Add ability to erode/dilate using structuring elements

[tlnagy]

In Matlab you can provide a structuring element to erode/dilate along: https://www.mathworks.com/help/images/ref/strel.html. This would be useful to have here too.

[timholy]

A foundation for efficient structuring elements can be found in the implementations of label_components. It's really old code but nevertheless may still be quite relevant.
This has come up in several places recently (e.g., #9), so it would be good to generalize that code. One option would be to consider an "inversion of control," and pass in a function f that then operates on the set of array elements that get passed. In some ways this is reminiscent of mapwindow in ImageFiltering, but not restricted to rectangular voxels.

For the foreseeable future my "community" time is going to be focused on bringing Julia a Matlab-like debugger (see ongoing work: JuliaDebug/ASTInterpreter2.jl#32, JuliaDebug/ASTInterpreter2.jl#37, JuliaDebug/ASTInterpreter2.jl#41). So it's best if someone else tackles it. Is there a taker? Or should we post this as a potential GSoC project? By summer I could imagine mentoring it, but right now I can't dive into this.

[martin-gutwin]

Don't know if it belongs here, but I wrote a quick generalized implementation of the _extremefilt! function, which can take a dist parameter, where dist=1 equals a 3x3 structuring element in 2D and dist=2 equals a 5x5 structuring element in 2D etc. I didn't write tests for dist>1 yet, but the existing tests for dist=1 all pass. I don't know if it helps, but if anyone is interested, here is my fork:

https://github.com/twinGu/ImageMorphology.jl

If there is interest for a pull request, I can also write up some tests and would be happy about some advice for improvements I could work in.

[timholy]

This seems as good a place as any to discuss structuring elements. Across the entirety of JuliaImages, we unfortunately seem to have a few conventions being used:

• dims = 1:n. In most places this implies "diamond" connectivity, but for erode/dilate it's equivalent to CartesianIndices (box/window)
• window = (3, 3, 3, ...). This is basically iterating over the CartesianIndices of a box around the center point.
• connectivity::AbstractArray{Bool}. Currently we don't insist that it be centered, and we interpret the center as the origin. When all values of connectivity are true, this is the same as window = size(connectivity)
• some languages support a dist integer or real number, typically a Manhattan or Euclidean distance

It would be nice to standardize these, or at least make them all mean the same thing throughout JuliaImages. Thoughts on how to unify them? connectivity is the most general of all of these but iterating over it is slow unless we convert to a Static variant; this has a runtime dispatch penalty but it should be pretty negligible for all but tiny images.

[johnnychen94]

This is supported by #65 and #80. The 0.4 release might be within a month if everything works smoothly.