@stropitek wdyt? Maybe we should remove the custom code and use transform with a scaling matrix?

I think it would make sense to use transform to have a consistent implementation of methods like resize, cropRectangle, rotate which all just iterate over a target image, calculate a position on the source image and then interpolate the value.

I would like to try to use transform matrix instead but I don't know how it work

https://en.wikipedia.org/wiki/Scaling_(geometry)

With a 2x3 matrix, it should be

[
  [ factor   0    0 ]
  [   0    factor 0 ]
]

It's the best I can do.
Nearest is working.
Bilinear is very close from the opencv result.

I let you review before resetting snapshots to let the tests pass.

Bilinear:

	PR	OpenCV	main
result
substract

Nearest:

	PR	OpenCV	main
result
substract

Do we still need custom implementation of the interpolation functions?

Bilinear is known to give different result compared to OpenCV. Solution is to add a margin of error with comment like here:

Use the smallest possible error margin

Do we still need custom implementation of the interpolation functions?

I removed It because I do not use it anymore.

Bilinear is known to give different result compared to OpenCV.

Ok I thought we should be one to one with OpenCV (And I found his less precise algorithm more smoother).

If you agree with theses code changes I think we can reset snapshots and merge

I found his less precise algorithm more smoother

Really? The difference is barely visible in the other tests (you have to quickly change from one image to the other to see that the colors change, but they are basically the same).

The difference is barely visible in the other tests

I did not check other tests, but take a lot of time to tweak trying to get result most similar result with opencv.
Comparing result with substraction, zoom-in and pixel grid ^^'

After snaptshots reset, some tests did not pass yet

There are still substantial changes to interpolateNeighbourBilinear which may explain it.

Yes, I'll adapt tests data

But are you sure about the change to the algorithm? It worked fine until now.

I'm not sure.
I found bilinear for the small image better and more similar to opencv. but for largeur image, black and white (like letters with rotation and matching) it seems more blurry. rotation seems to produce some color shifting with triangles.

Maybe I should rollback the bilinear but keep nearest.

This PR was in a dead end I preferred to restart from scratch and only fix nearest resize.
#454

This PR will be usefull for future as exploration of this topic.