1.7.402 Deconvolution > error diffusion

[alacant]

The help ? has notes for 'REGULARIZATION', but with a % slider rather than a number.
@ Ivo, could you tell us what it is?

Cheers and clear skies folks,
Steve

[admin]

Hi Steve,

Documentation/help is not up-to-date in alpha (development) versions. Error diffusion (name tentative) is a re-working of the Regularization parameter in 1.6 to be more intuitive and useful.

Hope this helps!

[hixx]

Hi Ivo,
For me, 'Error Diffusion' parameter introduces coarse grain on any other value than 0. The result looks like pushed TRI-X-PAN (if anyone still knows that look) . It seems to do exactly the opposite of diffusion. I hoped this to be rectified but recent releases up to 1.7417 show no change in behaviour. Am I doing something wrong? Do other users see the same?
cheers,
jochen

[admin]

Hi Ivo,
For me, 'Error Diffusion' parameter introduces coarse grain on any other value than 0. The result looks like pushed TRI-X-PAN (if anyone still knows that look) . It seems to do exactly the opposite of diffusion. I hoped this to be rectified but recent releases up to 1.7417 show no change in behaviour. Am I doing something wrong? Do other users see the same?
cheers,
jochen

Hi Jochen,

This is precisely expected behaviour!

Traditionally, deconvolution (of the iterative Richardson & Lucy variety) is a very difficult process to manage. The biggest problem has always been noise grain and artifact creation by the algorithm in response to noise in the dataset in a sort of runaway process. A regularization step is supposed to deal with this. As far as I know, StarTools' implementation is unique in that it will keep this runaway process in check 100% of the time and no noise grain will be introduced. That is to say, StarTools' implementation is convergent to an "optimal" solution (where "optimal" is the guaranteed suppression of visual noise grain). As a result of being convergent, you will/should notice that increase the amount of iterations will start to yield less and less improvement to the point of being unnoticable.

However, this behaviour can be overridden and the Decon module can allow for noise grain to be introduced like most other implementations I have seen. This, however may also improve detail further under some circumstances but at the expense of noise grain and artifact development.

Even in this scenario, StarTools attempts to be "clever" about this. The "Error" is the amount of grain StarTools thinks it will introduce, while the diffusion part pertains to how StarTools will attempt to scatter the grain in a way that is less obvious (similar to the way Denoise 2's noise grain equalisation allows noise to remain). The grain introduced in this manner can also still be effectively addressed by the Denoise modules.

I hope this helps and makes sense!

[hixx]

Thanks Ivo,
got it!
But I just can`t help - it looks very harsh on my dataset upon the slightest move of the fader. Other than Denoise 2.0 which creates a beautiful fine grain look, the 'error diffusion' pattern has very coarse blobs with an ugly "tech" structure to it. Moving the fader just a tad away from 0 will completely "destroy" any low contrast areas like nebulosity. So it looks like either the fader is working to extreme or the "diffusion" part is not working properly for me. The regularization parameter of 1.6 yielded much better results for me. I`ll produce some example data points for you to look at.
regards,
jochen

[admin]

Thanks Ivo,
got it!
But I just can`t help - it looks very harsh on my dataset upon the slightest move of the fader. Other than Denoise 2.0 which creates a beautiful fine grain look, the 'error diffusion' pattern has very coarse blobs with an ugly "tech" structure to it. Moving the fader just a tad away from 0 will completely "destroy" any low contrast areas like nebulosity. So it looks like either the fader is working to extreme or the "diffusion" part is not working properly for me. The regularization parameter of 1.6 yielded much better results for me. I`ll produce some example data points for you to look at.
regards,
jochen

Is this with the 1.7.417 version? The response of the slider has changed to act over a wider range.

Before 1.7.417, it acted as the inverse of the regularization parameter in 1.6 (e.g. if you think of a value of 1.0 in 1.6 as 100%).

[hixx]

Hi Ivo,
I have retested this on 1.7.421 now and the fader control shows much improved : the strange artifacts will not be produced using values below 50%.
Instead, there is a very slight but obvious increase in "detail" or "fine grain" which is the intended look. The fader curve seems just right now, even a slight dose of artifacts (above 50%) might be useful in some scenarios, because the Denoise process will eat some of this grain anyway.
regards,
jochen

[admin]

Excellent!
It's the little things. Some parameters elicit a logarithmic response (common with noise related things) or exponential response, so usually the finishing touch is to make the parameters behave linearly (much more predictable) rather than logarithmic or exponential, by taking the inverse.

[AndyBooth]

Just tried the latest decon routine on my latest mars with c11.
I think it is very good at suppressing the orange rind effect as much as it does,
Still there, but far Far less than other brands of sotware!