Switch to improved vips_resize, requires libvips v8.3.1+

[jcupitt]

Hi all, I've been looking at vips_resize(), the vips function that sharp uses for downsizing, and I think I've maybe found a couple of things which can improve quality, especially for small downsizes. Would anyone be able to help test the changes?

Code changes in this commit:

jcupitt/libvips@8506ff1

There's a test image here, plus sample shrinks from vipsthumbnail and convert. Put the images in separate tabs and try flipping between them.

http://www.rollthepotato.net/~john/skin/

The test image is an interesting one since it combines fine, high-contrast detail in the hair, plus a strong skin texture (I hope the subject doesn't mind me saying that). The downsize needs to minimise aliasing in the hair while retaining texture in the skin.

The old vips_resize() avoided anti-aliasing nicely, but had the effect of almost airbrushing the skin. The three changes in the patch:

• reduce the strength of the anti-alias filter slightly
• does not do the final sharpening pass if we skipped the anti-alias filter
• fixes a bug where the sharpening would be skipped if there was no box filter

Any thoughts? Would someone be able to compare to photoshop? I remember we had an interesting discussion the last time we went over this code.

[lovell]

Hi John, sharp doesn't currently use vips_resize 😱 although I'd very much like it to (e.g. #299).

One thing that prevents sharp going all-in is the need to support resizing without maintaining aspect ratio. (which _shrink and _affine both support). This could be as simple as something like sharp's approach of averaging the x and y factors - see https://github.com/lovell/sharp/blob/master/src/pipeline.cc#L536

[jcupitt]

You could check your resize code against these changes, I guess. I think you are using a pretty similar algorithm, is that right?

[jcupitt]

I'd forgotten about that request, sorry. I added non-square resizes to vips_resize():

jcupitt/libvips@2620f3b

It'll be poorer quality for non-square, but perhaps that doesn't matter.

[jcupitt]

I had a look through pipeline.cc and it looks like these changes aren't going to be relevant to sharp, you've already turned down the anti-alias filter. I'll close.

[lovell]

I feel this is still relevant to sharp. Your changes to vips_resize make it behave much more like the current (guesstimated) level of anti-aliasing.

If it's OK with you I'd like to reopen and make this the task to migrate sharp to using vips_resize.

[jcupitt]

Sure, reopen away. It's at least a useful test image.

Here's the previous resize quality issue: #121

It had these two other useful test images: wet hair, featuring a lot of high-contrast edges and some skin texture, though less than the one above, plus ghibli, which has tricky gently-angled lines.

http://www.rollthepotato.net/~john/wet-hair/
http://www.rollthepotato.net/~john/ghibli/

I've added new images for these new settings. wet-hair looks about the same, but the ghibli is a bit better.

[lovell]

Thanks John.

On a related note, I've been experimenting with the removal of the anti-alias filter for the linear(-ish) interpolators - see https://github.com/lovell/sharp/blob/look/src/pipeline.cc#L549 - sample outputs at https://github.com/lovell/sharp/tree/look/test/interpolators

This is due to the proposed switch, in the next release of sharp, to a default interpolator setting of bicubic (from bilinear). Avoiding the convolution step with bilinear resizing is ~20% faster and still produces fairly good results (esp. if we assume that using bilinear is itself "good enough").

[jcupitt]

Yes, vipsthumbnail has been bicubic for a while. I think you still need the anti-alias filter and the sharpen, even with bicubic.