52 replies to this topic

[2ghouls]

Inspired by a question in another thread about whether the Canon Ra does raw data scaling, I ran some tests suggested by Mark Shelley that he had previously run on his Nikon Z6.

 

Here is the method and results for anyone interested.

Method: 

1. turn off all lens corrections in camera

2. take flats from many stops underexposed to a bit overexposed. In this case from 1/100s to 1.6s.

3. subtract bias and separate the color channels

4. integer resample x2 in maximum mode

5. divide blue by the green channel and red by the green channel

6. very aggressive stretch with HistogramTransformation 

7. integer resample x4 in average mode

8. make contact sheet of the resulting images

 

Results:

Blue channel divided by the green channel:

 

Red channel divided by the green channel:

 

I also tried Mark's test of stretching a blue sky photo, but could not get colored rings to appear.

 

Cheers, Nico

[johnpane]

What is your interpretation of these results?

[2ghouls]

What is your interpretation of these results?

I've never seen 'Nikon-style' colored rings in any of my astrophotography with the Canon Ra, and this test didn't produce sharp obvious rings like in Mark's test of the Nikon Z6, so I don't think this would be a real world problem for the Ra. Apparently, It is also not a real world problem for the Nikon mirrorless cameras as long as you expose bright enough (histogram 1/2 way over). I do see the typical 'Canon-style' horizontal banding in the exposures longer than 1s., but then it seems to disappear in exposures shorter than 1s, which is interesting. No clue what that indicates. Canon horizontal banding mostly calibrates out with good dark frames (the issue sometimes with an DSLR/MILC is getting good dark frames though).

 

Meant to add to my original post:

 

-As far as Canon Ra read noise goes, I sent in my files to Bill Claff, and he confirmed that my Ra performed very similarly to his published result: https://www.photonst...Canon EOS Ra_14

 

-If anyone else has a test they would like see performed on the Canon Ra, I'd be happy to oblige.

 

-Nico

[Delta608]

You forgot the blue divided by red..Nico, Im a big fan, but hugh ....? You guys are taking all the fun out it...

Edited by Delta608, 03 December 2020 - 06:08 PM.

[sharkmelley]

That's an interesting set of results - thanks for doing them! 

 

There are certainly rings appearing but it's not obvious they are an artefact of raw data scaling and they don't have the obvious progression from one exposure to the next which we would expect from an algorithm applied to the data .  With the Sony and Nikon mirrorless cameras the artefacts that appeared were clearly caused by internal manipulation of the raw data - the sharp edges were one of the obvious signs.  We're not seeing those sharp edges in this Canon example which makes me think the rings probably have a different cause.  There's also no "join" down the middle of the sensor which we see in both Sony and Nikon. 

 

It might be worth repeating the test with lens corrections enabled just to see if there's any difference.

 

Mark

[Chris Cook]

Newton rings...   

[alan.dang]

The Ra is designed for astronomy and there was no time urgency to release it commercially. The R5 and R6 have RAW manipulation which is great for daylight photography. Canon could have gone with a RPa or R6a if that made sense. They didn’t... they chose the Ra as the platform.

The other thing is that we now have confirmation that at least two copies of the Ra have similar performance under Bill Claff’s analysis. We just need someone with a regular R to submit samples to him for analysis also. The more the merrier.

[sharkmelley]

Nico sent me the raw files used for his analysis and I get identical results.  But it also allowed me to investigate the cause in further detail.  For example, here is the RawDigger histogram of one of the files:

 

 

The thing to notice is the regular sequence of histogram gaps.  It is especially obvious in the green channel and is caused by the firmware applying a scalar multiplier to the raw data.  Certainly the blue and green channels are being scaled.  There are also other things going on in the raw data - look at the weird shape of the histogram of the blue channel for example.

