Ximagic Quantizer

[brodriguez]

If you continue testing you will get out of paper

About the DBS, no, I didn't heard about it, but as you mention it I look for information about it and it seems a very good method, at least from the error point of view.

I've find one good page, with some numbers
http://caca.zoy.org/study/part4.html

I'll try to implement it, although it can take some time, perhaps the next week. I'm not too much worried about its processing time, if the results are so good, this is something that could be solved later.

About your suggestion of new dithers, It is something that could be done configuring the threshold map of an ordered dither?

If it is, I plan to implement some kind of configurable ordered dither like the one included in imagemagick library.

You can find an explanation of that idea at http://www.imagemagick.org/Usage/quanti ... sholds_xml

If it is not, I'm ready for a longer explanation

I published a new version (1.83) which include the diag5x5 pattern mentioned in the imagemagick page, you could test it, although it seems too much artistic.
The other change in this version is the reduction in the error diffusion step from 10 to 5, to allow finer setting.

[eL_PuSHeR]

Nice! I will test it later when I came back from work. As for DBS, its results seems promising (at least for minimizing error-diffusion artifacts).

You are right about configurable pattern dithers. That would yield a lot of flexibility.

I have tried the diag5x5 dither and it seems good enough for printing line-art.

[brodriguez]

I just published a new version with DBS (1.90) including random dither and Direct Binary Search.

It takes me less time than expected because I've found an implementation available with source.

I try DBS a bit, it is artifact free depending on the previous process. If there are artifacts in the image before aplying DBS the result is bad otherwise it is good.

Although in the page explaining the library it seems the method should be started with some kind of dithering (f.i. random) the result when the method is applied inmediately after a 1 bit conversion is almost equal.

The tone don't match the original very well, sometimes are darker and sometimes lighter. I don't know if this is due to the implementation or is a problem of the method itself.

Try yourself

[eL_PuSHeR]

Thank you very much! I will try it and post results.

My findings so far:

1) With random ordered dither I got a very noisy output image
2) Starting with Gamma 1.8 (being 1 default or none), I got very nice results with ordered ort6x6 for the photo faced I posted earlier.
3) Upped Gamma to 2.0 (to compensate dot gain) and I got a too blackened image printed with DBS. I am testing now with higher gamma values. On the other hand, mixing a FAN error-diffusion dither + DBS got me a nice almost artifact-free image. Maybe I should try printing at 300 dpi instead. At 600dpi pixels are too near each other.
By the way, DBS took several minutes to complete. I have an AMD Phenom X4 9600 Black Edition, not the fastest, working on one core.

4) Gamma 5 (a lot) did the trick for DBS, but I prefer the grayscale result with ort6x6. Grays are more accurate. With DBS and error-diffusion in general, I got a lot of contrast.

[oops66]

-

[oops66]

...
http://www.ximagic.com

Hello,
Just to say: Congratulations, Ximagic Quantizer is a very interesting plugin !

PS: Within this new phpBB version I Do not know the way to delete a personal post now ( like above), any idea ?

[eL_PuSHeR]

I have thoroughly testing DBS and I find it no real-world benefits using it. At 600dpi it tooks a lot to complete and, I don't know why, all images are darkened a lot (at least for me). On the other hand, it is true that it breaks up almost all visible patterns (artifacts) from error-diffussion filters. I know it's in a very early stage, so we will continue testing.

[eL_PuSHeR]

I have compiled some information about dithering and such over the years. I don't know whether it proceeds to put it here or not, but I couldn't find a suitable way for sending it to you. I will put some excerpts for avoiding doing humongous posts.

More info about some error-diffusion filters:

The following info is from "Bitmapped Graphics" by Steve Rimmer,
published by Windcrest/McGraw-Hill ISBN 0-8306-3558-0. A second edition
has come out since but I don't have the ISBN for it. It's a great little
book and covers numerous popular file formats, dithering, screen and
printer drivers, PostScript etc.....

Ok, now to business. The Floyd-Steinberg filter looks like this:

x 7
3 5 1

In this case all the numbers add up to 16. Now you look at the pixel at
position x and threshold it, i.e. if your grey values are between 0 and
255 then you compare it to 127. If it's lower or equal to it then you set
the pixel to 0, if higher then you set it to 255 (or 1 if you're working
in a monochrome mode). Now you calculate the error as being the
difference between the original grey scale value and the value that you
changed it to. You then add 7/16ths of this error to the pixel to the
right, 5/16ths of the error to the pixel underneath etc... You then move
to the next pixel and do the same thing and so on and so on untill you've
done it to every pixel in the image. Note that the total amount you
spread around equals the error value, so your image intensity stays the same.

