I need an operation I’d call “soft dilation”.

If standard OpenCV dilation, using a binary kernel, works something like this:

dst(x, y) = max src(x + i, y + j) for i, j if kernel[i, j]

Then “soft dilation” would work like so:

dst(x, y) = max src(x + i, y + j) * kernel(i, j)

So, it would be like using a 2D filter, but with a max aggregator instead of sum.

I’d need the solution to work with large (e.g. 40×40) Gaussian kernels.

Looking through OpenCV API I don’t see anything supporting this. Dilation treats every kernel as binary. I don’t see other filters that use min/max aggregations…

I might be able to implement this using Cython and numpy, but I was hoping to find something already made, tested and optimized.