36 replies to this topic

[sharkmelley]

Here's the latest Sony A7S problem I'm working on.  Image taken with the Tak Epsilon under 20.8mag/arcsec/arcsec skies:

 

 

There's a wide purple swathe going top to bottom through the image left of centre.  This is a particularly bad example but all my images display a similar problem to a greater or lesser degree before I apply heavy gradient removal techniques.

 

I've already eliminated darks and bias frames as the cause - the problem remains when the darks are removed and when the bias is replaced by a synthetically generated bias.  So the issue is probably some kind of mismatch between lights and flats.  The flats are dusk sky flats, so banding from an artificial light source is not to blame.

 

In this particular example there is a fairly sharp change from purple to green at the left edge of the purple swathe.  It could be related in some way to the concentric coloured banding caused by the in-camera scaling applied to the digital data but if so, why would it appear as a broad vertical swathe?

 

Another possibility is the UV/IR filter I'm using or the anti-reflective coatings on the replacement glass I used when modifying the camera.  But why would that manifest itself as vertical swathes?

 

My current line of attack is to shoot various sets of flats at different ISOs and with different positions of the back of camera histograms.

 

If anyone has any other good suggestions then I'll be happy to hear them.

 

Mark

Edited by sharkmelley, 04 May 2017 - 05:13 PM.

[mmalik]

Can you share just a light for folks to look at; in short, how does a light look by itself? Regards

[sharkmelley]

A single 30sec ISO 10000 light looks just fine and so does the ISO 10000 flat.  Below is a jpg straight out of the camera (but scaled down in size) and also one of the flats (scaled down).  There's nothing odd about either of them.  Each calibrated light looks fine as well.  But the problem is quite obvious once 400 calibrated lights are stacked in PixInsight.

 

I've also tried performing a straight average of the 400 raw lights (not star registered) and dividing by a straight average of the 100 raw flats (subtracting a bias from both).  The result is also below.

 

The average of 400 lights looks quite normal and so does the average of 100 flats. But dividing one by the other followed by stretching of the data shows the problem.

 

The earlier issue I had with concentric coloured banding happened in a similar way - everything looked quite normal until one was divided by the other.  This revealed anomalies of very low amplitude that were not previously visible.  Maybe this is a similar kind of problem but there's no evidence for it yet.

 

Mark

 

 

 

[mmalik]

Try a different calibration/stacking program for experimentation sake, e.g., an eval of ImagesPlus. Regards

[sharkmelley]

Try a different calibration/stacking program for experimentation sake, e.g., an eval of ImagesPlus. Regards

I'll try that as a last resort but there should be no need because PixInsight stores all intermediate images in full floating point so it should not be introducing artifacts.

 

One of the things I always notice about my bias subtracted master flats is a left to right colour gradient.  It becomes apparent when the saturation is turned up high:

 

 

 

However I take my flats the gradient is the same: green on the left and purple on the right.  In theory this shouldn't be a problem because even if it originates in the sensor then it would affect lights and flats equally.  But it is a strange observation nonetheless.

 

Mark

[sharkmelley]

I have made another observation that might turn out to be important.  I now have a whole load of master flats.  They all cause purple and green colour gradients in the final image but some are worse than others.  The master flat I used at the beginning of this thread happens to be the worst one.

 

If I divide the worst (bias subtracted) master flat into the best (bias subtracted) master flat, this is what I see:

 

 

To the left of centre, notice the same purple swathe with the same green swathe on its left as we saw in the first image.  I think this is telling me that part of the problem (or maybe the whole problem) is in the flats.

 

So there are certainly weird purple and green gradients buried in the master flats.  The question is if those gradients are already in the data coming from the camera or if something in the processing sequence is generating those gradients. 

 

Mark

Edited by sharkmelley, 06 May 2017 - 02:47 AM.

[mmalik]

The question is if those gradients are already in the data coming from the camera or if something in the processing sequence is generating those gradients.

Try one more program other than PixInsight as a control, may be even DSS; once software is ruled out then you can focus on the acquisition. Reverse approach will be take new flats and use them as control against your existing data using PixInsight. Regards

[sharkmelley]

Although I had no reason to doubt PixInsight I tried processing the data in another package - my old favourite IRIS.  It gave the same result even though it doesn't operate completely in floating point format.  So that rules out software issues as far as I'm concerned.

 

Performing star alignment on 100's of lights is also far too time consuming for these test runs, so I'm skipping alignment for now.  Look at the difference between the following 2 examples where I've used different master flats:

 

 

 

The purple swathe in the first image has begun to disappear in the second image.  Instead, an additional purple coloured band is encroaching from the left and possibly from the right as well.  So what is that telling me?  It's a clue, certainly.  But a clue to what?  What exactly is going on in those flats?

