Proposal: add axes argument to ndimage API

Hi all,

Most functions in scipy.ndimage can be applied to n-dimensional images, but
applying the functions along a subset of axes is not as easy as it could
be. For separable filters there tends to already be a 1D variant where a
single axis can be specified, but the n-dimensional versions do not
currently have a corresponding axes argument.

I am proposing that we can expose an axes argument on functions to simplify
the process of filtering along a subset of axes (e.g. a common scenario
would be filtering along only the spatial axes of an array that may contain
other axes corresponding to color channels, timepoints or an image batch).
Users can technically apply filters to a subset of axes now by manipulating
the input arguments. For a few concrete examples of this:

For gaussian filter, set sigma=0 along any axis that should not be
filtered. For rank filters set size 1 along the non-filtered axes. For
convolution-based filters or morphology the weights array can have
singleton axes inserted.

None of these are particularly difficult to implement, but it may not
always be apparent to library users and an axes argument would be more
convenient. The proposed implementation would be purely at the Python API
level and is straightforward to implement and maintain.

I opened a proof of concept PR implementing only *gaussian_filter *here:
https://github.com/scipy/scipy/pull/18016

I am happy to extend this to the rest of the filtering module in a follow
up PR if the general approach looks good. It can also potentially be
extended to morphology or other submodules as well.

Let me know what you think

Cheers,
Greg

_______________________________________________
SciPy-Dev mailing list -- [email protected]
To unsubscribe send an email to [email protected]
https://mail.python.org/mailman3/lists/scipy-dev.python.org/
Member address: [email protected]