Jump to content

  •  

CNers have asked about a donation box for Cloudy Nights over the years, so here you go. Donation is not required by any means, so please enjoy your stay.

Photo

Sony A7S - Star Eater Algorithm?

  • Please log in to reply
22 replies to this topic

#1 sharkmelley

sharkmelley

    Apollo

  • -----
  • topic starter
  • Posts: 1226
  • Joined: 19 Feb 2013

Posted 24 April 2015 - 06:41 PM

A discussion has started on another thread regarding the probable existence of a hot pixel suppression algorithm within the Sony A7S:

http://www.cloudynig...lpha/?p=6553412

 

I now think this subject is important enough to have its own thread. 

 

I have done some further experimentation that throws more light on this "star eating" behaviour.  The evidence being seen in raw files was that bright pixels would never appear singly but always paired with an adjacent pixel (in the Bayer matrix) with the same value and same colour.  The same behaviour for dim pixels - they never appear alone but always paired with an adjacent pixel (in the Bayer matrix) with the same value and same colour.  This pairing of identical value pixels is very reminiscent of the infamous Nikon "star eater" behaviour.  Broadly speaking, the algorithm appears to reduce the value of any bright pixel down to the maximum value of its immediate neighbours of the same colour

 

The experiments indicate that this hot pixel suppression algorithm is only applied in "Bulb" mode.  The exposure length seems immaterial - I have seen it from 1sec up to 300sec.  I haven't yet checked if all ISOs are affected - I was using ISO 2000. 

 

Here are some test shots taken of an artificial star with a 35mm lens at F8.  The images show the raw data before the Bayer matrix colour conversion has been applied.

 

Manual mode with 30sec exposure at ISO 2000:

manual200iso30sec.jpg

 

 

Bulb mode with 30sec exposure at ISO 2000:

bulb200iso30sec.jpg

 

 

Bulb mode with 30sec exposure at ISO 2000 with LENR (long exposure noise reduction) switched on:

bulb200iso30sec_lenr.jpg

 

You can see that the two images using bulb mode have the bright star centre completely punched out. 

 

Using diffraction limited optics at a faster F-ratio then the star would be tighter and would almost certainly be destroyed entirely.

 

This is such a disappointing result for a camera that has incredible sensitivity and performance for astro-imaging.

 

Mark


Edited by sharkmelley, 24 April 2015 - 07:01 PM.

  • ccs_hello likes this

#2 AnakChan

AnakChan

    Messenger

  • -----
  • Posts: 493
  • Joined: 01 Sep 2014

Posted 24 April 2015 - 07:06 PM

Good finding Mark. Out of curiosity, which raw converter was used? Adobe Raw, Sony's Raw Viewer, etc?



#3 sharkmelley

sharkmelley

    Apollo

  • -----
  • topic starter
  • Posts: 1226
  • Joined: 19 Feb 2013

Posted 24 April 2015 - 07:16 PM

I used IRIS to load the raw files, which uses Dave Coffin's DCRaw. 

 

Mark



#4 bwallan

bwallan

    Viking 1

  • -----
  • Posts: 528
  • Joined: 08 Jul 2010
  • Loc: Alberta, Canada

Posted 24 April 2015 - 11:46 PM

A discussion has started on another thread regarding the probable existence of a hot pixel suppression algorithm within the Sony A7S:

http://www.cloudynig...lpha/?p=6553412

 

I now think this subject is important enough to have its own thread. 

 

I have done some further experimentation that throws more light on this "star eating" behaviour.  The evidence being seen in raw files was that bright pixels would never appear singly but always paired with an adjacent pixel (in the Bayer matrix) with the same value and same colour.  The same behaviour for dim pixels - they never appear alone but always paired with an adjacent pixel (in the Bayer matrix) with the same value and same colour.  This pairing of identical value pixels is very reminiscent of the infamous Nikon "star eater" behaviour.  Broadly speaking, the algorithm appears to reduce the value of any bright pixel down to the maximum value of its immediate neighbours of the same colour

 

The experiments indicate that this hot pixel suppression algorithm is only applied in "Bulb" mode.  The exposure length seems immaterial - I have seen it from 1sec up to 300sec.  I haven't yet checked if all ISOs are affected - I was using ISO 2000. 

