Other possible options for handling 16 bit images:

1. just divide by 256 before casting to uint8, possibly loosing dynamic range (10 and 12 bit images are common).
2. allow pil_to_array() to return uint16 arrays. The docstring says that uint8 is returned.

I like the idea of pil_to_array returning a 16 bit array when the input image is 16-bit. It allows the caller to deal with downsampling to 8-bit in the most appropriate way. We can update the docstring to indicate that the dtype of the output depends on the input file or some such. I think it's reasonable to make small backward-incompatible changes like this on master (just add to the changelog). While we're at it, can we add a test for this? Showing that a 16-bit input image is correctly returned as a 16-bit array?

I have one small note about testing. This doesn't (I believe) change the fact that PNG images read through the _png extension (i.e. not through PIL) are still converted to 8-bit. We should fix that as well.

This PR looks useful but discussion seems to have died down. I'd like to get this in the bugfix release. @cgohlke , any chance you could change the src to v1.1.x rather than master. Do you have any feedback from @mdboom comments?

I don't think all of these changes can go into v1.1.x. Not flipping PIL images or allowing 16 bit images might break existing applications. The _png extension should be fixed in another PR. It is currently broken in master (test_png .py fails).

I agree this shouldn't be on 1.1.x, but it should get merged into master.

I've added an issue #915 to track adding 16-bit support to the _png extension.

As far as I can tell, the test_png.py failure is currently obsolete -- it is passing on master now.

I think the following would make this mergeable:

1. Add a few asserts to the unit test to ensure the data is read correctly. I think assert that "sum" of the pixels is the correct value would be adequate to ensure there were no endianness problems, etc.

2. Rebase this on current master

3. Make sure the regression tests still pass.

@cgohlke -- it looks like your fork of matplotlib has been removed, making it impossible for me to review this pull request. Any idea what might have happened?

I was able to get the diff manually from here:

https://github.com/matplotlib/matplotlib/pull/616.diff

so all is not lost -- I should be able to verify and merge this by hand.

Merged into master

Thanks. My matplotlib fork got corrupted.