Similarities between ImageInpainting.jl and CloudClean.jl about cloudclean.jl HOT 5 OPEN

Hi, thanks for brining ImageInpainting.jl to my attention. The method there is very cool, but also very different from this algorithm. I would be happy to consider merging this there if there were different algorithm options and the different applicabilities were well clarified.

from cloudclean.jl.

Thanks for your interest @andrew-saydjari , let me see what's needed to make things easier to incorporate newer algorithms in terms of API in ImageInpainting.jl. By the way, are you on slack?

from cloudclean.jl.

Yes. Happy to join the slack discussion if that works better. [email protected]

from cloudclean.jl.

In the example in ImageInPainting.jl,

using ImageInpainting

# inpaint image within mask using algorithm
inpaint(img, mask, algo)

img: This is always likely to be there
mask: Again always supposed to be there I think to define the ROI
algo: This for your case could be CloudClean which is of InpaintAlgo Type and its parameters can also be specified in kwargs as needed.

If you see: https://github.com/JuliaImages/ImageInpainting.jl/blob/master/src/criminisi.jl

struct Criminisi <: InpaintAlgo
  px::Int
  py::Int
end

for your case, it will be something like:

struct CloudClean <: InpaintAlgo
  parameters1::Int
  parameters2::Int
  parameters3::Int
end

from cloudclean.jl.

So I think that would be straightforward for proc_continuous, but for our proc_discrete mode, we pass a list of ROI centers, not just the mask. This improves the infill because you are choosing your "training" data to be optimally centered around the regions that you want to infill. I am not seeing an obvious way to fit that into what you describe. It would be hard for me to think about that list of ROI centers as parameters of the algorithm, but maybe an optional keyword to inpaint that (via multiple dispatch) is only allowed for CloudClean.

from cloudclean.jl.