this module supports inpainting of clipped channels of raw images. it works only for raw images, and on raw data. a kind of wavelet pyramid is used to obtain the inpainted colour, trying to keep colour ratios smoothly varying.
whitethis lets you manually overwrite the clipping point. ideally
everything would be normalised to clip at 1.0.desatcontrols the amount of desaturation of highlights. it is a parameter because there are cases for both extremes. some images come with chromatic aberration artifacts near clipped highlights and would thus only blur false colour magenta into the clipped region. you want to desaturate this completely. however, things such as saturated skin colours should be inpainted with the original skin tone, i.e. not be desaturated at all.softcontrols the amount of softness of the inpainting. essentially this weighs different scales of the wavelet pyramid. no softness means the highest frequencies get a good share of the result. all soft means pretty much only the coarsest scale, i.e close to constant colour reconstruction. again, when inpainting chromatic aberration artifacts you may choose to go soft, but when reconstructing a smooth gradient of a sunset you want a specific kind of gradient.
reconstruction runs in two steps: blur and combine.
the blur is clipping aware, i.e. it disregards clipped pixels and renormalises results to include only unclipped pixel values.
recombination then takes the original image and the blurred result as input images. it rescales the blurred version in brightness to match any potentially available unclipped pixels.
gauss pyramid:
- downsize half size, disregard clipped pixels (all clipped -> output clipped)
- blur similar to llap, but disregard clipped pixels
recombine:
- upsample low res version
- compare to hi res: rescale low res to match hi res unclipped channels
- fill in
this is done by reduce.comp and assemble.comp. the last step, going from xtrans/bayer to rgb image
is done in a halfsize scaling kernel half.comp/doub.comp