Binning v pixel size

[traveller]

Not sure if this has been discussed before, but here goes.
Just want to test my line of thinking:
With larger pixel cameras, the sensitivity is increased with decreased resolution.
Smaller pixels ccd cameras with binning can achieve the same effect.
Therefore larger pixel count ccd with smaller pixel sizes with binning can achieve similar resolution as smaller pixel cameras with larger pixel sizes.
Hope that made sense.
:shrug::help::question::thanx:
Bo

[Merlin66]

More data required....
a 6 micron pixel, 2 x 2 will be the equivalent of the 12 micron pixel.
Comparing that to a "native" 12 micron pixel camera, would require knowledge of QE etc.

[gregbradley]

Yes, that is my experience. It does depend on the CCD as KAF8300 does not do a clean binning and there is some loss. But Sony ICX694 bins cleanly and 2x2 binning does seem to emulate a 9 micron pixel very well only cleaner and more sensitive.

Greg.

[uwahl]

Hi Greg

You caught my attention with your comments re clean binning. Could you amplify a little? I am presuming there is some loss due to the architecture of the chip but is there any data on the size of the loss?:question:

[uwahl]

Hi Greg

You caught my attention with your comments re clean binning. Could you amplify a little? I am presuming there is some loss due to the architecture of the chip but is there any data on the size of the loss? :question:

[Bassnut]

Larger native pixel sites on the chip have a larger well depth than small pixels. I dont think binning small pixels increases well depth.

[Shiraz]

well depth is 4x with 2x2 binning, but the transfer and output electronics usually cannot handle that much signal - I think that the makers generally turn down the gain by about 0.5 to protect the output stages, so that you get a compromise of up to about 2x the effective well depth - seems that it varies with the chip design. Still quite usable if the read noise is low - but sometimes the read noise also suffers with binning. Overall, binning is a useful extra facility, but not quite the same as a "real" larger pixel. As Greg points out, the icx694 comes pretty close though.

Software binning is possibly a better option overall - if you expose the 1x1 subs for long enough to bury the read noise you get all the advantages (and disadvantages) of bigger pixels - 4x the sensitivity, 4x the well depth and 1/2 the linear resolution.

[PRejto]

Ray,

If you don't mind, how is software binning done? I assume it's not all all the same as reducing the image size in Photoshop.

Thanks,

Peter

[Shiraz]

Hi Peter.

I use software binning in Nebulosity - presume that it is just averaging data from 4 pixels to get one new pixel (the pixel noise stats indicate that this is what it does). If the read noise is low enough, this is functionally the same as you would get from a perfect 2x2 binning process, only with signal and noise scaled by 0.25.

Reducing size could be done the same way - but it could also be done by just taking one out of the four pixels to represent the new pixel data, or using a surface fitting process - would need to know how Photoshop does it under the hood.

[traveller]

Thanks for all the informed comments.
Crappy weather makes time for more education...
Bo

[gregbradley]

What I meant by clean binning is theoretically 2x2 binning should be 4X the signal to noise. On KAF8300 because of something to do with the horizontal registers, see here:

http://www.ccd.com/ccd103.html

What I have read is that with the KAF8300 chip there is some issue with the horizontal register such that it does not do a proper 4X addition of the 4 pixels that make up 2x2 binning. So it loses something in the process. You can this when you use a KAF8300 camera in that 2x2 binning gives a gain but not the 4X you might expect more like 2X.

On the other hand a chip that does 2x2 binning very well like the Sony ICX694 you do see pretty much a 4X improvement and with its small pixels it emulates a 9 micron pixelled Kodak chip quite well but with increased sensitivity and lower noise. The downside is its a small chip and has a narrow FOV which is useful for some shots (galaxies).

Ray I have only experimented briefly with software 2x2 binning and noticed it also reduces resolution noticeably. So no free lunch there.

Greg.

[Merlin66]

Greg,
I use 2 x2 binning when chasing faint spectra....
There's no loss of resolution as long as the binned pixel size still meets the Nyquist requirements.

