Summary

We currently use a relatively complicated image resampling algorithm because we want to preserve data values above vmin and vmax but theagg resample routines only allow values between 0 and 1. Agg also does not preserve NaN, so we need to remask at the end. This makesimage._make_image quite long and convoluted.

Proposed fix

Pillow allows 32-bit one-channel images, with no apparent restrictions on the values, and has aresize method that seems to do resampling as we would like.

The only issue that I can see is that Pillow has fewerresampling kernels than we do

• nearest
• box
• bilinear
• bicubic
• hamming
• lanczos

Whereas we allow quite a few more currently (inherited fromagg):

methods = [None, 'none', 'nearest', 'bilinear', 'bicubic', 'spline16',           'spline36', 'hanning', 'hamming', 'hermite', 'kaiser', 'quadric',           'catrom', 'gaussian', 'bessel', 'mitchell', 'sinc', 'lanczos']

I frankly think very few people are using the more exotic filters, or need to. If we want to change to pillow we could

• keep the old resampling around for the other filters (perhaps deprecating them)
• add the other filters to Pillow

The patch for the giant resampling code is:

+                im = Image.fromarray(A)+                output = np.asarray(im.resize(out_shape, Image.BICUBIC))+                output = self.norm(output)+                out_alpha = np.ones_like(output)