Now Rimmer says that that you can get better results by using filters
which "communicate" with more of the surrounding pictures.

Here's a few more filters he gives:

Stucki
x 8 4
2 4 8 4 2
1 2 4 2 1

This one is slower than Floyd-Steinberg since it can't be done with bit
shifts and subtraction.

Burkes
x 8 4
2 4 8 4 2

This one is a compromise between Floyd-Steinberg and Stucki, spreads
error over more pixels but still can be done with shifts and subtractions.

Sierra
x 5 3
2 4 5 4 2
2 3 2

Jarvis, Judice and Ninke filter:

x 7 5
3 5 7 5 3
1 3 5 3 1

Stevenson and Arce
x 32
12 26 30 16
12 26 12
5 12 12 5

Usually gives the best results but it's slow as all hell.

Rimmer also says that using a serpentine scan (ie left to right on first
line, right to left on next, left to right, right to left etc...)
produces "different" results. He also recommends some good reading on the
subject.

Oh yeah, the book also discusses how enlargening images before dithering
them often makes them look heaps better.

Unfortunately in this edition he doesn't discuss things like 256 color ->
16 color, but I believe that the technique is the same, you just have to
match each color to the "closest" 4 bit color available and spread the
error for each RGB component. If you get a color version working I
wouldn't mind seeing the results.

For printing at 600dpi B/W, Stevenson & Arce filter's hexagonal grid provides the most accurate grays.

More info
More info 2

A little off-topic. Has anyone heard of Edge-directed Interpolation. It seems similar to B-Spline and looks cool.

Regards.

PS - Crappy forum text display. All filters are bad (misaligned)

[brodriguez]

Hello,
Just to say: Congratulations, Ximagic Quantizer is a very interesting plugin !

Thanks

My findings so far:
1) With random ordered dither I got a very noisy output image
2) Starting with Gamma 1.8 (being 1 default or none), I got very nice results with ordered ort6x6 for the photo faced I posted earlier.
3) Upped Gamma to 2.0 (to compensate dot gain) and I got a too blackened image printed with DBS. I am testing now with higher gamma values. On the other hand, mixing a FAN error-diffusion dither + DBS got me a nice almost artifact-free image. Maybe I should try printing at 300 dpi instead. At 600dpi pixels are too near each other.
By the way, DBS took several minutes to complete. I have an AMD Phenom X4 9600 Black Edition, not the fastest, working on one core.
4) Gamma 5 (a lot) did the trick for DBS, but I prefer the grayscale result with ort6x6. Grays are more accurate. With DBS and error-diffusion in general, I got a lot of contrast.

Yes, the output of the random dithering is not nice, it is only included because the DBS page I've found seems to recommend it as the starting point of the algorithm. Anyway it is used sometimes as a learning I tool. The result with color images are not so ugly. Try with your test image (evawyrwa) using only random dithering with 6,6,6 levels (216 colors)

I've also found DBS needs too much correction of gamma, which is not what one could expect from an algorithm that take HSV (Human vision system) into account. It should be something wrong in the implementation.

I know DBS is very slow, and it is using only 1 core, I use a q6600 (overclocked to 3,2) for the tests. If I could find what it is wrong and get better results I'll try to make it use all the cores, but otherwise it don't make sense to spend time in this algorithm.

I have thoroughly testing DBS and I find it no real-world benefits using it. At 600dpi it tooks a lot to complete and, I don't know why, all images are darkened a lot (at least for me). On the other hand, it is true that it breaks up almost all visible patterns (artifacts) from error-diffussion filters. I know it's in a very early stage, so we will continue testing.

I suspect it is not usable for printing but it gives nice results for display. Somewhere I read it is used for dithering pattern design.

I have compiled some information about dithering and such over the years. I don't know whether it proceeds to put it here or not, but I couldn't find a suitable way for sending it to you. I will put some excerpts for avoiding doing humongous posts.

Its it clear you know what you are talking about.

You can send me a private message using the PM icon which you can find below my username in one of my posts.

Stevenson and Arce
x 32
12 26 30 16
12 26 12
5 12 12 5

Usually gives the best results but it's slow as all hell.

I added Stevenson & Arce to the TODO list.

Although it is not very popular, I found SnagIt includes this dithering.

