ImageIntegration noise weighting?

[mschuster]

On a particular target my raw dithered subs have almost equal MRS noise values. After calibration the same. But after registration resampling the noise values vary quite a bit. The weightings in ImageIntegration end up spanning the range from roughly 0.7 to 1.9. Also I noticed that these weights do not correlate well with the raw sub's FWHM measurements. Some subs with larger FWHM get weighted more than some with smaller values. Does it make sense to go ahead and use the ImageIntegration noise weighting option given this situation or should I do something else?

Thanks,
Mike

[Ignacio]

Interesting, I have made similar observations. Not with so much variation on the weights, but I do see little correlation between what I consider a good subexposure (low fwhm, low background level, nice shaped stars, etc.) and the weighting factors.

Are the weights determined from noise estimates, or from SNR estimtes?

BTW, another observation is that the batch processing script doesn't seem to pick the user selected reference sub at the integration step. Don't know what happens at the registration step, if it uses or not the right frame.

best
Ignacio

[mschuster]

The weights are determined from noise estimates and not from SNR estimates. I believe the MRS noise estimates measure background noise (read noise plus dark current shot noise plus sky flux shot noise).

The NoiseEvaluation script produces a noise estimate as well as the number of pixels selected for the estimate, where the selected pixels do not contain significant signal as determined by the MRS algorithm.

I plotted MRS noise versus pixel percentage for both my registered and calibrated subs. You can see a clear correlation for the registered subs but not for the calibrated sub. This is unexpected. I wonder if this indicates a problem in the MRS algorithm.

It would be helpful to see a map of selected pixels to investigate further.

[georg.viehoever]

Maybe it is somehow related to this http://pixinsight.com/forum/index.php?topic=3320.msg22928#msg22928 ?
Georg

[mschuster]

Georg, thanks for the helpful thread. MRS noise convergence failure in ImageIntegration remains a problem, reported here. I have seen the failure again on another, more recent H-alpha integration.

The sets of noise pixel locations used for noise estimation varies across the subs and the integration. Choosing these locations is not a noise free process. I wonder if unifying these locations makes sense. Here is the idea: Do an integration and determine noise pixel locations in resulting integration. Redetermine all sub noise values using this set of noise pixel locations. Redo the integration using weights derived from these noise values. It seems to me the weights now correspond better to the final measurement. I think this is an example of using a priori information, as suggested more generally in the original Starck-Murtagh paper.

Mike

[georg.viehoever]

...The sets of noise pixel locations used for noise estimation varies across the subs and the integration. Choosing these locations is not a noise free process. I wonder if unifying these locations makes sense. Here is the idea: Do an integration and determine noise pixel locations in resulting integration.... I think this is an example of using a priori information, as suggested more generally in the original Starck-Murtagh paper.
...

Some time ago I read a paper (see http://pixinsight.com/forum/index.php?topic=4161.msg29198#msg29198,  "Removing noise from astronomical images using pixel-specific noise model") where they determine "noisy" pixels and use this information to created a maximum likelihood image. While I dont like the procedure outlined in the paper,  I think that's an idea worth to follow. It may be possible to determine noisy pixels from the darks that we need to shoot anyway.

Georg