[Shiraz]

+1

the resolution loss with software binning is exactly the same as with hardware binning - 1/2 the linear resolution when you use 2x2 binning. The free lunch of software binning is that you get more dynamic range.

minor point Greg, but perfect binning gives you 4x the signal, but only 2x the SNR if you are shot noise limited (which is usually the case).

[uwahl]

Hi Greg

After your comment I fired off an e mail to QSI (I have a QSI 683). Their reply:

Ulrich,

Kodak doesn't publish full well capacity figures for the HSR and Output Register on the KAF-8300, but we have been able to determine effective values through Photon Transfer analysis. Well depth of the HSR is about 37,000 e-. Well depth of the OR is about 55,000e-. Low gain on the 683 is set to 1.1e-/ADU to maximize the dynamic range of the Output Register.

Take a look at this Knowledge Base article for a bit more detail about how this relates to Flats.

http://store.qsimaging.com/kb_results.asp?ID=28

Regards,
Kevin

Essentially while 2 X 2 binning is able to collect 100,00 photons only 55,000 can be read through the output register.

Ulrich

[gregbradley]

Thanks for that. That sounds about right.

Greg.

[traveller]

Ok, more questions,
I looked up Celestron Nightscape 8300 which uses the KAF 8300 sensor and here is what I lifted from their website

"Full Well Capacity 25,500 e-
Imaging Sensor Kodak KAF-8300 Color Sensor
Mounting 2" barrel and t-thread
Operating Environment 40C° to -40C° (104F to -40F)
Optical Window High transmission Schott B270 glass
Quantum Efficiency 33%@630 nm; 40%@550 nm; 33%@470 nm
Read Noise (RMS) 8 e-"

Am I missing something here :help:
Bo

[Merlin66]

Is this a 16bit camera???

[traveller]

Yes, 16 bit A/D conversion.
Full specs on the website http://www.celestron.com/browse-shop/astronomy/astroimaging-cameras/nightscape-8300-ccd-camera
Click on the specs and expand.
Bo

[Shiraz]

that's entirely consistent with earlier posts.

Binning occurs in 2 steps.

1. First, the signal from pairs of adjacent pixel is added by putting the charge from 2 pixels (rather than one) into each cell of the shift register - for a saturated 8300, this results in trying to stuff 51,000 electrons in each SR cell. But, the shift register cells cannot hold that much and saturate at 37,000 electrons - so you lose ~1/3 the available full charge in this step.
2. The (37,000 electron) charge packets in the SR cells are then transferred into the output register in pairs, so 74,000 electrons turn up at the output register for each binned pixel. The output register cannot handle more than 55,000 electrons, so some of the available electrons are again lost to saturation and you are again down by about 1/3.

ie, in the two processes, you lose about half of the total number of electrons available from the 4 saturated pixels - you start out with 25,500x4 and end up with ~55,000, simply because you cannot squeeze all of the available electrons from 4 saturated pixels through the transfer electronics.

The final step is that the gain of the output stage is normally set to give 64000 ADU with only one pixel-full of electrons (25,500). With binning, it is presented with a bit more than double that many electrons per pixel, so the gain must be reduced by a roughly a half to keep the maximum reading to 64,000 ADU.

[traveller]

Thanks Ray,
I had to read that 2-3 times, but I think I got it.
:D
Bo

[rally]

Ken,

The A/D like many cameras may be 16 bits but the camera itself is really only capable of producing 11.6 bits of information because of the well depth and read noise.

Log2 (WellDepth e- / ReadNoise e-)

Although that is then artificially increased by the non linear antiblooming capability of the CCD circuit in terms of reproducing the dynamic range of the original target, the camera is still only giving you 11.6 bits of useable data.

Unfortunately some camera makers like to "oversell" to their consumers by quoting the A/D circuit capability ! I am being kind here !!

True 16 bits would be 64,000 e- well depth and 1 e- read noise or 256,000 e- and 4 e- read noise !

Rally

[Fizics]

As I understand it is important to have a good match of pixel size, focal length, and seeing conditions to optimize sensitivity and resolution. So it all comes down to arcseconds per pixal. Reference this link http://www.stanmooreastro.com/pixel_size.htm
and conveniently he has a link to calculate arc secs per pixal http://www.stanmooreastro.com/Pixel_size_calc.html

John

[rat156]

Hi Ray,

Great explanation. How does this effect a non-saturated signal, say one at about 5,000? Would you still get the full 20,000 if you binned?