Unfortunately in this edition he doesn't discuss things like 256 color ->
16 color, but I believe that the technique is the same, you just have to
match each color to the "closest" 4 bit color available and spread the
error for each RGB component. If you get a color version working I
wouldn't mind seeing the results.

I don't know if this text is part of the reference or of your message.
If it is part of your message, the plugin quantize and dither color images. In fact I didn't include B/W quantizations or ditherings initially. Thanks to your posts I pay more attention to the B/W question.

[brodriguez]

Because the last post was long enough I prefer to make a new one to make it more readable.

As it seems I'm lucky enough to get feedback about the plugin I've updated my TODO for it

• Work on DBS, gamma and speed.
  Allow configurable threshold dithering.
  Allow configurable error diffusion dithering.
  Allow grayscale input/output.
  Include Stevenson Arce dithering.

Suggestions

There are 2 issues which apparently can't be solved due to the structure of PS plugins.

• Output indexed or bitmap format. The output format should be equal to the input one.
  Change resolution (good for B/W and dithering). The output resolution should be equal to the input one.

Regards

[eL_PuSHeR]

Well the resolution changing is not a big issue as I always resize/resample my input images to match print resolution. That's another reason why DBS takes at least ten minutes to complete. Working at 600-1200dpi true colour is a heavy task.

I am anxiously waiting for those configurable threshold dithers. So far, I have got superb results printing directly to printer as Postscript at 1200 dpi resolution, which yields me a very nice output using ort 8x8 (150 LPI) and ort 16x16 (75 LPI). Quality is superb for my Laserjet 1200. I've never seen anything like that, no matter what drivers I have tried. Quality is also good at 600dpi.

Keep up the good work!

[brodriguez]

I am anxiously waiting for those configurable threshold dithers

You don't need to wait too much.

I've published a new version with configurable threshold dithering (2.00)

The threshold patterns can be loaded into the plugin from a file. The loaded patterns are stored with the plugin parameters and are available until the user deletes it. The format of the files is explained in the plugin web site (usage page).

This version is still limited in the sense the pattern should be the same for all the channels, as with the predefined threshold ditherings.

So far, I have got superb results printing directly to printer as Postscript at 1200 dpi resolution, which yields me a very nice output using ort 8x8 (150 LPI) and ort 16x16 (75 LPI). Quality is superb for my Laserjet 1200. I've never seen anything like that, no matter what drivers I have tried. Quality is also good at 600dpi.

I've patched an almost unnoticeable bug in the 16x16 ort pattern. One of the values at the center of the threshold matrix was wrong.

[eL_PuSHeR]

More testing! good.



EDIT: I don't understand anything about creating custom user threshold patterns. Could you please elaborate it a little more?

Well, I have created three filter purely by using trial & error. The results aren't so bad as I expected them to be.
Anyway, I am getting little banding and a somwhat bland tone. Something must be wrong (no surprise here). Care to comment?
My attempt of creating some 10x10, 12x12 and 14x14 ort filters.

eL_PuSHeR_10x10
10 10 100
073 074 075 076 077 078 079 080 081 082
072 043 044 045 046 047 048 049 050 083
071 042 021 022 023 024 025 026 051 084
070 041 020 007 008 009 010 027 052 085
069 040 019 006 001 002 011 028 053 086
068 039 018 005 004 003 012 029 054 087
067 038 017 016 015 014 013 030 055 088
066 037 036 035 034 033 032 031 056 089
065 064 063 062 061 060 059 058 057 090
100 099 098 097 096 095 094 093 092 091

eL_PuSHeR_12x12
12 12 144
111 112 113 114 115 116 117 118 119 120 121 122
110 073 074 075 076 077 078 079 080 081 082 123
109 072 043 044 045 046 047 048 049 050 083 124
108 071 042 021 022 023 024 025 026 051 084 125
107 070 041 020 007 008 009 010 027 052 085 126
106 069 040 019 006 001 002 011 028 053 086 127
105 068 039 018 005 004 003 012 029 054 087 128
104 067 038 017 016 015 014 013 030 055 088 129
103 066 037 036 035 034 033 032 031 056 089 130
102 065 064 063 062 061 060 059 058 057 090 131
101 100 099 098 097 096 095 094 093 092 091 132
144 143 142 141 140 139 138 137 136 135 134 133

