@alexandrejaguar a bit of rearrangement shows that the resulting threshold image should be integral_image * (1 - q), where q is a quantile:

result = (image - integral_image * (1 - q) > 0)
result = (image > integral_image * (1 - q))

The lower form is the format we want for applying a threshold, ie normally we would say image > t for some threshold t.

However, rearranging also shows that this is equivalent to niblack thresholding with k=0, so I'm not sure this addition needs to happen. Maybe a note on that function is sufficient. Otherwise, the implementation should just call that function, since it is n-dimensional rather than 2D only.

@scikit-image/core any thoughts on this?

Note also that the implementation above has a mistake: the integral image needs to be divided by the window size, otherwise we are looking at the sum of the surrounding window, not the mean as the docstring suggests.

@jni, @stefanv and I were discussing if this should be implemented. Based on your first comment, maybe it shouldn't. +1 on the note on niblack thresholding.

Better of we don't have to. Perhaps add a line to niblack stating the equivalence?

Sorry, I spoke too quickly: the q/bright_perc muddles things up. It's actually equivalent to Niblack times (1-q).

I don't see a problem on putting that on niblack's notes, and maybe an example on the docs. Is it okay?

yep!