I think you uncovered a very useful alternative workflow. Initial impression
was that at high ISO Denoise AI was best, but at base ISO sharpen AI which can
also do a light denoising. Now that is thrown on its head.
The overlapping tools need someone like you to suss out the optimal uses.
<<<Does the app make it up?
Yes, kinda sorta within the parameters of the pixel of interest and surrounding
pixels.
<<<how is it so accurate seeming?
---Well trained neural network with training set adequate for the task at hand.
<<<Does it do what deconvolution does, only from a completely new direction?
Not very similar at all. Deconvolution is all math based. The detail
recovery is "real" based on the assumed convoluting PSF (point spread
fxn)--this is from the real full wave solution of the lens/diffraction effects
in Canyon's DLO and
a Gaussian blur in FM, as best one can tell. In Gigapixel AI the new pixels
are manufactured out of thin air--can be spectacularly good if the training set
is adequate and massively beats interpolation methods in many instances.
I had a good working knowledge of backpropagation neural networks that I used
back in ISA board days--had an accelerator board just for this. I haven't had
a need to keep up with this but they are more powerful now of course.
The exact path for the NN to converge on a solution for a given pixel will
remain a black box for the foreseeable future--they just work.
Soon to be replaced by a NN augmented heuristic expert system? Mike
--
_________________________________________________________________
Options: http://lists.thomasclausen.net/mailman/listinfo/olympus
Archives: http://lists.thomasclausen.net/mailman/private/olympus/
Themed Olympus Photo Exhibition: http://www.tope.nl/
|