Star eater fix - what is the new algorithm?

sharkmelley · Jun 9, 2017

There are mixed reports coming out regarding the success or otherwise of the star eater fix in the latest firmware for the A7Sii and A7Rii:

http://community.sony.com/t5/Alpha-...-please-offer-alternative/td-p/619350/page/16

http://www.fredmiranda.com/forum/topic/1488852/8

The reports possibly indicate that Sony is now using a more sophisticated spatial filtering algorithm - one that removes single hot pixels but is less destructive to stars (which generally occupy more than one pixel). Nikon uses a similar approach to the same problem of reducing the noisy pixels in long exposures. By contrast, the infamous star eater algorithm would totally delete any star fully contained within a 2x2 area of pixels in the undebayered raw file.

Unfortunately, my camera is the Sony A7S for which there is (currently) no update. However, I'm still very interested in attempting to understand and/or reverse engineer the new algorithm. If anyone is happy to make available one or more of the following raw dark frames (i.e exposures taken with the lens cap in place) from an updated A7Rii or A7Sii I would love to take a look:

4 sec ISO 1600
bulb mode 1600 (4 seconds or more)
4 sec ISO 6400
bulb mode 6400 (4 seconds or more)

LENR should be switched OFF.

Regards,

Mark

PVCdroid · Jun 9, 2017

This FM post was revealing:

"I just did a re-test after fully initialising the camera. I still have star eater issues at 4.0s+. Spoke to a friend at Sony, he doesn't think this FW was meant to be star eater fix."

I think there was likely improvement with Long Exposure NR but it is likely not what some want. Jim K has said he will do some new long exposure tests with firmware 4.0 on the A7Rii in the near future to see if it has improved.

I do know is that when I'm shooting stars at night, the results are very good and I get what I want when post processing. You would think the stars wouldn't even show in your images >3.2 sec from some of the banter on this subject.

Using Sony cameras for over four years now, it is common for Sony to include "improves Long Expsure NR" in their firmware release descriptions. Maybe the next one will satisfy those that are still upset.

stadlereric · Jun 9, 2017

From what I've read the A7/A7R/A7S are only effected by the star eater issue when using bulb mode. If you set the exposure 4-30 seconds in camera the spatial filter issue is nonexistent. I could be wrong. I'm interested to see the results of your testing. Look it up though you may be out of the woods with a gen 1 A7S.

I'm going to shoot the milky way next weekend (only third attempt at Astro) with a "new to me" Sony Zeiss 24mm F2 lens that I haven't received yet. I'm going to do my own testing then. I'm already thinking I should have gotten the batis 25mm.

I have a couple average at best shots of the milky way on flickr if you'd like to see them. Link below

stadlereric · Jun 9, 2017

I should read posts more carefully before posting. I didn't realize we were talking about bulb mode.

Navman · Jun 9, 2017

sharkmelley said:
There are mixed reports coming out regarding the success or otherwise of the star eater fix in the latest firmware for the A7Sii and A7Rii:

http://community.sony.com/t5/Alpha-...-please-offer-alternative/td-p/619350/page/16

http://www.fredmiranda.com/forum/topic/1488852/8

The reports possibly indicate that Sony is now using a more sophisticated spatial filtering algorithm - one that removes single hot pixels but is less destructive to stars (which generally occupy more than one pixel). Nikon uses a similar approach to the same problem of reducing the noisy pixels in long exposures. By contrast, the infamous star eater algorithm would totally delete any star fully contained within a 2x2 area of pixels in the undebayered raw file.

Unfortunately, my camera is the Sony A7S for which there is (currently) no update. However, I'm still very interested in attempting to understand and/or reverse engineer the new algorithm. If anyone is happy to make available one or more of the following raw dark frames (i.e exposures taken with the lens cap in place) from an updated A7Rii or A7Sii I would love to take a look:

4 sec ISO 1600

bulb mode 1600 (4 seconds or more)

4 sec ISO 6400

bulb mode 6400 (4 seconds or more)

LENR should be switched OFF.

Regards,

Mark

Happy to do so when back home tomorrow night, Mark - msg me your email and will send via Dropbox.

- John

sharkmelley · Jun 10, 2017

PVCdroid said:
This FM post was revealing:

"I just did a re-test after fully initialising the camera. I still have star eater issues at 4.0s+. Spoke to a friend at Sony, he doesn't think this FW was meant to be star eater fix."

I think there was likely improvement with Long Exposure NR but it is likely not what some want. Jim K has said he will do some new long exposure tests with firmware 4.0 on the A7Rii in the near future to see if it has improved.

No-one seems to have spotted the fact that the description of the firmware update is different on the 2 cameras:

The firmware update (FW 4.00) for the A7Rii says: "Improves image quality when Long Exposure NR setting is OFF"

The firmware update (FW 3.00) for the A7Sii says: "Improves image quality when Long Exposure NR setting is ON"

In the case of the A7Sii this might be a typo. On the other hand, it is entirely possible that the recent updates do something entirely different than addressing the "star eater" issue. It could explain the mixed reports being obtained by users.