 

Here are some test shots taken of an artificial star with a 35mm lens at F8.  The images show the raw data before the Bayer matrix colour conversion has been applied.

 

Manual mode with 30sec exposure at ISO 2000:

attachicon.gifmanual200iso30sec.jpg

 

 

Bulb mode with 30sec exposure at ISO 2000:

attachicon.gifbulb200iso30sec.jpg

 

 

Bulb mode with 30sec exposure at ISO 2000 with LENR (long exposure noise reduction) switched on:

attachicon.gifbulb200iso30sec_lenr.jpg

 

You can see that the two images using bulb mode have the bright star centre completely punched out. 

 

Using diffraction limited optics at a faster F-ratio then the star would be tighter and would almost certainly be destroyed entirely.

 

This is such a disappointing result for a camera that has incredible sensitivity and performance for astro-imaging.

 

Mark

Has anyone processed your files in PixInsight?

 

bwa



#5 sharkmelley

sharkmelley

    Apollo

  • -----
  • topic starter
  • Posts: 1226
  • Joined: 19 Feb 2013

Posted 25 April 2015 - 01:48 AM

Has anyone processed your files in PixInsight?

 

bwa

 

 

I opened the above files in PI as well as in IRIS, if that's what you mean.  The files appear the same in both.

 

One other point I should have clarified is that raw (and jpg) files images shot with Bulb mode switched off appeared quite normal i.e. the raw files had the usual sprinkling of both bright and dark pixels with no twinned neighbours.

 

Mark


Edited by sharkmelley, 25 April 2015 - 01:59 AM.


#6 tazer

tazer

    Surveyor 1

  • -----
  • Posts: 1874
  • Joined: 22 Dec 2011
  • Loc: The Pitts (NC)

Posted 25 April 2015 - 07:29 AM

I'd be curious to see how it affects "larger" stars. My FWHM is typically around 4-5", so it's quite a bit larger than the artificial star you're experimenting with. I wouldn't think it would affect extended objects though.

 

Mark



#7 sharkmelley

sharkmelley

    Apollo

  • -----
  • topic starter
  • Posts: 1226
  • Joined: 19 Feb 2013

Posted 25 April 2015 - 07:52 AM

The only effect on larger stars would be that the central peak would be slightly truncated - truncated down to the values of immediate neighbours of the same colour.   Extended objects should be fine.

 

Mark



#8 bwallan

bwallan

    Viking 1

  • -----
  • Posts: 528
  • Joined: 08 Jul 2010
  • Loc: Alberta, Canada

Posted 25 April 2015 - 02:06 PM

Is there something I'm missing in this discussion?  Some particular camera setting that causes this, other than supposedly Bulb Mode?  I've spent a fair bit of time pixel peeping my images shot in Bulb mode and can't find the problem:California%2BNebula%2B%28NGC%2B1499%29%2

California Nebula (NGC 1499), Astro-Tech 65EDQ, Hutech HEUIB-II filter, full spectrum Sony A7S, Celestron AVX mount, 30x60 sec (Bulb Mode) @ ISO 3200, preprocessed PixInsight,

postprocessed in Lightroom, all in-camera noise reduction turned off, captured and processed from RAW subs, in-camera hot pixel map regenerated whenever hot pixels become obvious.

 

I'm interested in finding out the parameters I have to change/set to generate this apparent problem so I don't accidently set them at some point in the future.

 

Could the problem be originating from the compressed nature of Sony A7 series RAW's?  Is there a difference in firmware between cameras with and without the problem?  I'm running firmware Ver. 1.10.  Do the cameras showing the problem have a lot of hot pixels, i.e.: bad pixel mapping hasn't been done recently?

 

Interesting discussion but I can't find the problem.

