PhotometricColorCalibration results look too green

[sharkmelley]

I've been playing with PhotometricColorCalibration for my (modified) DSLR images.  It's a great tool and so far has worked flawlessly.  It gives consistent white balance multipliers across multiple images.  Credit to the very hard work put in by its authors!

However I do have a question.  Although the results are consistent from image to image, I am finding that the resulting images subjectively look too green.  Similar comments have been made on the following Cloudy Nights thread:
https://www.cloudynights.com/topic/586662-pixinsight-185-photometriccolorcalibration-localnormalization/

It's not just subjective, I have an objective data point as well.  If I take a daylight image of a ColorChecker test chart then I know what white balance and camera colour matrix will render the test chart colours correctly.  But if I use the same white balance and colour matrix on an astro-image and feed that image into PhotometricColorCalibration using "G2V star" as the white reference then PCC gives me white multipliers of approx. [0.76, 1.00, 0.73] i.e. it boosts the green channel quite noticeably from an image I believed was already correctly colour balanced.

I'm curious to know what might be the explanation of this apparent discrepancy.

Mark

[sharkmelley]

I have read the documentation now (https://pixinsight.com/tutorials/PCC/index.html).  It's an excellently written document and the examples are great.  I also take the point that a G2V white reference may not be ideal for many astro-images.

The document also gives me a better understanding of how PCC should work.  So if I'm supplying a properly daylight white balanced image to PCC then the population of G2V stars will be white in that image.  So looking at the graphs produced by PCC, on the B-V graph at the point on the x-axis where G2V stars appear (i.e. where B-V is around 0.6) the camera B-G will be precisely zero because the star will be measured as white.  And when finally it comes to reading the G2V adjustment for the white reference from the linearly regressed curve it will also be zero.
 
I'm now convinced that PCC is behaving correctly - I just need to examine my own workflow and carefully analyse the stars in my image to see where the problem originates.

Mark

[sharkmelley]

I have two versions of the same image - one using "standard" integration and one using Bayer Drizzle with x2 scaling.  I applied the same daylight white balance to both - the white balance obtained using a ColorChecker chart.  I then used PCC on both images.
PCC gives G2V white balance factors of [0.97, 1.0,  0.98] for the drizzled image which confirms that my daylight white balance is pretty accurate.
However,  PCC gives G2V white balance factors of [0.80, 1.00, 0.76] for the non-drizzled image.

Dividing the original drizzled image by the original (resampled) non-drizzled image gives a featureless result full of greenish stars, so this confirms what PCC is telling me i.e. that the stars (and only the stars) in the non-drizzled image are lacking green.

So something in my standard worflow is reducing the amount of green in stars.  I'm guessing this might happen either in the debayering or in the star alignment.  I'm using a Sony A7S on a 500mm focal length scope so the stars are undersampled (FWHM in the stacked image is 2.3px).  Some further detective work is required when I get some time.

Mark

[vicent_peris]

Hi Mark,

Thanks for the info. The problem should be then with the debayering algorithm. Which debayering method are you using? I guess you're using VNG. Could you try with Bilinear?

Best regards,
Vicent.

[sharkmelley]

It doesn't look as if VNG interpolation is the problem.

I've re-run the data using Bilinear, VNG and then 2x Bayer Drizzle so I can be absolutely certain everything is consistent.  Using PCC on daylight white balanced data, the white balance factors I obtained are:
2x Drizzle:  [0.94, 1.00, 0.97]
VNG:          [0.87, 1.00, 0.86]
Bilinear:     [0.88, 1.00, 0.87]

I was also very careful to set the PCC saturation threshold correctly in each case.  I'll carry on digging into it.

Mark

[sharkmelley]

Maybe the following image gives a clue as to what may be happening.  I have divided the VNG debayered and stacked image by the drizzled image.  Some of the brighter stars are saturated (they have grey cores) but if we ignore those stars, the most obvious feature is the green core and the purple/violet ring.  If I use a Bilinear debayered image instead of the VNG debayered image then the result is pretty much identical.

Maybe the debayering is spreading the light from the R&B channels over a wider area and somehow this ends up affecting estimates of magnitude in the 3 colour channels in the stacked image?

Mark