 

The sequence of coloured bands (in the second image I think I'm seeing purple, green, purple, green, purple) is reminiscent of the concentric coloured bands issue.  But I'm deliberately shooting my data to avoid the concentric banding triggered by the Sony histogram gaps .  But now I seem to be getting coloured vertical swathes.  I'm mystified!

 

Mark

Edited by sharkmelley, 07 May 2017 - 08:34 AM.

[sharkmelley]

The approach I'm now taking is to forget about the light frames but to create lots of different master flats, bias subtract them and then divide one by the other.  Three different results are observed:

• Concentric banding effects (which are already fully understood)
• Vertical stripes
• Flats behaving properly

For your entertainment, below are a couple of examples:

 

Division of 2 master flats at ISO 2000:

 

 

Division of 2 master flats at ISO 10000:

 

 

Of course the result has been severely stretched to reveal these effects, but the stretching is the same level of stretching required to reveal the IFN in the original stacked data - so it is a realistic test for deep sky astro.

 

The ISO 2000 example clearly shows the horizontal anomaly (or anomalies) on the left side of the split sensor but that has already been discussed elsewhere and my imaging technique is designed to avoid it occurring.

 

Usually when confronted with an issue, I can visualise the beginnings of a potential mathematical explanation and then perform experiments to concretise a full explanation.  This time I am left floundering.  I simply cannot understand how dividing one vignetted flat by another vignetted flat can possibly result in artifacts that are aligned vertically.  I'll carry on doing tests and await that flash of inspiration.

 

I'm not hopeful, however.  I'm beginning to think it may be the time to replace the Sony with a camera from another manufacturer.  The Sony A7S is just too full of flaws for deep-sky astrophotography.  I don't like to think about the amount of time I have spent investigating the various issues on this camera.

 

Mark

[sharkmelley]

Progess so far:  I've increased my shutter count by a couple of thousand taking sufficient flat frames to produce master flats at various ISOs and various levels of brightness.

 

One inescapable conclusion emerges:  two master flats at different brightness levels will not calibrate each other even if their ISO is identical.  By "different brightness levels" I mean the histogram peak is in a different place.  The calibration causes wide coloured stripes to appear in the data for which I have no explanation.  I'm guessing that some kind of data manipulation is taking place after the data has been read from the sensor.

 

Possibly, just possibly, it might be the case that if you take a whole series of master flats with the histogram peak in different places, it might be possible to find a master flat whose Red channel peak matches the Red channel peak of the master light.  Then look for another master flat where the Green channel peak matches the Green channel peak of the master light.  Then look for another master flat where the Blue channel peak matches the Blue channel peak of the master light.  Strip out the relevant channels from those master flats and re-combine them to create a new master flat that might work on the master light.

 

I'll give this a try. 

 

Mark

Edited by sharkmelley, 09 May 2017 - 12:25 AM.

[ccs_hello]

Mark,

 

If you see repeated pattern (post #9), I'd suspect processing caused aliasing (e.g., multiplication, especially on-SoC-chip digital multiplication/digital gain.)

 

RE: post #10

If DC bias is dynamically set (based on histogram peak), that has to be post-processing.

If you check Nikon DSLR's DSP processing, indeed the processing is per-color-channel based (though DC bias is fixed, as set by SoC config register.)

 

BTW, about FF sensor stitching as had been discussed in this forum, there is a set of mask pictures in this external blog article

https://landingfield...-image-sensors/

 

Clear Skies!

 

ccs_hello

[sharkmelley]

I have finally made some progress in understanding this problem.  After many false starts I eventually found a method of analysis that is giving useful information.  It stems from the earlier observation that there is left/right green to purple gradient in the flat frames.  Separating out the R,G&B channels from the bias-subtracted flat, it was possible to divide one by the other.  This showed that the R&B channels were in sync but the G channel diverged. Here is the G divided by the sum of R&B:

 

 

Note that it is not an even gradient from left to right but there are distinct vertical bands.

 

Doing the same to the sum of the unprocessed bias-subtracted lights gives the following:

 

 

Again there is a left to right gradient but with a vertical band that is different from vertical bands in the flat frame.  Since there is a mismatch in the vertical bands between the lights and flats, then green and purple vertical bands will appear in the final image, because the green is not in sync with the red and blue.