 

bwa



#9 sharkmelley

sharkmelley

    Apollo

  • -----
  • topic starter
  • Posts: 1226
  • Joined: 19 Feb 2013

Posted 25 April 2015 - 08:03 PM

Hi bwa, 

 

Thanks for your interest.  You raise some interesting points and I don't have complete answers to them at this stage.  My A7S is also running Ver 1.10 so I don't think this is the problem.  I am seeing the issue whenever Bulb mode is switched on without any other settings being changed.  This strongly indicates to me that Bulb mode is the culprit but I can't 100% rule out some other obscure camera setting that only activates when Bulb mode is used.  My camera is 2 weeks old so the bad pixel mapping would be pretty recent - I have very few bad pixels showing in non-Bulb modes.  It will be very difficult to see the effects of this supposed hot pixel suppression algorithm in a final stacked image like the one you posted (except for the absence and attenuation of faint stars - which may not be all that obvious).

 

Instead, it is easiest to see the effect in a long exposure dark frame opened in RawDigger, PixInsight, or IRIS without Bayer conversion. 

Here's an example:

A7S_raw_pixel_pairing.JPG

 

It is the the lack of single bright pixels is the important feature here - instead they always appear in pairs, most likely because a Nikon style hot pixel suppression algorithm has been applied.

 

However, this "pixel pairing" is not an issue in itself.  My issue is that some algorithm is being applied to my data that punches out the bright peaks of stars, turning them into rings - destroying valuable data in the process.  If my hypothesis is correct then the "pixel pairing" is simply another symptom of this same algorithm being applied across the whole image.  Especially since we have seen this same effect previously, in Nikon cameras.

 

Mark


Edited by sharkmelley, 25 April 2015 - 08:08 PM.


#10 bwallan

bwallan

    Viking 1

  • -----
  • Posts: 528
  • Joined: 08 Jul 2010
  • Loc: Alberta, Canada

Posted 25 April 2015 - 10:43 PM

Hi bwa, 

 

Thanks for your interest.  You raise some interesting points and I don't have complete answers to them at this stage.  My A7S is also running Ver 1.10 so I don't think this is the problem.  I am seeing the issue whenever Bulb mode is switched on without any other settings being changed.  This strongly indicates to me that Bulb mode is the culprit but I can't 100% rule out some other obscure camera setting that only activates when Bulb mode is used.  My camera is 2 weeks old so the bad pixel mapping would be pretty recent - I have very few bad pixels showing in non-Bulb modes.  It will be very difficult to see the effects of this supposed hot pixel suppression algorithm in a final stacked image like the one you posted (except for the absence and attenuation of faint stars - which may not be all that obvious).

 

Instead, it is easiest to see the effect in a long exposure dark frame opened in RawDigger, PixInsight, or IRIS without Bayer conversion. 

Here's an example:

attachicon.gifA7S_raw_pixel_pairing.JPG

 

It is the the lack of single bright pixels is the important feature here - instead they always appear in pairs, most likely because a Nikon style hot pixel suppression algorithm has been applied.

 

However, this "pixel pairing" is not an issue in itself.  My issue is that some algorithm is being applied to my data that punches out the bright peaks of stars, turning them into rings - destroying valuable data in the process.  If my hypothesis is correct then the "pixel pairing" is simply another symptom of this same algorithm being applied across the whole image.  Especially since we have seen this same effect previously, in Nikon cameras.

 

Mark

Please see: http://www.cloudynig...lpha/?p=6559181

 

bwa



#11 sharkmelley

sharkmelley

    Apollo

  • -----
  • topic starter
  • Posts: 1226
  • Joined: 19 Feb 2013

Posted 27 April 2015 - 07:39 PM

I tried the A7S on my Tak Epsilon tonight (focal length 500mm).  It was a typical cloudy night, so I placed an artificial star at the end of the garden.  The results are pretty much similar to the example in the first post.

 

In both examples below the left hand star is non-Bulb mode and the right hand star is Bulb mode with the same exposure length.

 

10 sec ISO 125:

A7S_StarEater1.jpg

 

 

2 sec ISO 2000:

A7S_StarEater2.jpg

 

I can't totally rule out a sub-pixel scope movement between the exposures but the loss of the bright pixels is quite dramatic.  These are two representative examples of a very consistent pattern.

