Question on BG Removal?

[dmcclain]

My images tend to develop what I believe are artificial rings of brightness near the edges of the image after background removal polynomial orders above 1. See attached example: left is before background removal, right is after processing. The ring is roughly the circumference of the intensity dome on the left.

The images were confusing at first, being images of regions in the plane of the Milky Way, up near Cassiopeia. But after seeing the ring-like structure in every image of a sequence of survey images over HA scans, I had to believe these rings were being artificially induced. My images are 7x10 deg FOV, so there will be lots of structure in the background. But not conveniently circular arcs of the size of my frames in every image.

I spent the entire day yesterday looking for flaws in my flat field process, and finally after obtaining the very finest bunch of flats I have ever seen -- including all the dust motes and slight corner vignetting, and using those in image calibration, I still obtain these rings in background removed images. In truth, there is no vignetting, but there is a very slight attenuation (e.g., 1 dB) near the frame corners.

I have arrived at the conclusion that these frame arcs are caused by:

  1) extreme nonlinear stretches,

  2) on data that are photon starved,

  3) and the 2D background curvature isn't a simple polynomial so that the residuals get stretched to produce these rings.

I thought about using Zernike polynomials, only because they restrict the polynomial to a radial coordinate (radial Legendre polynomials). But 2-D Cartesian polynomial fitting should really produce the same results, along with finding the center of curvature anyway.

So, taking a hint from Zernike polynomials and symmetry considerations, I tried progressive even-order background removal rounds in the L component of the image, orders 2, 4, 6, 8, until the background model stops showing hints of ring structure. While that helps somewhat, there are still traces of ring remaining in the images after background removal.

I'd hate to use cropping and throw almost half of the data away. Does anyone have suggestions for avoiding these image artifacts?

 

[aworonow]

Despite using flats, sometimes one seems to get this extreme vignetting. I think haze/high humidity may be the cause. Corrections are difficult, but some significant remediation is possible. You might have a look at http://www.lightvortexastronomy.com/2015/11/tutorial-pixinsight-reducing-light.html#SECTION1 as a starting point.

Here is what I have done: Using Dynamic Background Extraction, try boosting the Tolerance way up (maybe above 4.0 even), increase the number of samples/row, place a large number of samples in the corner areas, remove any samples clearly contaminated with stars or features of interest, reduce smoothing somewhat, and use division as the correction, not subtraction. Then try it. Remember, if you want to apply the same samples and parameters to another image, drag an instance of DBE to the desktop (drag the triangle to the desktop).

I wish I knew how the symmetry stuff at the top of DBE is used. Maybe it is the key...maybe not!

Alex Woronow

[dmcclain]

I did a bunch more work last night and found a number of interesting things...

1. Flats vary all over the place - changing dust motes, changing vignetting, even some evidence of camera electronic induced patterns.

2. When I perform ABE against reduced flats in PI, I found that in my case that the residuals show a declining MAD versus increasing fitting order, as would be expected.

3. When I look closely at the measured CCD parameters for the camera, and get all the scaling correct, my camera (EOS 6D in this case) is really only as good as slightly below 10 bits of dynamic range. That tells me that I should use an ABE estimator that produces residuals about 1e-3 or better, and that happens for order 8 and higher.

4. There is a trade-off between higher fitting order and wanting not to take out real structure in the background.

5. Given the limited true dynamic range of my sensor, I should not expect to stretch by more than 1e3, lest I end up stretching residual noise patterns and thinking I may have found true sky structure. Even for deep stacks, where you can potentially increase the background SNR, the fixed pattern noise will remain at the same level, regardless of the depth of stacking. PI allows us to amplify with abandon, like an electron microscope for image processing. But there is a limit to reasonable stretching.

And so, my big residual donuts are right down there at the level of the fixed pattern residual noise in the sensor. My histograms are nice broad bell curves at around 0.25 in a color calibrated and nominally stretched image. I should not try to stretch the body of that bell curve, but should only stretch above about 3 sigma and higher.

Photon starved is the key word here, and I shouldn't be trying to dig for more than my fair portion.