Mark

Tristimulus · Jun 10, 2017

Thank you for bringing up this topic.

In my case I would like to use Sony cameras for all my photography (do now use another brand for night sky photography and astrophotography).

Cling to my A7 cameras for now, and refuse to upgrade to another Sony camera until the star eater issue is solved once for all. In my case the 1.8/55mm lens, my Zeiss ZE lenses, and the Takahashi 106/530mm astrograph is severely affected by the star eater issue.

So following this space closely.

Ari Aikomus · Jun 10, 2017

Here is my "home made" star sky ;-)

With A7RII, RAW, Adobe camera raw, no NR:

3.2sec:

4sec:

200% crops (click original to see full size samples):

...I think star eater NR isn't so strong anymore.

Ari

--
- Ari Aikomus -
'Why should I feel lonely ? is not our planet in the Milky way?'

sharkmelley · Jun 10, 2017

Nice experiment!

Lining up the 2 frames in Photoshop and subtracting the 4sec from the 3.2sec gives the following:

Note the red and blue star remnants. It might indicate that the Red and Blue channels have been more heavily processed than the Green.

Mark

fred 76 · Jun 10, 2017

Ari, could you please test in Bulb mode ?

It should also be worth a try with a smaller focal length to get even smaller stars. The star eater was indeed very hungry of dot-like stars.

Fred

Ari Aikomus · Jun 10, 2017

fred 76 said:
Ari, could you please test in Bulb mode ?

It should also be worth a try with a smaller focal length to get even smaller stars. The star eater was indeed very hungry of dot-like stars.

Fred

This my test is with Samyang 12/2.8 fish eye....

I made also BULB test, but results is similar than with 4sec. That means= some improvement indeed, vs. v. 3.30 (because not much diffrences if compare to 3.2sec exposure).

Ari

daran · Jun 10, 2017

Ari Aikomus said:
fred 76 said:

Ari, could you please test in Bulb mode ?

It should also be worth a try with a smaller focal length to get even smaller stars. The star eater was indeed very hungry of dot-like stars.

Fred

Click to expand...

This my test is with Samyang 12/2.8 fish eye....

You could reduce star size by increasing the distance.

Ari Aikomus said:
I made also BULB test, but results is similar than with 4sec. That means= some improvement indeed, vs. v. 3.30 (because not much diffrences if compare to 3.2sec exposure).

Ari

So far I managed to record bright a single pixel "star" in the green and blue channel of a blue-ish white artifical star, even at 4s or bulb mode. This was previously impossible. The red part of the same star still got eaten though. Blue was fine at 3.2s.

Reasons could be

- red is the weakest channel for this light

- different handling per color channel

sharkmelley · Jun 10, 2017

Here's an example of an analysis of the Nikon D800 and D4 hot pixel suppression (HPS)algorithm performed by Bernard Delley:

https://www.dpreview.com/forums/post/50875417

In passing it also mentions the Nikon D7000 algorithm which is explained here:

In-camera processing of long-exposure RAW data

Have do you see one of the new NEFs from the D7000? At long exposures there are a lot of hot pixels isolated. It seems to me that something has changed in the way they process the RAW files now in-camera. Yes you get more noise as you open the NEF, but potentially more detail, because it seems...

www.dpreview.com

Previous to this, Nikon cameras used the same star eater algorithm that now afflicts Sony cameras. Back in those days, Nikon cameras were also avoided by the astro-imaging community.

Bernard explains that different colour channels have different thresholds applied for deciding to replace the actual value by a calculated substitute value.

It remains to be seen (and tested!) whether or not the new Sony algorithm for the A7Rii and A7Sii reaches this level of sophistication.

Mark

Navman · Jun 11, 2017

Ari Aikomus said:
fred 76 said:

Ari, could you please test in Bulb mode ?

It should also be worth a try with a smaller focal length to get even smaller stars. The star eater was indeed very hungry of dot-like stars.

Fred

Click to expand...

This my test is with Samyang 12/2.8 fish eye....

I made also BULB test, but results is similar than with 4sec. That means= some improvement indeed, vs. v. 3.30 (because not much diffrences if compare to 3.2sec exposure).

Ari

Ari Aikomus · Jun 11, 2017

Navman said:
Ari Aikomus said:

fred 76 said:

Ari, could you please test in Bulb mode ?

It should also be worth a try with a smaller focal length to get even smaller stars. The star eater was indeed very hungry of dot-like stars.

Fred

Click to expand...

This my test is with Samyang 12/2.8 fish eye....

I made also BULB test, but results is similar than with 4sec. That means= some improvement indeed, vs. v. 3.30 (because not much diffrences if compare to 3.2sec exposure).