 

To be fair, I should point out that on a longer focal length scope, where the star is adequately sampled (say 1000mm or more), this would not be such an issue - the central peak of the star would be flattened but the main body of the star would remain more or less intact.

 

Mark


Edited by sharkmelley, 27 April 2015 - 07:41 PM.


#12 Tonk

Tonk

    Cosmos

  • *****
  • Posts: 8383
  • Joined: 19 Aug 2004
  • Loc: Leeds, UK, 54N

Posted 28 April 2015 - 10:18 AM

the central peak of the star would be flattened

 

The examples you show aren't flattened as such but actually have a pit dug in them. Would adequately sampled stars also have a brightness "pit" in the centre as opposed to merely a truncated top?



#13 mmalik

mmalik

    DSLR camera modifications

  • *****
  • Vendors
  • Posts: 9061
  • Joined: 13 Jan 2012
  • Loc: USA

Posted 28 April 2015 - 10:46 AM

Borrowing medical terminology :), this is a good in vitro test. Next step would be to perform an in vivo test, i.e., test of both very small and bit larger "actual" stars. Regards


  • bwallan likes this

#14 sharkmelley

sharkmelley

    Apollo

  • -----
  • topic starter
  • Posts: 1226
  • Joined: 19 Feb 2013

Posted 28 April 2015 - 02:16 PM

 

The examples you show aren't flattened as such but actually have a pit dug in them. Would adequately sampled stars also have a brightness "pit" in the centre as opposed to merely a truncated top?

 

 

That's a great question and I actually know the answer. Depending on the size of the star in the image, the top of the star is depressed and has a crater ridge.  Here are a couple of typical star profiles from a stacked image of 80 exposures I did at a f/l of 2200mm:

 

A7S_starprofiles_2200mm.JPG

 

 

[Later edit - ignore the comment about the depression.  The odd effect in the shape graphs is caused by an elongation of the star in one direction.  The stars at a f/l of 2200mm actually look perfectly normal. ]

 

Mark


Edited by sharkmelley, 28 April 2015 - 05:43 PM.


#15 sharkmelley

sharkmelley

    Apollo

  • -----
  • topic starter
  • Posts: 1226
  • Joined: 19 Feb 2013

Posted 30 April 2015 - 12:30 AM

Repeating a post made on the DPReview forum, here are the results of an experiment to determine by experiment how destructive (or not) the star eating algorithm really is in practice

 

All the evidence I've seen so far, points to the fact that the Sony A7S Bulb mode algorithm works in the same way as the original Nikon algorithm as follows:

StarEaterAlgo.JPG
 

For every pixel in the image, replace the original V0 value by: Min(V0, Max(V1, V2, ... , V8) )

 

It is this that removes single hot pixels and leads to "pixel pairing". Note that V1, ... , V8 are not immediate neighbours of V0 but are neighbours of the same Bayer matrix colour.

 

Now I already have some reasonably good data shot with a modified Canon 350D on a Tak Epsilon 180ED. Dithering was performed between each exposure during acquisition. So I stacked 36 exposures of 5min using darks, flats and bias frames as normal.

 

Then I took the same raw data and applied the above star eater algorithm to the raw lights and raw darks and performed a complete re-process. Looking at the results, it was impossible to tell the stacked images apart without "blinking" one against the other.

 

Here is a 1:1 scale crop of the original with a log intensity scale (to allow the full dynamic range to be displayed):

pleiades_logscale.jpg

 

Here is the "star eater" version:

pleiades_stareater_logscale.jpg

 

Only by blinking the two images can you detect that all the non-saturated stars have been dimmed down.

 

Here's the difference between the two - again log scaled:

pleiades_difference_logscale.jpg

 

This frame shows what data has been "removed" by the algorithm and makes it far more obvious that most stars have had their brightness reduced.

 

Dividing one image by the other, it also became clear that a great many (non-saturated) stars have had their central peak reduced by a factors of 1.5x - 2.5x

In the "star eater" image every star still had a central peak though in many cases it was severely flattened. I found no cases of an inverted peak (i.e. a central depresssion).

