DBE - How to verify the model

[Stephane Murphy]

Hi, reading several good posts I now understand why we want to seperate processing by linear/non-linear steps. During my pre-processing step, I always use flats (sky or light box), darks and bias and create a final master RGB file.

My next steps are in the linear space and usally STF and DBE, Background neutrazlization and color correction. After depending of the type of image I may apply a deconvolution.

My question is how do I make sure that the DBE model generated is correct and subtracting the model will keep my data in the linear space? Any quantive ways the verify that?


Sorry for the long introduction.

Thanks for the help,
Stephane

[Juan Conejero]

Hi Stephane,

how do I make sure that the DBE model generated is correct

Excellent question. These things should not be taken for granted without verification.

After subtracting a good background model, the image should be uniformly illuminated (what we often call a flat image informally). To evaluate uniformity, you can use different strategies. Personally, I think that a careful visual inspection is the best practical approach. I often apply a wild stretch and color saturation boost to a temporary duplicate of the corrected image. A strong nonlinear stretch immediately shows which areas of the image are poorly corrected, either overcorrected, which causes too dark areas, or undercorrected, which means that the light pollution gradients have not been properly removed.

But one can be more exhaustive, of course. A more objective way to test for illumination uniformity is to define a different DBE instance, that is an instance with different samples over background regions, and generate a second background model. The generated model should be essentially a constant function (an image filled with a constant value). This method applies a counter-test criterion to validate a background model.

Yet another way --100% qualitative but very sensitive-- is to generate a 3D profile of the image. You can use the excellent 3DPlot JavaScript script by Andrés Pozo and David Serrano. The script cannot be used with a large image unless you have a power machine, but you can downsample the image strongly before applying the script, because the background illumination profile is an extremely smooth function.

Hope this answers your question. If not, we can feed our brains to find more and better ways

[Stephane Murphy]

First thanks for your detailed answers. This give me some material to think of. I think visual inspection can work for me but it take trained eyes to know if the data is still linear. I am not sure to understand yet why a 3D profile is needed but I'll try the suggested script. Why not just a 2D intensity profile of the image?

I am thinking maybe a function to create a before and after intensity profile and plot the two profile image by superimposing them.

Not sure if I am thinking this correctly.

Thanks
Stephane


 

[Stephane Murphy]

Hi Juan, I am looking at the 3DPlot JavaScript script by Andrés Pozo and David Serrano. I now understandwhy why you suggested this solution, very nice work.

I am thinking to create a new version that will superimpose two 3d profile usinf different color for each of them, what do you think?

Thanks again,
Stephane

[Niall Saunders]

Hi all,

If my understanding is correct, Juan seems to be suggesting that the image be processed with DBE - TWICE - using two different sample sets.

Then the two resulting images are either subtracted from, or divided into, each other - and the resulting image examined with the 3D viewer.

If the DBE routines have been successful, then the final image should have a 'constant' ADU across the field.

Is that what you were suggesting Juan?

Cheers,

[Juan Conejero]

Hi Niall and Stephan,

The idea is to generate a test background model after applying the real model generated with the first DBE instance. If the image was well corrected for gradients, the test model should be flat (within tolerances). This is a sort of counter-test. Of course, the test DBE instance must be different from the real one (it must define different background samples).

You can also apply DBE iteratively two or more times if you want; I do that occasionally for the sake of accuracy.