A bit of history and the bilateral filter

[Andreas Mueller]

Hi Emmanuelle.
Thanks for the info about the Patent. I didn't know about that.
I was also thinking about the Ponce algorithm for denoising,
but I thought the non-local means was easier. Have you used
the dictionarly learning yourself?
I heard a talk by Ponce about it and it looked pretty cool.
I haven't had time to really look at the dictionary learning
in sklearn yet but it looks as if this is nearly what we want.
As far as I can tell, one "only" needs to add the patch-similarity
term in the optimization. I don't know how straight
forward that is.
But this would definitely be a cool feature to have.

What do you think about the lena denoising example
in sklearn?
 From what I saw in papers, I would expect the
denoising to work better, even without the
improvements from the Ponce paper.
Do you think the result would be better other
optimization strategies? Or do you think the
dictionary is to small?
It seems the patches where only trained
on the Lena image. In the talk, Ponce
said it would be good to train on many images
as an initialization and then train again only
on the image on wants to denoise.
Do you have any experience with that?

Cheers,
Andy