forced palette for RGBA images

[Paul Spencer]

Thomas, Assefa,

how is this approach different from the quantization that Steve/Assefa  
implemented?

I believe we are currently using the gd-based quantization with a  
limited palette of primary colours that we want to preserve and a  
number of open colour slots that are filled by gd for covering the  
antialiasing values, and that it works quite well to reduce image size  
to better level without causing horrendous loss of image quality.

Does this new algorithm work the same way?  Why would it be used over  
the gd method?  It seems to be substantially slower than the gd  
quantization.  One use case for quantization that we have is in  
generating map tiles for a large, static cache.  This additional  
overhead would be pretty substantial for this use case.

Cheers

Paul

On 12-Dec-07, at 12:02 PM, Daniel Morissette wrote:

> Could we please have a ticket for this referred from the SVN change  
> comment, summarizing the info from this thread, just so that we have  
> a place to point users to when they ask about this change in a few  
> months?
>
> Personally I'd like to see tickets for anything that's not a trivial  
> bugfix... but that's just me of course.
>
> Daniel
>
>
> Steve Lime wrote:
>> Thomas, these are great additions. I'm looking forward to using  
>> them myself as I'm dealing
>> with image size issues with AGG/OpenLayers. On your questions, if  
>> you think it might be disruptive then perhaps users should opt in  
>> (e.g.
>> --enable-experimental-png).
>> I'd go ahead and commit. No sense waiting and that gives more folks  
>> a chance to test.
>> Steve
>>>>> On 12/12/2007 at 2:46 AM, in message
>> <d2b988930712120046n1c0dce28ye568b9c693040f0d at mail.gmail.com>,  
>> thomas bonfort
>> <thomas.bonfort at GMAIL.COM> wrote:
>>> A quick update on this.
>>> I've now also added quantization for rgba images ( based on the  
>>> median
>>> cut implementation of pngquant / ppmquant:
>>> http://www.libpng.org/pub/png/apps/pngquant.html ).
>>>
>>> examples here (the first two images are actually transparent. you'd
>>> have to open them in an image editor to see that):
>>>
>>> the full rgba png:
>>> http://www.terriscope.fr/ms_tmp/rgba.png : 129k
>>>
>>> the quantized rgba png (may look funny under oldish MS IEs)
>>> http://www.terriscope.fr/ms_tmp/palette.png : 39k
>>>
>>> the output produced with the current gd quantization
>>> http://www.terriscope.fr/ms_tmp/rgbapalettegd.png
>>> I'd personally enable the pngquant quantization for rgb images  
>>> to , as
>>> it produces higher quality results, but there's a tradeoff on
>>> rendering speed (rough times for quantization only: 1.5 sec for
>>> pngquant vs 0.3 sec for gd, on a 2000x2000 image). Or this could be
>>> yet another formatoption selectable option.
>>>
>>> Dan: all this will be documented. We've talked with Assefa about
>>> either writing a complete reference on the different formatoptions,
>>> and/or a howto on reducing image size with agg (palette,  
>>> quantization,
>>> and tradeoffs beween speed/imagesize/flexibility)
>>>
>>> Now for the questions:
>>> * how should I add this? I'd go for a configure option
>>> (--disable-experimental-png) with ifdef'd code as I don't know how
>>> disruptive this would be on all platforms.
>>> * can I add this soon? It's stable on my setup, but there might be
>>> issues on other platforms. I'd like to have it tested as much as
>>> possible before 5.2
>>>
>>> cheers,
>>> thomas
>>>
>>> On Dec 7, 2007 10:46 PM, Daniel Morissette  
>>> <dmorissette at mapgears.com> wrote:
>>>> Sounds like a cool feature. We'd need to make sure it is well  
>>>> documented
>>>> and not just hidden in the source. A short RFC or ticket for this  
>>>> maybe?
>>>>
>>>> Daniel
>>>>
>>>>
>>>>
>>>> thomas bonfort wrote:
>>>>> hello list,
>>>>>
>>>>> I've added the possiblity to force a color palette for RGBA  
>>>>> images. As
>>>>> gd does not support this kind of images, I had to resort to  
>>>>> directly
>>>>> use libpng functions for the actual saving of the file.
>>>>>
>>>>> The steps needed to actually use this would be:
>>>>> - create a 24 bit rgba image representative of a typical image  
>>>>> that
>>>>> would be produced
>>>>> - use an external program to create an rgba paletted version of  
>>>>> the
>>>>> image (pngquant, pngnq, photoshop...)
>>>>> - extract the palette from the created image and save the values  
>>>>> in
>>>>> the form r,g,b,a in a text file (one per line)
>>>>> - use palette_force and palette formatoptions in the  
>>>>> outputformat block.
>>>>>
>>>>> This is rather complicated, but usefull on production setups to
>>>>> minimize image size when using clients that overlay layers (eg
>>>>> openlayers).
>>>>>
>>>>> There are still a few quirks I'd have to work on before the code  
>>>>> is
>>>>> robust enough for inclusion in the trunk, so I'd like to know if  
>>>>> this
>>>>> is a feature we'd definitely like to have before putting too much
>>>>> effort into that.
>>>>>
>>>>> cheers,
>>>>> thomas
>>>>
>>>> --
>>>> Daniel Morissette
>>>> http://www.mapgears.com/
>
>
> -- 
> Daniel Morissette
> http://www.mapgears.com/

+-----------------------------------------------------------------+
|Paul Spencer                          pspencer at dmsolutions.ca    |
+-----------------------------------------------------------------+
|Chief Technology Officer                                         |
|DM Solutions Group Inc                http://www.dmsolutions.ca/ |
+-----------------------------------------------------------------+