Example for a 2D Median Filter

[RoyiAvital]

I find StaticKernels.jl to be a jewel for low level image processing in Julia.
It is odd, as Julia is very good choice for image processing yet the landscape of packages is rather small.
Unless you go the JuliaImages.jl path, which I don't like due to the choice Arrays of Struct as the underlying type and large dependency.
I prefer working with Matrix{T} and Array{T, 3} for images.

This is how I implemented Median Filter with StaticKernels.jl:

mI = rand(10, 10); #<! Image
mK = Kernel{(-2:2, -2:2)}(@inline w -> median(Tuple(w))); #<! The kernel (Window function)
mF = map(mK, mI); #<! Filtered image

The Tuple(w) is required as in #8 (Iterating the window).

There are optimized algorithms for 2D Median, yet I couldn't find implementation for simple arrays in Julia.

[stev47]

I think you meant mF = map(mK, MI) (kernel first) in your last line.

For some kernels like median there are more clever algorithms which scale better with window size such as the Herk-Gil-Werman algorithm. Keep in mind that StaticKernels will not be able to do that.

The MR implementing median for StaticArrays is still open, but you can hand-roll your own median e.g. like this

@inline function mymedian(w::StaticKernels.Window)
  v = SVector(Tuple(w))
  v = @inline sort(v; alg = StaticArrays.BitonicSort)
  isodd(length(v)) ? v[end ÷ 2 + 1] : ((v[end ÷ 2] + v[end ÷ 2 + 1]) / 2)
end
k = Kernel{(-2:2,-2:2)}(mymedian)

which should be slightly faster than what you did.

[RoyiAvital]

Yes, I am aware of the more efficient way for rolling medians (I wrote that above).
But in the current eco system it means one must implement by himself.

Your package gives me the ability to do things quick.
So it is better to have something than potentially better :-).

So this is just a post for appreciation and others, in case of need, will be able to use.