eL_PuSHeR_14x14
14 14 196
001 002 003 004 005 006 007 008 009 010 011 012 013 014
052 053 054 055 056 057 058 059 060 061 062 063 064 015
051 096 097 098 099 100 101 102 103 104 105 106 065 016
050 095 132 133 134 135 136 137 138 139 140 107 066 017
049 094 131 160 161 162 163 164 165 166 141 108 067 018
048 093 130 159 180 181 182 183 184 167 142 109 068 019
047 092 129 158 179 192 193 194 185 168 143 110 069 020
046 091 128 157 178 191 196 195 186 169 144 111 070 021
045 090 127 156 177 190 189 188 187 170 145 112 071 022
044 089 126 155 176 175 174 173 172 171 146 113 072 023
043 088 125 154 153 152 151 150 149 148 147 114 073 024
042 087 124 123 122 121 120 119 118 117 116 115 074 025
041 086 085 084 083 082 081 080 079 078 077 076 075 026
040 039 038 037 036 035 034 033 032 031 030 029 028 027

[eL_PuSHeR]

Testing version 2.10 - It seems load from image doesn't work, at least for the supplied examples.

[brodriguez]

EDIT: I don't understand anything about creating custom user threshold patterns. Could you please elaborate it a little more?

From the patterns you have created, I suspect you don't need the explanation, but I'll try to make one for those who know less about the subject.

A good and short explanation of dithering methods:
http://users.ece.utexas.edu/~bevans/pro ... thods.html

These halftone methods may be classified into three categories--- amplitude modulation (AM), frequency modulation (FM) and AM-FM hybrid halftoning. In AM halftoning, the dot size is varied depending on the graylevel value of the underlying grayscale image while the dot frequency is held constant, e.g. clustered-dot ordered dither. FM halftones have a fixed dot size and shape, but the frequency of the dots varies with the graylevel of the underlying grayscale image. Conventional digital FM halftones have a fixed dot size of one pixel, e.g. those produced by dispersed-dot ordered dither and error diffusion. AM-FM halftones have variable dot shape/size, and variable dot frequency that depends on the graylevel value to be reproduced. Examples of AM-FM halftones include "green-noise" halftones by Levien, halftones on space filling curves, and halftones with texture control.

So let classify dithering by purpose and method

• 1.- Get the best matching image with a reduced color set.
  • FM (dispersed-dot) The most used are Bayer method. The void and cluster seems better but is is covered by patents. (Diffusion patterns in the plugin).
    AM (clustered-dot) There are a lot of method is this area. The most used ones are those with circular/eliptical clusters. (Halftone patterns in the plugin)
    AM-FM. The following article refers to it as 2nd generation FM http://webstaff.itn.liu.se/~sasgo/Print ... e_2008.ppt The plugin doesn't support this kind of dithering.
  2.- Get an artistic transformation of an image.
  • A.- Some kind of "artistic" clustered-dot (like the Diag 5x5 in the plugin)
    B.- Mixing the image and the dither pattern.http://www.iro.umontreal.ca/~ostrom/pub ... echArt.pdf (Not supported yet)

So, you can use User threshold to create your own void & cluster implementation (http://caca.zoy.org/study/part2.html) or create your own clustered dot with or without artistic purpose.

I'm working now to implement function 2.B (mixing image with dither pattern).

Well, I have created three filter purely by using trial & error. The results aren't so bad as I expected them to be.
Anyway, I am getting little banding and a somwhat bland tone. Something must be wrong (no surprise here). Care to comment?
My attempt of creating some 10x10, 12x12 and 14x14 ort filters.

It look great on color image.

The only think mistake I saw in the patterns is mine, the divisor should be max + 1.

Testing version 2.10 - It seems load from image doesn't work, at least for the supplied examples.

I added this functionality because I think it will be easier to work dither patterns as images, and it will be a must for artistic dithering type 2B.

You can download the Diffusion 8x8 and Halftnone 8x8 patterns as images from the usage page.
I have also converted your patterns to png images


This could be a problem in the plugin or in the instructions.

Lets try with the instructions

I'm not sure I give clearly enough instructions. The load from image get the dither pattern from the image open in the editor, so the process should be:

• Open the dither pattern image in the editor (o make it the selected image)
  Start the plugin
  Press Load button.
  Fill pattern name, select "from image" in the Load Dialog and press OK button.
  Now the pattern should appear in the user threshold drop down.
  Cancel the plugin

Now you have the pattern loaded and you can use it to dither images.

If the above steps don't load the dither pattern, or the pattern don't work ok, it should be a bug in the plugin.

Can you explain why it don't work with more detail?
The pattern is not loaded or it is not applied correctly ?
I also need to know the image editor you are using.

Regards