Super-fast Kuwahara image filter (for n-dimensional real or complex data)

This is a well-known filter that preserves edges by 'analyzing' the statistical variation in the environment.
The filtering procedure involves two convolutions: if the chosen kernel is relatively large, the convolutions are carried out by a multiplication in the Fourier domain.
If the input data is scalar, the convolutions are carried out in a single convolution, by processing the second convolution in the imaginary channel of the first convolution.
Compared to Luca Balbi's implementation, it is twice as fast for the smallest kernel, and up to five or six times as fast for larger kernels.
https://www.mathworks.com/matlabcentral/fileexchange/15027-faster-kuwahara-filter
The demo files compares this method on speed and accuracy with two other methods published on file exchange (on a 2D image).