Understanding Noise - Process Flow...

[MortenBalling]

I've been thinking about this.

Both ATWT and Deconvolution are sort of noise reduction and detail enhancement in one.

Yesterday I tried my new Atik 428 Mono for the first time ever (sweet camera btw!), and I'm currently fooling around with the data. I took my Light-integration, and tried to get the best possible result either Deconvolving or ATWT'ing it. The results are quite different, so afterwards I used a classic trick and made an average of the two. Problem solved. Sort of...

Morten

[jkmorse]

In light of the discussions above, I have been experimenting with the noise reduction in both linear and non-linear formats and have to say that, while I recognize that you can tackle noise in the linear stage, it seems to create all kinds of problems and artifacts (at least for me) that I do not see when saving noise reduction to the non-linear stage. 

As such, I am adjusting my workflow to save all noise reduction for post HistogramTransformation.

I would love to hear others' views on this.

Thanks,

Jim

[pfile]

what denoise algorithm are you using? for TGVDenoise a local support image is pretty critical when applying the process to a linear image. if you do autoSTF on your image and use the computed midtones and shadows values, that's a good support image.

rob

[jkmorse]

Rob,

I was using ATWT based on a tutorial I read from Juan but it really messes up my background when I go non-linear.  I get much much cleaner results by waiting until I get to the non-linear stage.

Jim

[MortenBalling]

Some parameters are kind of hysterical when you work with linear data. Especially Deringing. My experience has been that Deconvolution on linear OSC data was difficult, but it worked fine after HT. With linear mono data Deringing still needs very precise settings (dial in a low number instead of using the slider and test it on a preview), but the results are fine, once you find a proper setting.

The point here, I think, is that whatever method works on ones images is the way to go. There is probably a method that is theoretically correct for all images, but I haven't found it (yet).

Morten