Selecting reference image for integration

[TrickyDicky]

Selecting reference image for integration.
If all images are identical – it doesn’t matter, but let us assume that I have a series of images taken with different exposure times, ISOs, or F-stops
I was just wandering if selecting different references for integration will produce different results.
By default, integration uses the first image from the list, but there is an option to select a particular reference image.
When a reference image is selected, it sets its “weights” values to a “1”, and the other images’ weights are set accordingly.
So, which image should we select? Shorter, or longer exposure, higher, or lower ISO, which of them would be the best reference? Or, does it matter?
Thnx ....

[jerryyyyy]

Selecting reference image for integration.
If all images are identical – it doesn’t matter, but let us assume that I have a series of images taken with different exposure times, ISOs, or F-stops
I was just wandering if selecting different references for integration will produce different results.
By default, integration uses the first image from the list, but there is an option to select a particular reference image.
When a reference image is selected, it sets its “weights” values to a “1”, and the other images’ weights are set accordingly.
So, which image should we select? Shorter, or longer exposure, higher, or lower ISO, which of them would be the best reference? Or, does it matter?
Thnx ....

Not being very experienced, let me say what I do.  If I am collecting H-alpha, I use the first H-alpha I have collected.  I find this works and I cannot forget what I chose if I add lights later.  The integrations all come out the same as you add in more lights.  That being said, you probably could use any reference images as long as they properly plate solve and you get all the images to overlie correctly, so if you do not have H-alpha or Luminosity, I choose the first Green image. 

BTW, am integrating all my stuff from last night in the background as I write this and have found that I also get the best results with the linear fit integration parameter... but for that you need 15-20 lights... a lot of work...

[Juan Conejero]

Hi,

As for image weighting, the choice of a reference image is theoretically irrelevant under ideal conditions. However, given that we have to work under less-than-ideal conditions, these points can help you to select an optimal reference image for integration:

- If there are varying gradients in the data set, select the image with the least/weakest gradients. Gradients complicate calculation of fundamental statistical properties such as scale and location.

- Try to select the best image in terms of SNR. In general, this corresponds to the image with the least noise estimate.

- Avoid selecting a reference image with strong artifacts, such as plane and satellite trails, etc.

The SubframeSelector script by Mike Schuster, and the Blink tool by Nikolay Volkov, are your best friends in all image analysis and selection tasks.

[Eddy Timmermans]

I used the first frame of a series and it didn't work.
Then I used the Subframeselector tool and let it run on the entire stack of lights.
It turned out that my first 3 frames had the worst FWHM values.
I then chose the one with the best value as reference frame and everything went smooth from then on.

I can't check it right now, but is there a selection in the subframeselector that lets you sort images by SNR-values ? Is it a better choice than FWHM ?

Eddy

[GJL]

I check the images with CCD Inspector and use the one with the best FWHM.

Gerhard

[TrickyDicky]

Thanks all for sharing your experiences.

So far I’ve always had one image of lower exposure, which I used as reference for registration and then integration. But I wasn’t sure if that image of definitely lower weight wouldn’t somehow ruin the maths behind asserting weights to all other images.
So, here is question to Juan: is my logic right?

PS. I haven’t tried the SubframeSelector script yet.

[Ignacio]

The arguments here seem to apply more to registration than integration (i.e., pick the "best" sub as reference). I always thought that weights in image integration were relative, so it didn't matter which one had a value of "1".

Ignacio

[pfile]

i think probably the #1 factor that juan gave (pick the image with the least gradients) is probably the most important for II. if this throws off the scaling/normalization of the remaining frames, then the outlier rejection can't be as accurate.

[Ignacio]

I still don't get it. In image integration, don't you compute all parameters for each image independent of the rest, and then compare with a reference image parameters to come up with relative weights/scaling?

Ignacio

[jkmorse]

Just a quick add that if you aren't using the SubFrameSelector (SFS) tool you owe it to yourself to try it out.  I probably get carried away, but when I run the ImageIntegration tool I create small incremental changes to the sigma clipping settings (one parameter at a time, of course) and then repeatedly test to results with the SFS to hone in on the best result.  Once I get one sigma setting set, I do the same with the other.  SFS is the way to go since it is right there in the PixInsight work environment and makes my workflow that much easier 

JIm

[TrickyDicky]

The arguments here seem to apply more to registration than integration (i.e., pick the "best" sub as reference). I always thought that weights in image integration were relative, so it didn't matter which one had a value of "1".
Ignacio

Thanks Ignacio, but let’s put the registration aside.

If you say: weights .. were relative ..    so they were relative in comparison with what? If they don't matter, what are they used for?

I still don't get it. In image integration, don't you compute all parameters for each image independent of the rest, and then compare with a reference image parameters to come up with relative weights/scaling?

That makes two of us ..

[Juan Conejero]

I still don't get it. In image integration, don't you compute all parameters for each image independent of the rest, and then compare with a reference image parameters to come up with relative weights/scaling?

In the image integration task the images have to be normalized. Normalization makes the images statistically compatible, so that comparisons between estimates of dispersion and central tendency can be meaningful. Put in simpler words, normalization allows us to compare the histograms of all of the images, ruling out differences in signal strength and mean background values. You have an example in this post, and I describe this in much more detail in the new version of the reference documentation for the ImageIntegration tool, which I am finishing right now.

Accurate and efficient estimates of scale (dispersion, or variability) and location (central tendency) are crucial for image normalization. If an image has strong gradients, valid scale estimates can be very difficult or impossible to calculate, since the gradients "mask" the true information contents of the image. Large amounts of noise also complicate estimation of meaningful statistics. Since all images are normalized to match the reference image, inaccurate or meaningless reference scale estimates can be problematic and degrade the result. So it is important that the reference image be the "best" possible one in terms of gradients and noise, especially if these defects are comparatively large for specific images in the set being integrated. Obviously, these considerations can be irrelevant for good-quality data sets, where the algorithms work under nearly ideal conditions.

[Ignacio]

Thanks, Juan, I will look into the new documentation when available.

Ignacio

[jkmorse]

Juan,

In the typical workflow, ImageIntegration comes before applying DBE.  Would the same be true when creating a synthetic Lum from the RGB channels using ImageIntegration or should you first apply DBE to the RGB channels before running ImageIntegration to create the synthetic LUM to get the cleanest possible image without gradients?

Thanks,

Jim