Replace resize method

[homm]

Ok, that is why we all there.

Affine-based im.resize implementation should be replaced with convolution-based im.stretch for Bilinear and Bicubic filters (Lanczos was always there). im.stretch uses all pixels from source image for downscale (actually it uses each source pixel in several destination pixels depend of filter), while im.resize uses only fixed amount of pixels for each destination pixels (2x2 for bilinear and 4x4 for bicubic).

Examples:

This also fixes #986, #964, and possible other related bugs. im.stretch also faster for upscaling (especially on Bicubic).

As im.resize and im.stretch is part of python API (although not documented), we can't just remove one of them. Also im.stretch has own Nearest implementation, which is Nearest for upscaling and supersampling for downscaling. Therefore I suggest to leave both functions, replace im.resize implementation to convolution-based for all filter except Nearest and use ImagingTransformAffine for Nearest.

[homm]

Oh, I don't even run the tests. Let's discuss first.

[hugovk]

I recommend enabling Travis (and perhaps Coveralls) on your own repo.

See also https://github.com/python-pillow/Pillow/blob/master/CONTRIBUTING.md

[homm]

@wiredfool Any thoughts?

[wiredfool]

I'm thinking about the interfaces.

We have very similar imaging.c:_resize and imaging.c:_stretch functions. At this point, I don't think that there should be a distinction between them. If there is a distinction, it's lost on me.

Nothing in Pillow calls imaging.c:_stretch anymore. We should drop that function. If we wish to retain API compatibility, we should point _imaging.c:3030 {"stretch", (PyCFunction)_stretch, 1}, at _resize instead. Alternately, we could make _stretch just wrap _resize and issue a deprecation warning. I consider _imaging.c to be essentially a private api, so I'd be willing to just drop it completely.

I think that the lower level calls in Antialias.c should be called ImagingResize*, instead of ImagingStretch*. That way, the calls all the way down the stack are resize, and there's no trace of stretch anywhere in the code.

Other than that, it looks good to me.

[homm]

Distinction in NEAREST method only. _stretch produces supersampling for downscaling instead of real NEAREST. This should be a bit faster than BILINEAR (that is why someone can use it), but much slower than separate supersampling implementation can be. I suggest to leave two versions in 2.7 and remove _stretch in 3.0. What is the best way to produce a warning from C code?

[homm]

@wiredfool I've fixed test a bit and make stretch point to _resize.

What you think about renaming ImagingStretch to ImagingResample and Antialias.c to Resample.c? Because resample is exactly what that happening. In opposite, resize is a method who can choose what to do: make a resample, or fast affine transforms.

[wiredfool]

Resample sounds like a good rename. Will look over the code after dinner.

[wiredfool]

Ok, Good to 41029f0.

[homm]

renamed

[coveralls]

Coverage increased (+0.03%) when pulling bc0f896 on homm:replace-resize into e0b94d6 on python-pillow:master.

[wiredfool]

Looks good to me to bc0f896. Anything else to do?

[homm]

@wiredfool No, I suppose.