TGVDenoise Local Support for NonLinear Images

[tstephens]

I have seen a couple of posts on this topic for Linear images using settings that are taken from STF, however, once you are in the NonLinear mode there seems to be nothing on this. Does anyone have a logical approach to setting Local Support parameters for TGVDenoise for NonLinear Images? Is there any reason why TGVDenoise is better applied to Linear versus NonLinear images?

[pfile]

i have not bothered with a support image for nonlinear images. as far as i can tell it's not strictly necessary. without the support image TGVDenoise really destroys linear images, in my experience.

since the point of the controls of the support image is to stretch the linear image, i suppose for a nonlinear image all you have to do is just turn on the support image without messing with the transfer functions. this is because the nonlinear image is, after all, already stretched.

rob

[Carlos Milovic]

Well, the reasoning behind the local support is to tell the algorithm where the SNR is higher, so more of the original image is keeped. You may think of this as a mask, but this works inside the algorithm, in each iteration, in a more elaborate way.
The algorithm in TGVDenoise in it's nature assumes that the image contains homogeneous gaussian noise, so this local support is one way to "fool" the algorithm, and take into account that in reality we have Poisson noise, with other components.

 

[tstephens]

OK. I am getting two messages: 1) don't mess with local settings with a nonlinear message, and 2) adjust the settings to fool the algorithm into using poison noise model. So again, what is the best way to adjust the local support settings in TGVDenoise for a nonlinear image?

[niteman1946]

For non-linear image, I use default settings for TGVDenoise, no Local Support, and I usually reduce Strength from 5 to 2.

This works for me.  It really is a useful tool.

Good luck,

Mark

[tstephens]

Yes, I agree a very useful tool. Even when the settings are not so obvious.

[Carlos Milovic]

I think that the answer is: "it depends". In many cases, with non-linear images you won't need to use the local support, specially if the noise amplitude seems to be quite the same across the image. If not, then a simple mask to protect the zones with higher SNR can be enough. I would say that the use of the local support in those images is intented only as an advanced option, pursuing for the best result.

I'm sorry that the settings are not obvious (specially those in the local support section). Please understand that this is a state-of-the-art algorithm,  and we are discovering it almost at the same time that you. We tried to give the best default settings, and the most meaningful names from our little experience with it. It is one thing to understand the mathematics behind it, and implement it, and another to put a name and intuitive behavior to all the parameters. We'll definitively be using your feedback to design a better tool in the future. Right now I'm coding again the new tgv based deconvolution tool, and later will begin refining the tgv toolset. So, thank you for your feedback and suggestions, and also for your questions. This shows us where to put our emphasis in the upcoming documentation.

[tstephens]

I love the tool and have already learned that it is more effective than ACDNR. I have found local support without modification of the settings produces a luminance map like the image, and to be expected with a nonlinear image. Even this changes the outcome of the TGVDenoise procedure. Not sure why but in the images I have been working with it gives a better result. For now, this will work for me.