It's ridiculous in places and still a bit slow, but I'm quite happy with the idea underneath it.

It maps the number of all finite (not NaNs, not infs, not having indices beyond the array edges) neighbours of each pixel in the given array. Then it loops through the masked pixels which have the _largest_ number of adjacent neighbours available, picks the closest neighbours in a 5*5 matrix around each pixel, fills the masked pixel with distance-weighted Gaussian average of said neighbours, then updates the neighbour count, loops through the pixels which have the largest number (out of 4) of neighbours again, (...).

I realised I should do inpainting this way while riding my bicycle from cbase. My previous implementation simply looped over indices along one axis, disregarding the number of pixels used for filling the given one. Hence ugly flux shape deformations when a masked region is directly above or below the galaxy in question.

But now it seems to be fine! So finally I get to do some science, instead of coding 24 if's and exceptions in one function.

It maps the number of all finite (not NaNs, not infs, not having indices beyond the array edges) neighbours of each pixel in the given array. Then it loops through the masked pixels which have the _largest_ number of adjacent neighbours available, picks the closest neighbours in a 5*5 matrix around each pixel, fills the masked pixel with distance-weighted Gaussian average of said neighbours, then updates the neighbour count, loops through the pixels which have the largest number (out of 4) of neighbours again, (...).

I realised I should do inpainting this way while riding my bicycle from cbase. My previous implementation simply looped over indices along one axis, disregarding the number of pixels used for filling the given one. Hence ugly flux shape deformations when a masked region is directly above or below the galaxy in question.

But now it seems to be fine! So finally I get to do some science, instead of coding 24 if's and exceptions in one function.

## No comments:

## Post a Comment