Allow arbitrary extrapolation methods

[bergfried]

Whenever a function needs to extrapolate values of non-existing pixels (i.e. pixels outside of the image), it should be possible to specify the extrapolation method to use.

The extrapolation methods I consider to be most useful are:

• replicate border
• replicate border, with alpha set to 0
• arbitrary constant
• wrap around
• arbitrary Python function

See the introduction to http://docs.opencv.org/modules/imgproc/doc/filtering.html for further explanation and other extrapolation methods. There are several Pillow functions where this comes in handy. For example,

image.transform(size, PIL.Image.AFFINE, data, PIL.Image.BICUBIC, extrapolate=REPLICATE_BORDER_ALPHA0)

might produce a rotated and scaled version of the source image on top of a transparent background so you can paste this transformed image into other images easily. Another example:

image.filter(filter, extrapolate=WRAP_AROUND)

might produce a blurred image that you can use for a tiled wallpaper without ugly tile borders.

If you need to prioritize, I consider Image.transform to be more important than Image.filter. (This is because it is quite easy to work around this issue for Image.filter as soon as this is implemented for Image.transform.)

[nathancahill]

Does this have to do with the fill parameter? I'm digging into the _transform2 function. It seems that fill is passed to the function, but passed as 1 to ImagingTransformAffine, ImagingTransformPerspective and ImagingTransformQuad.

I lose track of fill in the generic transform function ImagingTransform, I don't know enough C to know what the line if (fill) memset(out, 0, imOut->pixelsize) is doing here. Besides that line, I might be able to submit a pull request to at least support transparent fill, if not extrapolate.

Edit: Reading more about it, it looks like memset(out, 0, imOut->pixelsize) writings a pixel's worth of zeros (black?)

[radarhere]

I might be able to submit a pull request to at least support transparent fill, if not extrapolate.

Transparent fill should already be the behaviour.

from PIL import Image, ImageTransform
im = Image.open("Tests/images/hopper.png").convert("RGBA")
transform = ImageTransform.ExtentTransform((0, 0, 150, 150))
im.transform((100, 100), transform).show()

gives this image (it's hard to see because it is transparent, but there is transparency to the right and bottom of the image).

[Yay295]

Of course, it's only transparent if the image has an alpha channel. Otherwise it's black.

Also there are two different parameters to some of the methods: fill and fillcolor. Some of the methods take one, or the other, or both. So maybe that could be something to make more consistent.