 

Further analysis appears to show that the pattern and intensity of the vertical bands (in the G divided by R&B) differs according to ISO and even when the ISO is the same it changes according to how much light was collected.  Further work is required to get a better understanding of this.

 

It might also be relevant that this vertical banding appears to be confined to the left side of the split sensor - i.e. the same side where the horizontal bands cause problems.

 

Mark

Edited by sharkmelley, 13 May 2017 - 10:01 AM.

[mmalik]

It stems from the earlier observation that there is left/right green to purple gradient in the flat frames.

Is calibration necessary; why not skip it. I only do limited dark subtraction (~3 darks only). This approach is quite feasible with longer FLs. Regards

Edited by mmalik, 14 May 2017 - 02:03 AM.

[sharkmelley]

 

It stems from the earlier observation that there is left/right green to purple gradient in the flat frames.

Is calibration necessary; why not skip it. I only do limited dark subtraction (~3 darks only). This approach is quite feasible with longer FLs. Regards

 

For this image of the Integrated Flux Nebulosity, the stacked data needs to be stretched quite highly which means that the vignetting and dust spots need to be corrected to a high level of accuracy.

 

In any case, even without calibration, my analysis above shows that the stripes exist in the light frames.  So even on a theoretically perfect optical system with zero vignetting and zero dust, the data stretching will reveal those stripes.

 

Mark

[ccs_hello]

quote: It might also be relevant that this vertical banding appears to be confined to the left side of the split sensor - i.e. the same side where the horizontal bands cause problems.

 

Probably a good explanation.  On the other hand, it begs for future SONY full-frame sized image sensor development mandate.

IMO, SONY should have designed such split image sensor as if there were two completely independent halves.

This could avoid any type of left-> right (or right-->left) signal hand-off imperfections.

 

Mark, sorry you got the bad end of the deal (on that misaligned split image sensor.)

The high cost of such sensor perhaps caused SONY lowering its QC standard which may in the end, hurt its reputation.

 

Clear Skies!

 

ccs_hello

[t_image]

Random question.

Mark you are finding all of this on your self-modded camera?

https://www.cloudyni...-2#entry6571090

 

My wonder is whether you not having a Sony anti-aliasing filter in front of the sensor could be some of the cause,

maybe the in-camera processing of the data is purposefully meant to accommodate for what the AA filter does?

Just a thought?

[sharkmelley]

Random question.

Mark you are finding all of this on your self-modded camera?

https://www.cloudyni...-2#entry6571090

 

My wonder is whether you not having a Sony anti-aliasing filter in front of the sensor could be some of the cause,

maybe the in-camera processing of the data is purposefully meant to accommodate for what the AA filter does?

Just a thought?

Yes - the camera is my self modded one.  Without taking the camera apart again it is impossible to exclude that the anti-reflective coated glass replacement might somehow be causing a left to right colour shift across the frame.  The same gradient is found using lenses as well as the scope so it's not a scope related issue.

 

However, anti-reflective coatings, Newton's rings and all similar explanations cannot possibly explain whey the character of the vertical stripiness in that gradient changes from ISO to ISO and changes according to the amount of light recorded and changes between 12bit and 14bit modes.  This can only be an effect of the internal data processing.

 

That's why I have excluded that possibility.

 

Mark

[t_image]

Without taking the camera apart again it is impossible to exclude that the anti-reflective coated glass replacement might somehow be causing a left to right colour shift across the frame.  The same gradient is found using lenses as well as the scope so it's not a scope related issue.

Mark, my point was not your replacement filter but that Sony may be intentionally "cooking" the data intentionally based on calibrations around the physics of their anti-aliasing filter that is also the hotmirror of the a7s [which you removed].

 

Maybe if you give me very detailed instructions of what to do to repeat your methodology, I could take sample RAWs with my unmodded Sony a7s and send them to you for analysis to differentiate if what you are seeing is a general Sony sensor problem or maybe amplified by something particular to your situation???

 

-mind you I don't have a TAK, only lenses and a OST, and an AVX that can give me 2 minutes on a good day.

Edited by t_image, 16 May 2017 - 01:08 PM.

[sharkmelley]

 

Maybe if you give me very detailed instructions of what to do to repeat your methodology, I could take sample RAWs with my unmodded Sony a7s and send them to you for analysis to differentiate if what you are seeing is a general Sony sensor problem or maybe amplified by something particular to your situation???

 

 

A single flat frame taken at ISO 100 with the back of camera histogram well over to the right should be sufficient to show any overall colour gradient and any other anomalies on the left side of the sensor.  A second flat frame at ISO 100 with the back of camera histogram in the middle would allow one flat to be divided by another (once the bias of 128 or 512 is subtracted from each) and this might begin to show the stripe issue.  It would only require 2 RAW files.