The reason I ask is that you guys turned me away from taking my RGB at 2x2 binned because you argued (quite convincingly) that I was better off getting RGB at 1x1 on a KAF8300. But if the argument is only for saturated signals, really it doesn't matter to me as these are usually just the stars, which I would tend to take shorter subs for and overlay the saturated ones in PS.

And, yes, cloudy days are great for expanding you knowledge rather than actually doing any astronomy.

Cheers
Stuart

[Shiraz]

very well summarised Rally - not widely accepted though, which is a pity.



Hi Stuart. with wells of 5,000 ADU, the binned signal would be 20,000 ADU, but that would then be scaled by about 0.5 (along with the noise). Net result is better SNR from binning. If you use HDR methods, the loss of dynamic range on stars is not a problem.

However, apart from the loss of dynamic range, I understand that binning with the 8300 is problematical for 2 reasons:
1. the read noise increases for binned operation (probably not big issue, but requires longer subs than expected)
2. the saturation in the registers somewhere along the line result in one sided blooming on brighter stars (I have no experience with this chip, but the effect is widely documented). Presumably there are parts of the chip that are not protected by ABG - they would not normally need to be, but are pushed into saturation by the higher net signal.

In summary, imaging at 1x1 with subs that are long enough to bury read noise allows full resolution for targets with small bright features that might benefit from high res. If you decide that you don't have enough colour data at 1x1, you could alternatively use software binning on this data to get the full binning advantages of 4x dynamic range, 2x SNR and no artefacts. So I guess that the basic question is why bother with on-chip binning.

Clouds?*++%. I have a new mount sitting here and an AO to try out - naaah not going to happen any time soon.

[RickS]

Only if there is a significant read noise advantage? Not the case with many of the current Kodak/TrueSense sensors.

[uwahl]

Noise is only read once if you on chip bin whereas software binning results in 4X noise if 2X2 binning - the software reads the noise from all 4 pixels. Better to on chip bin.

[Shiraz]

read noise adds in quadrature, so 2x2 binning in software only results in 2x the read noise, not 4x. In addition, the 8300 produces significantly more read noise in hardware binning mode than in 1x1 mode so there is actually not a lot of read noise advantage from on-chip binning.

However, if you take 1x1 subs that are long enough that the read noise is insignificant cf the shot noise, read noise is not an issue at all with software binning and there really is is no advantage to on-chip binning, only disadvantages.

[RickS]

Good point, Ray. I also like to do my RGB unbinned so I have the option of extracting additional luminance data.

Cheers,
Rick.

[JohnH]

Hmmm - the reason to bin in hardware is surely to get a reduction in exposure time or better s/n for the same exposure time (given you have small pixels and can trade away the resolution). Any software binning after the event is not going to give you signal that is not there (ie that is is lost in the background noise).

If we assume our CCD has a readout noise of 10e and out signal is twice that, then in the single pixel example each pixel is readout with a noise of 10e hence we achieve a signal to noise ratio of 2:1 (20e/10e). Even if we subsequently sum the four pixels in a computer after readout the signal to noise ratio becomes 4:1. In adding the four pixels we sum the signal (4 times 20e i.e. 80e) and the noise is added in quadrature i.e. square root of the sum of the noises squared (square root of 4 times 10 squared i.e. 20e). In the binned example there is no noise until the signal is readout by the amplifier so the signal to noise ratio is 8:1(80e/10e) i.e. twice as good as the single pixel readout mode.

Or at least that is my understanding....

[Shiraz]

I don't quite understand what you mean by signal being lost in background noise - apart from the dark current pedestal (which could possibly be considered to be noise), noise is simply variability in the signal - it is not something that in some way can obliterate the signal. Binning is just one way to average numerous independent measures of the signal to reduce the uncertainty (the noise). It doesn't matter if you do it on the chip or in software, the same signal will be there in either case.

