[Image-SIG] python PIL 16-bit tiff files

[Fredrik Lundh]

This isn't really a full solution, but the following patch at least
allows PIL to read 3x16-bit RGB TIFF files, converting them on the fly
to 24-bit RGB.  Note that it requires new binaries to handle little
endian (intel) files:

http://hg.effbot.org/pil-2009-raclette/changeset/45c2debe0fc3

Unfortunately, your sample use "contiguous" pixel storage, which makes
the descriptor manipulation tricks I mentioned harder to implement
efficiently for PIL 1.1.7.  I played a bit with libtiff's tiffcp
utility to see if that could be used as a preprocessor (which is
otherwise a great way to handle more "obscure" TIFF flavours), but at
least the version I have doesn't handle 16-bit images well either.  I
need to think a bit more about this, I think...

</F>

On Thu, Apr 22, 2010 at 10:57 PM, Fredrik Lundh <fredrik at pythonware.com> wrote:
> Oh, missed that there was one in your first post.  I'm a bit busy
> right now, but I'll take a look when I find some spare time.
>
> </F>
>
> On Thu, Apr 22, 2010 at 10:56 PM, Fredrik Lundh <fredrik at pythonware.com> wrote:
>> On Tue, Apr 13, 2010 at 10:52 AM, Dan Blacker
>> <dan.blacker at googlemail.com> wrote:
>>> Hey guys,
>>>
>>> Thanks for your input,
>>>
>>> The image is only of a tiny cropped area of a long strip of color kodachrome
>>> film - I will send a better example with some more color in it when I get a
>>> chance.
>>>
>>> I was under the impression that PIL handled 16 bit images (experimentally)
>>> but does this only apply to 16-bit grayscale images?
>>>
>>> Am I going up a dead end trying to read my images with PIL?
>>
>> The current PIL release only supports 8 and 32-bit/pixel internal
>> storage; that's enough to hold e.g. RGB triplets or 32-bit signed
>> integers, but not 3x16 bit pixels.  I'd love to support more storage
>> formats (machines are a lot bigger now than when the internal,
>> intentionally very simple storage model was designed) including HDR
>> formats (float16 etc), but rearchitecting the internals without
>> breaking all existing code is a pretty big project...
>>
>> There is some limited support for 16-bit storage, by packing two
>> pixels per 32-bit storage unit, but not all operations support this
>> (it's mainly intended to support working with huge, memory mapped
>> single-layer images, such as satellite data).
>>
>> There are some non-standard tricks that may help you with your
>> specific case, though.  All codecs do things in two steps; the first
>> is to identify the file format and build a descriptor of where the
>> image data is in the file (the "tile" map).  The second step then
>> loads pixel data on demand, using that descriptor.  You might be able
>> to tweak the descriptor before loading the image, to load one layer at
>> a time.  Do you have any samples?
>>
>> </F>
>>
>