 

Another useful thing would be to average 25 flats at ISO 2000 with the histogram over to the right and another average of 25 flats at ISO 2000 with the histogram at the 1/3 position.  These averages would need to be performed in floating point and saved as floating point TIFF files.

 

Mark

[sharkmelley]

Here's another example of stacked data with the same purple stripe running through it:

 

 

In this particular case it was very difficult to process because the dust cloud itself is also discoloured by the purple.  PixInsight DBE couldn't deal with it because you can't put sample points in the middle of a dust cloud!  I had to import the linear stacked data into Photoshop and use the technique of an "anti-purple" brush which could be applied iteratively to the both the background and the dust.

 

The result is here:

 

 

This was certainly the most difficult image I have ever processed.  It does show that it is still possible to produce decent results from my A7S despite the "purple stripe" issue but it is unnecessarily difficult.  Imaging really shouldn't have to be this hard!

 

Larger version of the finished result is here:  http://www.markshell...ust20161130.jpg

 

Mark

Edited by sharkmelley, 16 May 2017 - 06:30 PM.

[AnakChan]

Mark, that is a fantastic job in recovery there. All your effort there worked out well.

Regarding this A7S purplegreen though, I've not noticed it myself but mine is somewhat different that it's been CDS modified/cooled and I opted for the Hoya UV/IR filter during the modification request whereas I believe yours is a full spectrum.

I'm curious if other A7S full spectrum owners have the same findings as yours.

[sharkmelley]

Regarding this A7S purplegreen though, I've not noticed it myself but mine is somewhat different that it's been CDS modified/cooled and I opted for the Hoya UV/IR filter during the modification request whereas I believe yours is a full spectrum.

I'm curious if other A7S full spectrum owners have the same findings as yours.

I'm curious as well.  Yes, mine is a full spectrum mod with a replacement filter of plain glass having anti-reflective coatings.  It was needed to correct the optical path distance so that legacy lenses could still focus on objects at infinity and also so it was compatible with the very precise back focus distance of my Tak Epsilon.  However I really don't think the filters (either the factory fitted ones or the replacement ones) can result in the apparent non-linear behaviour I am observing.

 

All my previous A7S issues (star eater, split sensor, horizontal artifacts on left hand side, concentric coloured banding) have been duplicated by other people.  But this one requires either a well controlled test experiment setup or hundreds of stacked calibrated exposures followed by extreme stretching.  I just wonder how many people are using the A7S in that manner?

 

I know for a fact that the A7Rii suffers from the split sensor and left side horizontal artifacts.  I don't know if the Mark 2 cameras suffer from the concentric banding - it is possible that Sony doesn't apply the same digital scaling to the data.  It is certainly the case that firmware upgrades can change the character of the histograms i.e. change the internal processing of the data.  Jim Kasson has proof of this on the A7Rii.

 

If Sony lends me a batch of their cameras I would be more than happy to put them all through a whole suite of tests.  Including the A9

 

Mark

Edited by sharkmelley, 17 May 2017 - 12:30 AM.

[mmalik]

 

...mine is a full spectrum mod with a replacement filter of plain glass having anti-reflective coatings.

I would have recommended 'Removed LPF-2' mod... on a7S. More and more Sony cameras are now coming bonded, means 'Removed LPF-2' mod is not an option for most.

 

 

a7S is one of those rare and very desirable cameras out there that offers 'Removed LPF-2' mod, quite amenable to cooling..., very compact size, and has 'no' IBIS, a possible source of IR pollution. Regards

[sharkmelley]

I've just found on my hard drive the only flats that I took before the A7S was modded. They were taken on my C11 Telescope.

 

Here's a single ISO 100 flat, colour balanced to white then severely saturated in Photoshop:

 

 

Note the same green to purple gradient across the flat.

 

Here is a (bias subtracted) ISO 2000 master flat divided by a (bias subtracted) ISO 100 master flat:

 

 

Unfortunately the result is dominated by concentric banding but I think it is just possible to make out a single wide vertical stripe immediately left of centre.  This confirms (to me at least) that it is nothing I have done to the camera which is causing these artifacts.

 

Mark

[t_image]

Nice Mark, you beat me to it!

Quite scientific as it is the same device.

I never meant to imply your results were fault of you,

I was just trying to understand the logic behind Sony cooking the data,

as your results all along more indicate it is intentional data manipulation and not just a fault of the sensor (although the melding of two wafers than happens with most full frame sensor chips is partly to blame)....