 

Raw data scaling of this type will often lead to concentric rings in the image and this is the main cause of the rings we are seeing in Nico's analysis.  Unfortunately it also makes it pretty much impossible to determine if there is any kind of lens corrections being applied to the data in addition to the channel scaling.

 

The obvious question of interest is whether or not these rings might appear in astro-images, especially for deep sky objects where we stack exposures, subtract light pollution and stretch - all these operations increase the likelihood of making these artefacts visible.  It's difficult to be certain but I think as long as low ISOs and underexposure are avoided then it shouldn't be a problem.  The problem was far worse on the Sony A7S.

 

For anyone interested in the technical details of how raw data scaling causes rings, the most useful discussion took place over on DPReview:

https://www.dpreview...s/post/59378650

A one sentence summary is that raw data scaling causes ripples in the photon transfer curve that manifest themselves as subtle changes in brightness in the image.

 

Mark

Edited by sharkmelley, 13 December 2020 - 09:56 AM.

[2ghouls]

The obvious question of interest is whether or not these rings might appear in astro-images, especially for deep sky objects where we stack exposures, subtract light pollution and stretch - all these operations increase the likelihood of making these artefacts visible.  It's difficult to be certain but I think as long as low ISOs and underexposure are avoided then it shouldn't be a problem.  

I haven't seen them yet in real-world imaging, but the minimum ISO I've used is 800, and typically I use 1600. Would it be worth trying an imaging session at ISO100 or 200 to see if I could make the rings appear? I have so few clear nights, I am hesitant, but I will try to remember to try it.

 

For anyone interested in the technical details of how raw data scaling causes rings, the most useful discussion took place over on DPReview

I didn't understand all of that, but one aspect in particular confused me. Could you explain what is meant by 'adequately dithered' in the context of that discussion? Here's is your quote responding to Jack Hogan: "However, I argue differently. I'm arguing that data with regular histogram gaps might already contain visible coloured banding even though it is both adequately dithered and unposterized." Are you applying a dithering algorithm in manipulating the raw file? How would one know their sample is 'adequately dithered', etc.?

[sharkmelley]

I haven't seen them yet in real-world imaging, but the minimum ISO I've used is 800, and typically I use 1600. Would it be worth trying an imaging session at ISO100 or 200 to see if I could make the rings appear?

It would be interesting experiment if you want to try it.  The flat frames give some idea of what signal level is required to produce the rings so you should aim to expose the green channel to that level.

 

I didn't understand all of that, but one aspect in particular confused me. Could you explain what is meant by 'adequately dithered' in the context of that discussion? Here's is your quote responding to Jack Hogan: "However, I argue differently. I'm arguing that data with regular histogram gaps might already contain visible coloured banding even though it is both adequately dithered and unposterized." Are you applying a dithering algorithm in manipulating the raw file? How would one know their sample is 'adequately dithered', etc.?

In astronomy we use the word dither to mean spatial offsets between successive frames.  But it has a different meaning in the context of quantization of data.  Posterization occurs when there is not sufficient image noise to hide the discrete steps in data values and sometimes noise (or dither) is added to an image to prevent posterization occurring.  Sometimes sufficient noise already exists in an image. The point I was making to Jack was that there is certainly sufficient noise in the image to hide the image quantisation so there is no image posterization but there is not sufficient image noise to hide the larger scale variations in brightness caused by the raw data channel scaling.

 

Mark

[alan.dang]

https://www.dpreview...s/post/64259159
Yes, Bill talked about the digital scaling 4 months ago.

Isn’t this just white balance scaling to account for the intense Ha sensitivity?

https://blog.kasson....ikon-d850-ptcs/

[alan.dang]

No free lunch.

 

I think I finally get it.   I am getting spectacular dust lanes from my EOS Ra / VSD100 combo from relatively short exposures.  ~1 hour from Bortle 6/7 and I can get dust lanes off M45.  I can barely see Polaris to the naked eye on most nights.  [The comma in the upper right is dust on the sensor. One of the disadvantages of sensor auto-cleaning and being lazy is that when I went to take flats the following day, the dust was no longer in that same spot].