Your analysis is correct in that perfect on-chip binning provides 2x the SNR of software binning. However, the 8300 does not do perfect binning - from a few sources, the single read noise in binning is somewhere around 1.3x that in 1x1 mode, so some of the SNR advantage is lost that way.

But the main point is that, if you get enough signal in the subs that shot noise dominates (and that is what you should aim for anyway), then it doesn't matter much what the read noise is in software or hardware binning. In your example let's take a more realistic signal of 1600 in each pixel, for which the total shot noise will be 80 with 2x2 binning. For your read noise of 10, the combined read/shot noise with on-chip binning would be 80.6 and for software binning 82.5 - ie there is only a minor advantage to on-chip binning. That very slight advantage would likely be overwhelmed by the increased single-event read noise in the on-chip binning mode of the CCD (this is certainly so for the 8300). In these circumstances software binning provides better dynamic range, a little bit more SNR and more flexibility.

[JohnH]

I must have expressed it badly for a given temp/ccd/gain/exposure noise will be fixed whereas signal will vary with the brightness of the object, thus for a short exposure faint details are lost as they cannot be distinguished from the noise. We use longer exposures or multiple exposures to deal with this.

I take your point re s'ware binning but it has been my experience that s/n is better with hardware binning - I guess it depends on the chip.

[Bassnut]

This has been a very interesting thread.
The 8300 seems to not be the best chip to demonstrate the advantages of hardware binning, and good at showing when its NOT usefull!.

So, yes, it does seem to depend on the sensor and I agree its all about read noise (when the sensor allows a binning advantage).

My 6303 (NABG) sensor has high read out noise but large well depth. 100000 per pixel cell and 330000 for register cells.

Anecdotally anyway, I get better S/N with hardware bin 2*2 because the 6303 has high read noise and its difficult to overcome read noise in dark skies. It requires very long exposures.

Binning with urban imaging, even with short exposures with an 8300 sensor does appear to be fruitless.

[gregbradley]

Binning is a bit like megadata. Ideally you alway use 1x1 but sometimes its time efficient to use 2x2.

2x2 is fine for non critical detail in images like O111, S11, Ha if its not full of detail or the main item in the image.

If you want the best possible image and have tons of time, clear skies, no bad weather and reliable gear then sure go for 1x1 and add up the signal to noise ratio.

Greg.

[traveller]

Keep them coming gents, I am learning tonnes just by reading through the posts. :thumbsup:
Bo

[Bassnut]

Well, thats a bit sweeping, dunno about that.

For colour non res critical data your still better off, depending on sensor (as discussed) , and if your not using RGB to add to bin1 Lum, as in NB, and you are imaging in dark skies, even huge NB colour data is better in bin 2 unless you also have long, as in or more than 1hr subs.

[uwahl]

I have been at this for a very short time so limited experience but my impressions from listening and reading are as follows:-

1 The major reasons to bin or not are driven by seeing, focal length, pixel size and the need to avoid under or over sampling.

2 When a CCD is read the charge in each pixel or group of pixels if binned is shifted to the output registers and then read. It is at this point that read noise is introduced. Therefore whether you read 1x1, 2x2 or 3x3 pixels you only introduce read noise once. This means that your signal to noise ratio increases if you on chip bin and that's good.

3 Binning however sacrifices resolution and that may be not so good.

4 There are other sources of noise (eg dark current) but I am under the impression these are not affected by binning :question:

Can anyone tell me if I have that approximately right and if not where my thinking is off track.

There have been suggestions that the 8300 chip suffers from extra read noise above that experienced when not binning if on chip binning occurs. I would be grateful if anyone can explain how this happens and better still put some sort of value on it. I am afraid I am one of those irritating nerdy types who just HAS to know how things work:shrug:

[Terry B]

Another reason to bin is the faster read time and smaller files. I do time series photometry and will usually use 2x2 binning unless my target is saturating with the binned image and a reasonable length exposure (10 secs).
The binned data has less time between exposures and are much faster to process.
Even binned I have 1.3 arcsec/ pixel so don't under sample.

[Shiraz]

some of the posts in this might be worth a read as well: http://www.cloudynights.com/ubbthreads/showflat.php/Cat/0/Number/5495336/Main/5494372