Ari

--
- Ari Aikomus -
'Why should I feel lonely ? is not our planet in the Milky way?'

Click to expand...

Ari, my further testing over the weekend makes me pessimistic that Sony have delivered the improvement that we were hoping for. I need to check my results further, but I'm not feeling very positive that it will be a good outcome. I'll be back with more soon.

-John

Hi,

And we are talking now about A7R2 camera?

Do you mean BULB exposures or all kind of exposures longer than 3.2sec?

...Because I think there is "some" improvements over v.3.30 firmware, the algorithm is not so aggressive with firmware v.4.00 (> 3.2 sec and BULB)...

But I look forward to your experiments, with interest.

regards,

Ari

--
- Ari Aikomus -
'Why should I feel lonely ? is not our planet in the Milky way?'

sharkmelley · Jun 11, 2017

First of all many thanks to John (Navmann) for providing the raw files to me for analysis.

In brief, it looks like there has been no change to the algorithm applied to the Red and Blue channels - the star eater algorithm is applied in the usual way. However there has been an improvement in the Green channel - the two green channels G1 and G2 support each other in a manner that makes the algorithm less destructive than previously.

What follows is a more detailed description.

Here is a small highly magnified crop of a 6sec A7Rii exposure (firmware version 4.00) exposure of a star field, debayered without interpolation, so you can see the colours of the individual pixels:

Debayered without interpolation

Now look at the red pixels only:

Red pixels only

Notice that there are no individual bright pixels. Brighter pixels always appear in pairs because the value of any single bright pixel is reduced to the maximum value found in its 8 immediate neighbours. This is how the "star eater" spatial filtering works. The converse also happens - dark pixels appear in pairs because any single dark pixel has its value increased to the minimum value found in its 8 immediate neighbours.

Exactly the same is happening in the blue channel.

Now look at the green pixels only:

Green pixels only

Notice that we can see bright pairs of immediately adjacent diagonal pixels. This did not happen under the old algorithm. The old algorithm would take the G1 and G2 channels separately and a bright pixel would be reduced to the max of its 8 neighbours in that channel. It looks like the new algorithm is combining the G1 and G2 channels (as in the image above) and a bright pixel is now reduced to the max of its 12 neighbours. This means that a greater number of brighter green pixels survive than previously.

In the context of astrophotography, what is the effect of the new algorithm on a star field? The old algorithm would completely erase any star completely contained in a 2x2 box. Remember that a 2x2 box will always contain 2 Green pixels and 1 each of Red & Blue. With the new algorithm the R & B components will still be destroyed but as long as the star falls on both G pixels then the G component will survive. However this is just one illustrative example of the many ways a small star will fall on a group of adjacent pixels.

With the new Sony algorithm we therefore still expect to see a significant number of stars eliminated or severely attenuated. The ones that survive will be predominantly greenish in colour because those are the stars that happened to land across 2 green pixels and then had their blue and red components deleted or attenuated..

Given the short timescale that Sony had for producing a new algorithm, it does not surprise me that all they did was to make a slight modification to the existing one. A wholesale change to the algorithm would have been risky.

This is a preliminary analysis of the new spatial filtering algorithm and it is entirely possible that there are subtleties that I may have missed. So please feel free to free to correct me if I have made any errors.

Mark

VLreviews · Jun 11, 2017

This sounds plausible and also explains the green color cast on all small "stars" (dots of light) in Ari's test images. Thank you very much for your investigation!

Graham Gibson · Jun 11, 2017

Wow, excellent analysis and explanation. At least it's a good start. Sony hasn't historically been great about updating camera firmware, so this was a welcome surprise and hopefully Sony continues to develop their NR algorithm (and push it out to older cameras).

sharkmelley · Jun 11, 2017

I forgot to add that with firmware version 4.0 the A7Rii bulb mode behaves the same as manual exposures of 4seconds and greater i.e. bulb mode also uses the new spatial filtering algorithm.

Mark

VLreviews · Jun 11, 2017

sharkmelley said:
I forgot to add that with firmware version 4.0 the A7Rii bulb mode behaves the same as manual exposures of 4seconds and greater i.e. bulb mode also uses the new spatial filtering algorithm.

Mark

Do you - by any chance - know how the A7II (without R or S) behaves with firmware 3.3 or the new 4.0? I was looking around and couldn't find any info if this model ever was affected by star eater.

Star eater fix - what is the new algorithm?

Senior Member

Senior Member

Active member

Active member

Navman

Guest

Senior Member

Veteran Member

Veteran Member

Senior Member

Active member

Veteran Member

Senior Member

Senior Member

Navman

Guest

Veteran Member

Senior Member

Forum Enthusiast

Leading Member

Senior Member

Forum Enthusiast

Similar threads

Keyboard shortcuts