 

 

Roger Clark talks about the color of interstellar dust in Figure 8.

https://clarkvision....erstellar.dust/

There is an awful lot of relative intensity from dust in that 650 to 700 nm range.

 

There is also a lot of blue sensitivity in that range for the Ra.

 

It is my belief that the sensitivity in this >650-700nm range gives the Ra better-than-expected ability to pull out data from interstellar dust.  (And the trade off is the asymmetrical flares).

 

The Ra doesn't behave like most astro cameras.  It's massaging the RAW data, is adding a lot of near-IR sensitivity, and has a lot of banding in warmer climates.  But the end-stack seems to be punch above its weight, when the end-stack is an artistic astrophoto as opposed to photometry.

 

It would be interesting if someone could figure out how to install the R firmware on the Ra, which is presumably easier than doing the opposite.  The white balance pre-scaling in blue/green could change and it would be interesting to see if it has anything to do with the 650-700nm range.

 

@Mark, can you take a few ISO 1250, 30 second exposures on M45 on your Z6/Tak Epsilon setup for a cumulative 1 hour exposure time?  It would be interesting to compare.  The Tak Epsilon should pull in way more data than the VSD100 and Ra.  Not sure about your Bortle situation.  Maybe @Nico can do the same?  I have rain in the forecast, so that's all I can offer for comparison.

Edited by alan.dang, 14 December 2020 - 01:28 AM.

[sharkmelley]

https://www.dpreview...s/post/64259159
Yes, Bill talked about the digital scaling 4 months ago.

Isn’t this just white balance scaling to account for the intense Ha sensitivity?

https://blog.kasson....ikon-d850-ptcs/

Yes, we knew about the digital scaling months ago.  I should have looked more closely and warned about the possibility of coloured concentric banding. My bad!   

 

White balance pre-conditioning is very common in Nikon cameras, as you point out from Jim's blog for the Nikon D850.  So there is no link between Ha sensitivity and WB pre-conditioning in general. The WB preconditioning on the Nikon D850 is better designed than the EOS Ra because the histogram gaps are more closely spaced, so they are far less likely to cause imaging artefacts.  This is also true of the Nikon Z6 - the problems on the Z6 are caused by a different raw data manipulation which I think may be a colour shading correction and it's not visible in the histograms. 

 

The white balance preconditioning is present on the EOS R - see the ripples on the photon transfer curve (PTC):

https://www.photonst...htm#Canon EOS R

This shows Canon is not applying it because of increased Ha sensitivity.

 

The PTC ripples on the EOS Ra have a slightly different shape:

https://www.photonst...tm#Canon EOS Ra

It shows that the scalars for the WB pre-conditioning are different between the R and the Ra.  This difference could be a design decision resulting  from the stronger red channel of the Ha.

 

The mystery is why WB pre-conditioning exists at all.  Why don't camera manufacturers simply incorporate it into the WB multipliers which are applied during the raw conversion?  Why instead do they destructively scale the raw data?  Until recently we could rely on Canon not to interfere with the raw data in this manner.  What brought about this unfortunate change in policy?

 

Mark

Edited by sharkmelley, 14 December 2020 - 05:15 AM.

[2ghouls]

No free lunch.

 

It is my belief that the sensitivity in this >650-700nm range gives the Ra better-than-expected ability to pull out data from interstellar dust.  (And the trade off is the asymmetrical flares).

I guess that's possible, but remember the Astronomik L2 let's in most of that range (starts cutting off around 690nm) at a very high transmission (over 95%), and also eliminates the asymmetrical flares. It seems to me that Canon could have used an IR-cut filter that is more similar to the L2 bandpass, and you would have your free lunch, no?

