PixInsight DBE - Background gradients in nebulous regions

[sharkmelley]

PixInsight DBE (Dynamic Background Extraction) is a fantastic tool for removing background gradients in images.  You drop markers on the points that you know should be black background sky and it "magically" calculates the background gradients and subtracts them.  But what if my image has large areas of faint nebulosity, so there are very few points that I consider to be background sky?  I often have a background gradient but dropping a marker on an area that contains faint nebulosity will eliminate the nebulosity along with the gradient.

Does anyone else have this problem?  What's the best way around it?

Mark

[wimvb]

Clone the image and apply an extreme stretch to it. A STF with the channels not linked will be a good starting point. Make sure that you see all the nebulosity. Then use this image to place samples.

When done, drag an instance of the DBE onto the workspace. Close DBE and the clone. then apply the instance of DBE that you just created to the original (unstretched) image.

I find Alejandro Tombolinis website a good reference for tricky situations, expecially the process examples

http://pixinsight.com.ar/

Hope this helps,

PS

Just thought of another way, that I haven't used myself, but might work:

google an image with the same object as your image. Choose an image that also shows the nebula. Use that image as a reference when placing samples. You can not use this image to place samples on, but you can have it as a reference.

Just a thought.

[sharkmelley]

40 minutes ago, wimvb said:

Clone the image and apply an extreme stretch to it. A STF with the channels not linked will be a good starting point. Make sure that you see all the nebulosity. Then use this image to place samples.

 

Thanks for your reply.  Now we're getting to the crux of the issue.  The structure of the background gradient is quite obvious to the eye, sitting "behind the nebulosity".  But there are insufficient "gaps" in the nebulosity in which to place the samples I need to adequately capture the gradients

Mark

[wimvb]

If you judge the gradient to have some symmetry, you can use the symmetry settings in DBEs, target view panel. Each sample is than matched with another part of the image, depending on where you put the symmetry axes.

This can work if you need to remove vignetting, or if your gradient goes off in a definite direction

I wrote down more details here:

http://wimvberlo.blogspot.se/2016/08/note-on-dynamic-background-extraction.html (target view)

[sharkmelley]

3 hours ago, wimvb said:

If you judge the gradient to have some symmetry, you can use the symmetry settings in DBEs, target view panel. Each sample is than matched with another part of the image, depending on where you put the symmetry axes.

This can work if you need to remove vignetting, or if your gradient goes off in a definite direction

I wrote down more details here:

http://wimvberlo.blogspot.se/2016/08/note-on-dynamic-background-extraction.html (target view)

Those are all helpful suggestions and your write up is a useful intro.

In my head I'm beginning to crystallise exactly what it is that I want.  I want DBE to generate the background model using points that I "know" contain only background sky.  After that I want the ability to place an additional set of "adjustment" points that allow me to interactively tweak the background model (I know interactivity isn't a strong point of PI).  I could then put some adjustment points in the places where the background model is "correct", to tie it down and then further adjustment points where I can interactively adjust an RGB offset and see the effect in real time.  A similar process in Photoshop would be to select an area, give it a large feathering and then adjust the black point for that selection - iteratively over the whole image.

A background model followed by interactive tweaking sums it up quite nicely I think.

Mark

[wimvb]

You can always do multiple passes of dbe, refining the model as you go. Otherwise, if you choose to create a new image during dbe, you can inspect that and adjust samples in the original. Dbe doesn't have a preview option like other tools in PI.

[sharkmelley]

Playing around further I've discovered that DBE can already do what I need, up to a point.   It is possible to place a sample and tick the "Fixed" box for that sample.  Override values can then be manually entered into the R/K,G,B boxes and the background model will use those values instead of estimating them from the pixel values in your sample.  In this way it becomes possible to have control over the background model even in areas where there are no gaps in the nebulosity showing pure background.

By choosing to place fixed points only, it's actually possible to have full control over the background model - it's just very awkward to do it.

Mark

[wimvb]

That's a neat trick. But as you say, it's awkward to do this. Lots of trial and error.

[ollypenrice]

While I understand why you'd want to be sure not to include nebulosity in placing the markers I find that, in reality, you can get away with it quite well provided the threshold setting is not too high.

Because I work from a dark site I'm rarely dealing with significant brightness gradients, it's colour gradients that are the problem. (I realize that they are essentially the same thing but the eye is incredibly sensitive to colour gradient.) One thing's for sure, you don't need many markers for DBE to work well. Have you tried settling for second or third best in terms of placements and seeing how it performs?

Using a reference image from a dark site can be useful, as Wim says. I only know this because I've worked on guests' images from LP sites at their homes and used my own as reference for markers. When stuggling for marker points I aim for one in each corner and one in the middle, or as close to that as I can get.

Olly

[sharkmelley]

5 hours ago, ollypenrice said:

Because I work from a dark site I'm rarely dealing with significant brightness gradients, it's colour gradients that are the problem. (I realize that they are essentially the same thing but the eye is incredibly sensitive to colour gradient.)

You've hit the nail on the head there.  Yes, the eye is remarkably sensitive to colour gradients and that is exactly where I see both the problem and a potential solution  With an image of a galaxy in a black sky it's straightforward to place the DBE markers and remove gradients.  Once you have Integrated Flux Nebula, Nebulosity and Dust, the eye continues to see gradients in areas where placing a marker is all but impossible without damaging real structures.  So this afternoon I started playing with an alternative idea using Photoshop.  The idea is to stretch the image to within an inch of its life (using "Levels" layers to preserve data linearity) and then add 3 "Curves" layers - one each for R,G,B and each having a "Hide All" layer mask.  The "red" layer is given a "Black" offset in the red curves channel and similarly for the green and blue layers.  Then a judiciously chosen brush size and opacity level can be used to completely remove any background tint (simply by painting it away) beginning with large areas of background colour which require broad brush strokes and then continuing on to the smaller scale problems with smaller brush sizes.  Such an approach uses the power of the eye both to detect and to subsequently remove background colouration - all interactively.  If a particular broad brush stroke is too aggressive then undo that step, change the brush parameters and try again.  This is something a human is very good at. The end result is a complete background gradient model.  Once this is subtracted from the original linear data the result is linear data with the background gradients removed.

I'm sure none of this is new to experienced Photoshop users but I'm new to Photoshop and the power of such a technique is quite a revelation to me. 

Mark

[ollypenrice]

I must try this. I have never done anything like it because I don't have the need, but I'd still like to understand the technique.

Olly