refactor: update image resizing implementation to use OpenCV and add …

[r-bit-rry]

closing this issue:
#253

removing scipy dependency
chaging resize_image from scipy.ndimage.zoom to opencv resize
adding unittest to avoid regression

[maxlund]

This might be worth taking a look at
https://zuru.tech/blog/the-dangers-behind-image-resizing

[Blaizzy]

@r-bit-rry could you share some benchmark results with the old and new method?

[Blaizzy]

Also, why can't we use PIL for this?

[maxlund]

Also, why can't we use PIL for this?

From the link I posted above, PIL seems to be the best choice

[r-bit-rry]

I guess it doesn't really matter.
I chose opencv for the speed (at least the documented one)
I don't even know if the scipy method zoom that was used before was equivalent to the quality by PIL

One last note, the article is from 4 years ago. it might be worth retesting what he's done, or just take PIL whatever takes less time

[r-bit-rry]

@r-bit-rry could you share some benchmark results with the old and new method?

checked for regression and added a test, not for benchmark.
benchmarking can be seen here but obviously it was not made for this specific usecase:
https://github.com/libvips/libvips/wiki/Speed-and-memory-use

generally speaking I'm not sure that now is the time to optimize performance, especially when the bottleneck is most likely the mlx model itself.

looking over the updated benchmark it looks like PIL is the proper choice if using latest versions

[Blaizzy]

Awesome!

Lets migrate to PIL then, it's already widely used in the project.

Could you update the PR @r-bit-rry?

[Blaizzy]

Hey @r-bit-rry

I went ahead and tested this and it works, thank you very much!

However, there are more places where the change was needed and after testing different resampling techniques, I found that LANCZOS4 performs best overall with CV2 compared to CUBIC, INTER_AREA and others. It captures more details on images. #301

And for PIL bilinear was the the best.

However, they don't numerically match scipy.ndimage.zoom outputs.

PIL (Bilinear) Mean pixel difference between methods: 2.64
CV Mean pixel difference between methods: 49.28
Allclose with rtol and atol (1e-05): False

Here is an open issue that we can work towards closing that gap: #302