 

@Mark, can you take a few ISO 1250, 30 second exposures on M45 on your Z6/Tak Epsilon setup for a cumulative 1 hour exposure time?  It would be interesting to compare.  The Tak Epsilon should pull in way more data than the VSD100 and Ra.  Not sure about your Bortle situation.  Maybe @Nico can do the same?  I have rain in the forecast, so that's all I can offer for comparison.

I'm not sure what that test would suggest due to different sky conditions and different processing styles. This is not a knock on your photo at all, which is very well done, but I've never found the dust around M45 a particularly difficult target, and have captured it with a variety of cameras from the Canon 60D to the Canon T3 to the Nikon D800 over the years. All that it really requires is a fast focal ratio like your VSD100 has at f/3.8, and processing the photo with the dust in mind. Here's an example of one of my photos of the region, just a 25 minute integration (48*30") with the Nikon D800 at f/1.4:

From experience, I know that photo would look roughly the same with any of the modified cameras I own if shot under the same sky with same lens, and used the same processing. I don't think there is anything special about the Ra in this regard.

[sharkmelley]

Pierre (CN member piaras) sent me a series of ISO 100 EOS Ra flats taken with an FD lens on an adaptor i.e. a non-CPU lens and therefore unrecognised by the camera firmware. For the low exposures, the results were similar to the OP i.e. at low exposures there were rings caused by channel scaling.

 

However there's a much bigger story - a set of circular steps caused by another correction being applied to the raw data.  Here's the contact sheet of the in-camera JPGs to allow you to judge the exposures:

 

 

The histogram of the image in the centre has its peak 1/4 from the left, which is fairly typical of a light frame.

 

This time I used a different method of analysis.  In PixInsight I used the raw data and processed each flat as follows:

1) Subtract the known bias level of 512

2) Debayer as Superpixel - which preserves the values in the original data

 

I then divided each flat by the top left flat (i.e. the brightest one) and binned by a factor of 8 using IntegerResample with Average as the downsample mode.  This is to increase the SNR.  Stacking 64 frames would have a similar effect.

 

Here's the result for the middle row, with a ScreenTransferFunction applied:

 

 

There's not much to see until the red channel is stretched.  Here's the PixelMath formula used:

($T[0]-mean($T[0]))*100+0.5

 

Here's the result:

 

 

Circular steps, bullseye pattern or concentric rings - however you wish to describe it.  These rings only appear in the red channel - not in the green or blue.

 

I then went back to Nico's data and did the same thing.  Sure enough, similar rings appeared:

 

 

So these circular steps appear whether or not the lens is recognised by the camera firmware.  

 

Does this have a practical consequence?  Yes, I think so.  For certain levels of exposure, either the flats or the lights (or both) will end up with a set of circular steps encoded into the red channel of the raw data.  As more and more sub-exposures are stacked, the rings will become more and more visible - it's a very similar problem to what I have on the Nikon Z6.

 

If this is actually a real problem, it will most likely manifest itself as a dark central circular area in an image or as a central circular area where the hue is obviously different to the rest of the image.  It might also be accompanied by further circular changes in brightness or hue radiating outwards.  However it is only likely to be noticeable when stacking 50, 100, 200 images or more in order to image very faint areas of dust or nebulosity.

 

Mark

 

 

 

[alan.dang]

Thanks for doing these analyses.  I have stacked hundreds of images for my shots and haven't found dramatic shifts in hue/brightness -- but it may very well be masked by my light pollution.

 

One thing is that ISO100 is below unity gain, so perhaps piaras or 2ghouls can take those same shots at ISO 800, 1250, 1600 to see if there is a difference.

 

PhotonsToPhotos estimates unity gain at ISO 417 or so for the regular EOS R.

[2ghouls]

If this is actually a real problem, it will most likely manifest itself as a dark central circular area in an image or as a central circular area where the hue is obviously different to the rest of the image.  It might also be accompanied by further circular changes in brightness or hue radiating outwards.  However it is only likely to be noticeable when stacking 50, 100, 200 images or more in order to image very faint areas of dust or nebulosity.