 

So, in the end, does the star-eating behaviour matter? It really depends on your point of view. If your main interest is producing images of the night sky then I really don't believe you will notice the problem (which is probably why it has gone undetected until now). If you have a more scientific interest in star profiles, photometry, variable stars etc. then any camera that deliberately destroys data is clearly not the ideal camera to use, especially on an undersampled scope such as the one used here.

 

I ought to re-iterate that on a longer focal length scope then the "star eating" behaviour only leads to a minor reshaping of the star profile and is almost completely undetectable.

 

Mark


Edited by sharkmelley, 30 April 2015 - 12:34 AM.


#16 mmalik

mmalik

    DSLR camera modifications

  • *****
  • Vendors
  • Posts: 9061
  • Joined: 13 Jan 2012
  • Loc: USA

Posted 30 April 2015 - 10:54 AM

Good analysis Mark!

 

 

I am wondering if someone could share a Nikon dark (RAW as well as corresponding in-camera large format JPG) to include Nikon in this study as well; my preference would be D810 data since it would have ramifications for upcoming D810a. If yes, please take a 5-10 min dark at ISO 3200 or so and share both RAW and the JPG. Thanks in advance. Regards



#17 DuncanM

DuncanM

    Surveyor 1

  • -----
  • Posts: 1550
  • Joined: 03 Nov 2009
  • Loc: Arizona Sky Village or the rain forest

Posted 30 April 2015 - 02:21 PM

It's quite possible that the 'star eater' algorithm is ISO dependant. I would suggest trying low ISO exposures and examining the result. Hot pixel removal may be part and parcel of high ISO noise reduction.


Edited by DuncanM, 30 April 2015 - 02:23 PM.


#18 sharkmelley

sharkmelley

    Apollo

  • -----
  • topic starter
  • Posts: 1226
  • Joined: 19 Feb 2013

Posted 30 April 2015 - 03:53 PM

It appears to happen at all ISOs when using Bulb mode.  High ISO noise reduction is set to "Off" in the camera menu.

 

Mark



#19 calypsob

calypsob

    Vanguard

  • *****
  • Posts: 2430
  • Joined: 20 Apr 2013
  • Loc: Roanoke, Virginia

Posted 03 May 2015 - 09:58 AM

Email this report to Sony Mark. I am sure they would be interested



#20 AnakChan

AnakChan

    Messenger

  • -----
  • Posts: 493
  • Joined: 01 Sep 2014

Posted 24 May 2015 - 02:26 AM

Sorry if I missed this but Sony released v1.20 in late March. Have you tried to see if this is a problem with 1.20 too??

#21 sharkmelley

sharkmelley

    Apollo

  • -----
  • topic starter
  • Posts: 1226
  • Joined: 19 Feb 2013

Posted 24 May 2015 - 11:40 AM

I haven't tried version 1.20 (I'm using 1.10) but nothing in the 1.20 release notes suggests any change to this functionality. 

 

Moreover, since it affects a range of cameras (not just the A7S) I suspect it is deliberate functionality and not something Sony would consider to be a bug.  For a consumer camera I can certainly see why they might want to do it - it makes long bulb exposures appear less noisy. 

 

Mark



#22 Denimsky

Denimsky

    Apollo

  • *****
  • Posts: 1007
  • Joined: 21 Jan 2007
  • Loc: BC, Canada

Posted 24 May 2015 - 11:50 AM

An option to turn off this behaviour will be nice.

#23 ccs_hello

ccs_hello

    Cosmos

  • *****
  • Posts: 8775
  • Joined: 03 Jul 2004

Posted 24 May 2015 - 11:57 AM

All camera mfgs have the habit of doing something different.  Some we like and some features we really wish they go away.

Unfortunately, it is the business side of the camera mfgs that made the call.

After all, they sell cameras for a living, not image sensors.

 

Clear Skies!

 

ccs_hello




CNers have asked about a donation box for Cloudy Nights over the years, so here you go. Donation is not required by any means, so please enjoy your stay.


Recent Topics






Cloudy Nights LLC
Cloudy Nights Sponsor: Astronomics