Software Binning and WBPP

[Cygnusx1]

I've got a lot of frames I'm running through WBPP w/wo Astrometric Solution enable, w/wo downsampling, dynamic cropping, autocropping in WBPP and anything else I can think of that might affect image solving. I'm finding odd behaviour in Image Parameter initialization and inheritance and what you can change without creating a new ImageSolver script instance (like resoultion). I figure if I can get a rough estimate of average frame S/N and it's consistent across frames then I'll be comparing apples to apples. Ultimately, here, I'm just trying to understand how the ImageSolver script works. I have found virtually no documentation for the script (404 error on Pixinsight) so I'm trying to figure it out the old fashioned way, trial and error. Below are 10 frames taken over 2 days. The plate centers seem very close to me. So I thought I'd look at a measure of frame S/N. Figuring out what parameters this process is most sensitive to, Documentation sure would be nice.

 

[Cygnusx1]

Just finished analysis of a few variables affecting image solve for color calibration. I used Red frames taken with an ASI2600MM. I integrated 60 minutes of frames (3minx20 frames). I ran 8 cases through WBPP with Bin1x1 and resampled to yield Bin2x2. For the 2 binnings I ran with Astrometric Solution on and off and AutoCrop on and off. Here are the two summary matrices :

	NGC5457 Red Bin 1x1​
	RA​	Dec DMS​	S/N(dB)​	Solved​
No AS/No AC​	'14 03 10.688'​	'+54 20 49.14'​	34.58​	Yes​
No AS/AC​	'14 03 14.898'​	'+54 22 27.22'​	34.75​	No​
AS/No AC​	'14 03 10.685'​	'+54 20 49.22'​	34.58​	Yes​
AS/AC​	'14 03 30.219'​	'+54 20 59.53'​	34.75​	No​
	AS=Astrometric Solution​	AC=Auto Crop​

	NGC5457 Red Bin 2x2​
	RA ​	Dec DMS​	S/N​	Solved​
No AS/No AC​	'14 03 14.898'​	'+54 22 27.22'​	41.58​	No​
No AS/AC​	'14 03 14.898'​	'+54 22 27.22'​	41.99​	Yes​
AS/No AC​	'14 03 10.681'​	'+54 20 49.17'​	41.59​	Yes​
AS/AC​	'14 03 11.988'​	'+54 20 54.82'​	41.99​	Yes​
	AS=Astrometric Solution​	AC=Auto Crop​

Here are links to the processed data files if you feel so inclined to play :
1x1
https://drive.google.com/drive/folders/1ADMqQXAGzMkp8iXuTiBT4HloZvo52Dt_?usp=sharing
2x2
https://drive.google.com/drive/folders/1SGv4gaOuAfZT7F99M0OKURnSZ-wvC58A?usp=sharing
I'd appreciate any feedback. Just thought this might be useful for anyone who likes to play in the data.

 

[fredvanner]

I would be interested to know what definition of S/N you are using. I know a variety of ways of estimating the noise level, but I don't know any meaningful way to turn this into an SNR; each pixel is effectively a separate sensor, but with widely varying signal input, so an "all image" SNR makes very little sense. If you just average the ratio of each pixel signal value to the estimated noise level this will tell you very little about the image, since it will depend strongly on how much of the image is background (mainly noise).
If you are using a multilevel support method to estimate noise, one of the data structures produced by this method is a support array for the non-noise component of the image. Restricting the SNR estimate to this part of the image would make sense, but I don't know of any utility to do that.

 

[Cygnusx1]

Fred, ABSOLUTELY! I've been looking for a tool, model, script ... anything that could help me get my arms around S/N in any frame to use as a comparative metric (probably the most important). I'm still looking. I certainly never expected A single number for a spatially distributed 24 mega pixel frame, which is what I was given with the PI SNR script I found. Again, no documentation (that I have found) on the analysis or methodology used for generating the "SNR" from the script. I very much hesitated to use it here but since it was a comparative metric (I could have called it "Apples") I included it. If you know of any tools out there to use in PI that can give a comprehensive analysis of the signal and noise characteristics of a data frame (I'd love to see an analysis of an image's spatial correlation/decorrelation properties) that would be wonderful. By non-noise component do you mean a set of pixels above a set threshold or offset. BTW I am looking for the parameters that affect Image Solver plate solving. In this case the 'feature' was SNR (Apples) which did not seem to affect the result. Good stuff. all!

 

[fredvanner]

the PI SNR script I found.

What script was that?

 

[Cygnusx1]

This zip-file contains the script SNR.js

Move 2637 Bytes SNR.js to \\NASA\Sources\Scripts\Release\SNR.js


Hartmut
hvb356@hotmail.de

 

[fredvanner]

This script appears to use the definition:

SNR = (mean square signal) / (mean square noise)

where "mean square signal" is exactly that: the average of the squared pixel data values, and the noise statistics are taken from the multilevel noise estimation routines now built into the PI image object.

This can provide a comparative SNR for different images of the same region, but has no meaning I can see outside this comparative context.
... but that is what you are using it for, so it looks OK.

 

[Cygnusx1]

This script appears to use the definition:

SNR = (mean square signal) / (mean square noise)

where "mean square signal" is exactly that: the average of the squared pixel data values, and the noise statistics are taken from the multilevel noise estimation routines now built into the PI image object.

This can provide a comparative SNR for different images of the same region, but has no meaning I can see outside this comparative context.
... but that is what you are using it for, so it looks OK.

Can you please repost the link on multilevel noise estimation?

 

[fredvanner]

The link is in this post. I've attached a copy of the paper.

 

[Cygnusx1]

The link is in this post. I've attached a copy of the paper.

Much appreciated.

 

[fredvanner]

Once you have read the paper you can try and make sense of the PixInsight code.
I tried to experiment with this, using MMT to generate the separate wavelet layers - but PI keeps renormalising the data. I want the raw unnormalised data for each layer so that the separate layers add up to the original image.

 

[Cygnusx1]

Once you have read the paper you can try and make sense of the PixInsight code.

I tried to experiment with this, using MMT to generate the separate wavelet layers - but PI keeps renormalising the data. I want the raw unnormalised data for each layer so that the separate layers add up to the original image.

Well you've got me going. Never used the AtrousWaveletTransform, but I'm digging into it and the code now along with the RANSAC method and how it's implemented in PI. Also, StarAlignment. I'm thinking maybe doing a "forced" fit to Image solve. I'm assuming there used to be documentation for the ImageSolver script but it has vanished into the infamous Error 404 abyss.

 

[fredvanner]

The ImageSolver documentation was basiclly user interface documentation, not documentation of the design and implementation, so I'm not sure it would have helped you.

I'm thinking maybe doing a "forced" fit to Image solve.

I'm not sure what you mean. If I have persistent problems with ImageSolver (usually of a poor quality image), I usually resort to ManualImageSolver.

 

[Cygnusx1]

Actually, I was referring to image registration (in WBPP). I have not used ManualImageSolver and have tried to avoid it after reading the procedure. I want to understand the other processes like the RANSAC method and AtrousWaveletTransform first.