Allow use of complex numbers with denoise_nl_means

[shirawerman]

The code of non_local_means.py converts the dataset to 'float', but I have a dataset with complex numbers. Is there a possibility to support more data types?

Thanks

[grlee77]

It would likely require substantial work to refactor the underlying Cython code to support complex types directly, but there are a couple of other options:

1.) Run non-local means independently on the real and imaginary components.

2.) If the data is 2D it might also be possible to stack the real and imaginary data along a new axis (e.g. np.stack((image.real, image.imag), axis=-1) and call it with multichannel=True. This would consider both components together when determining patch similarity and ensure that the same patch weights were used for both components. If your data is 3D, we would need to add support for 3D multichannel for this to work. but I do already have an implementation for that which I never officially submitted for review. Let me know if your data is 3D and I can see about reviving that.

[shirawerman]

Thank you for the quick reply!
I am working on 3D data so if you can let me know when that code is ready, that would be great!
For now I will work with the first solution, although is will generate different weights for the real and imaginary part.

[grlee77]

Hi @shirawerman, I opened a PR for 3D+multichannel (and 4D) support in #5488. If you want to compile scikit-image from that branch you can try it out.

We are hoping to make a 0.19 release soon, but I don't know that the PR will be reviewed in time for that.