Edge detection about scikit-image HOT 7 CLOSED

The sobel and prewitt stencils look at the 1st and 2nd nearest neighbors. So a "bad" pixel would affect the gradient results of those neighbors. It may not be the best solution, but it seems these filters extend the mask to cover up this spreading.

That said, all the masking calculations can be moved into the if-block so that they don't affect the non-masked case. I'll take a crack at this and add some tests to actually check the masking behavior described above.

from scikit-image.

I would personally remove this functionality completely as it is a bit too magic and its behaviour is not described in the doc string.

from scikit-image.

The point of mask erosion is for when measuring edges/textures in a labeled
image, where one wants to ignore any values that aren't within the
foreground region. That's its original intent from the CellProfiler source.

I would fix the docstring, rather than remove the functionality.

from scikit-image.

I do not mean to remove the mask parameter, but the binary_erosion part. I find this quite confusing.

from scikit-image.

I think the erosion is probably necessary for some applications,
particularly if foreground and background intensities are widely separated.
Otherwise, foreground-background edges are likely to contaminate edge-based
texture calculations.

from scikit-image.

Yes, that makes sense. But we should definitely mention this in the doc string.

from scikit-image.

Docstring, tests, and refactor for these edge filters in PR #298.

from scikit-image.