Thanks for doing these analyses.  I have stacked hundreds of images for my shots and haven't found dramatic shifts in hue/brightness -- but it may very well be masked by my light pollution.

Like Alan, I haven't seen this come up in processing. I just looked through my shots again. My longest integration from a dark site is this one of dust in Cepheus, about 3.5 hours from a Bortle 3. I went back to just the calibrated, registered, stacked integration, separated out the red channel, and gave it a very dramatic stretch:

 

I don't see the circular pattern (ignore the registration artifact; this was taken over two nights on a star tracker). I think if the dark bulls-eye shading was there, it would be seen in that faint dust cloud towards the center of frame. Just in case, anyone want to take a look at the stacked integration, here it is.

[agavephoto]

Interesting results. I did have problems with flat fields in my Ra that resulted in colour shifts across the frame, following a circular pattern as above. Not all the images I made showed this, and the ones that did were not fully consistent in intensity.

[alan.dang]

Like Alan, I haven't seen this come up in processing.

 

...

 

I don't see the circular pattern (ignore the registration artifact; this was taken over two nights on a star tracker).

 

What was your ISO on that specific picture?

 

I have seen a stack of ISO 200 x 60 sec shots with a lot of banding when doing an aggressive stretch.  It wasn't my shot, but I was helping someone trouble shoot their Ra results and I suggested bumping up to unity gain and even going a bit higher given Canon's low-ISO banding.

[2ghouls]

What was your ISO on that specific picture?

ISO1600. I haven't done any actually imaging below ISO800 for the reason you identified, the normal Canon horizontal banding seems to pose more of a problem at lower ISOs.

[sharkmelley]

Like Alan, I haven't seen this come up in processing.

I don't pretend to completely understand what is going on, so it's very difficult to predict when and where we will see this effect.  However, the number of circular steps definitely depends on the red channel data values.  The brighter the red channel, the more steps there are until they are so close together that they are difficult to distinguish.  If the red channel is too weak then there may be no steps at all.  Broadly speaking, every one stop increase in exposure will give twice as many steps.

 

I forgot to add the RawDigger histogram to my earlier analysis.

 

Here is that middle row again:

 

 

The left hand panel has four obvious steps.  Interestingly the red channel of the raw data histogram has four obvious spikes:

 

 

Each of the four spikes corresponds to each of the four steps.  Similarly the red histogram of the middle panel has two spikes and the histogram of the right-hand panel has one spike.  I guess that must be telling us something but I'm not sure what!

 

By the way, if you want to see how bad it can be on the Nikon Z6, take a look here:

https://www.cloudyni...ting/?p=9610122

and here:

https://www.cloudyni...ing/?p=10158668

 

Mark

Edited by sharkmelley, 15 December 2020 - 04:37 PM.

[sharkmelley]

Now some good news.  Pierre sent me a new set of flats, this time using ISO 800 instead of ISO 100.  This time I couldn't find any hint of the rings.

 

This probably certainly explains why Alan and Nico have not hit this issue in practice.

 

By the way, the bias level at ISO 800 is 2048 which is different to the bias of 512 at ISO 100.

 

Mark

[2ghouls]

Now some good news.  Pierre sent me a new set of flats, this time using ISO 800 instead of ISO 100.  This time I couldn't find any hint of the rings.

 

This probably certainly explains why Alan and Nico have not hit this issue in practice.

 

By the way, the bias level at ISO 800 is 2048 which is different to the bias of 512 at ISO 100.

Thanks Mark for continued analysis. How do you calculate the bias level from the flats?

[sharkmelley]

How do you calculate the bias level from the flats?

RawDigger gave me the bias level from the EXIF.

 

Mark

[piaras]

Thanks to Nico and Mark for their work on helping to decipher this camera’s idiosyncrasies.

Pierre