
>
> What we want to do is define the hidden color of these pixels as
> something other than black.  Some color that is a better representation
> of the real colors of the visible object in the image.
>
> There are a couple of methods that can be used for this, but the first
> step is always the same.
>
> First we need to figure out what pixel have a properly defined color
> (anything which is opaque or semi-transparent) and what parts do not
> have a properly defiend color.
>
> First thing is to take the original image and threshold the alpha channel
> so that any pixel that is NOT fully-transparent, becomes fully-opaque.
> That is convert it to a boolean transparency, which gives us a mask
> defining the pixels that we know has a 'good' color that we will not want
> to change.
>
>   convert circle_halves.png \
>           -channel A -negate -threshold 0 -negate +channel \
>           defined_color.png
>
> Now we need to some how spread that good color to the other undefined
> color pixels.
>
> One method is to blur the 'mask of good colors' with a -channel RGBA setting.
> This will make each pixel an average color of its neighbours, and the
> color will spread into the undefined transparent areas.
>
> Note that we don't want to change the properly defined colors
> so we also overlay the good colors on top of the result, to resutl them
> back to there correct values.
>
>   convert defined_color.png \
>           \( +clone -channel RGBA -blur 0x5 \
>              -channel A -negate -threshold 0 -negate +channel \) \
>              -compose DstOver -composite      blurred_color.png
>
> Note that the above blur did manage to color most of the undefined
> pixels, and the color were quite a reasonable result, at a distance from
> the original image. however it fails close in to the original image,
> producing a sharp discontinuity.  This is because it is average of all
> the defined colors, especially the inside colors, rather that a
> spreading of just the edge colors.
>
> The better solution is a actually to blur the image using a radius of
> only 1 to 2 pixels at most, so that new colors will only get a average
> of the closest neighbours.
>
> A mophological operation to average neighbours.
>
> By repeating this small blur, and reseting the defined pixel colors
> we can incrementally spread the objects edge color out to produce a
> better result.
>
>   convert circle_halves.png \
>           -channel A -negate -threshold 0 -negate +channel \
>           spreading_color.png
>   for i in `seq 10`; do
>     convert spreading_color.png \
>             \( +clone -channel RGBA -blur 1x65000 \
>                -channel A -negate -threshold 0 -negate +channel \) \
>              -compose DstOver -composite     spreading_color.png
>   done
>
>   convert spreading_color.png  -scale 500%  show:
>
> That worked but not very nicely. We got a 45 degree join, rather than
> the original angle of the edges in the original image.  A different
> mophological operation may have resulted in a horizontal or vertical
> joint instead.
>
> Repeating the above with a radius of 2 actually produces a better result
> also not doing the thresholding also improves the result.
>
>   convert circle_halves.png  tmp.png
>   for i in `seq 35`; do
>     convert tmp.png \( +clone -channel RGBA -blur 2x65000  \) \
>              -compose DstOver -composite     tmp.png
>   done
>   convert tmp.png -channel A -negate -threshold 0 -negate spreading_color2.png
>
>   convert spreading_color2.png  -scale 500%  show:
>
> Now that you have your defined colors, all you need to do is re-add the
> original images transparency, which is a very difficult thing to do.
>
> All the normal Alpha Composition methods appear to fail.
>
> The ONLY solutions I have found for this is either using the slow -fx
> operator or to seperate and combine the image channels.
>
>   convert spreading_color2.png circle_halves.png \
>           -channel RGBA -separate -delete 3-6 -combine \
>           circle_halves_restored.png
>
> Of course that looks exactly like your original image, but then that is
> the point.  You can verify that the transparent pixels have a nicely
> defined color using, just as we did before...
>
>   convert circle_halves_defined.png  -alpha off  show:
>
> Be warned however that almost any typical IM operation will likely reset
> those transparent colors back to the undefined state of black.
>
> ---
>
> An alternative for all this is to pull out all the edge pixels of the image
> and input them in the new   -sparse-colors  operator
>   http://www.imagemagick.org/Usage/canvas/#sparse-color
>
> As this operator is designed to deal with a very simular problem.
>
> The question of course is how to extract a list of edge pixels and there
> colors from an image.
>
>
>   Anthony Thyssen ( System Programmer )    <A.Thyssen at griffith.edu.au>
>  -----------------------------------------------------------------------------
>    Judith looked at the demons and shook her head.  "I'm glad we didn't
>    build four processors.  I'm not sure I could take a barbershop
>    quartet!"                      -- Rick Cook, "The Wizardry Compiled"
>  -----------------------------------------------------------------------------
>      Anthony's Home is his Castle     http://www.cit.gu.edu.au/~anthony/
>   

-- 
+33 (0) 6 63 20 03 56  Cedric Pinson mailto:mornifle at plopbyte.net http://www.plopbyte.net