Albert Theuwissen continues his series of binning vs averaging comparison articles. In the most recent one, "
the situation for a shot-noise limited condition is considered. And actually the story can be very short : it does not matter when or where the binning is done, in all cases the result is exactly the same."
The full story includes a proof of that.
You can simply consider the shot noise as read noise, then you have the same result either by binning or by averaging...
ReplyDeleteThis statement does not hold for binning in the charge domain ! See the previous blog. When the noise is read noise limited, there is a major difference.
ReplyDeleteIf the photon shot noise dominates, then the readout noise can be ignored. In this case, when you add charge you will get sqart(2) more shot noise. So it's that easy and don'est need any complexe demonstration. :)-
DeleteThe role of binning is to improve the performance in low light and low light is not shot noise limited. It still makes a lot of sense to do binning in the charge domain.
ReplyDeleteNot really true. Low light signal noise can be well above the read noise floor, but still not enough for suitable SNR. Example, 1.5e- rms read noise, 25e- signal. Shot noise limited but binning may still be desired.
DeleteIn thermal imaging, it's almost always background shot noise limited. Binning can give better NETD since the SNR is always improved by binning.
Delete-yang ni
Charge binning in most of the cases reduces the conversion gain hence increases read noise. The advantage of charge binning is not always that obvious I think.
ReplyDeleteThe real charge domain binning ( CCD like) will not change the conversion gain.
Deletepeng
But in CMOS you can't, and like Adi says you'll always lower the CvG.
DeleteI think you should approach the problem from the other direction. The CG (1/FD cap) is always tailored wrt the full-well capacity of the photodiode. Now, *if your application* demands low light situation, you can *choose" binning (i.e., you dont have to necessarily make your CG capacitance 2X). But the camera can also handle medium or high-light situation without binning. i.e., the same camera is useful to be applied in various (static)conditions. This is quite desirable for industries, where re-usibility of cameras is required. Of course, HDR application is quite different and there your logic applies.
DeleteAbout charge domain binning in CMOS : if you apply this with a shared pixel concept, then the conversion factor is not changed, or am I wrong ?
DeleteI am happy to read all these comments.
DeleteIt shows that the subject of binning versus averaging is quite interesting, and apparently there is a need for a more precise definition of what is what.
Keep following the blog, there is much more to come about butting and stitching !
"About charge domain binning in CMOS : if you apply this with a shared pixel concept, then the conversion factor is not changed, or am I wrong ?"
DeleteYou need to design your FD cap to be 2X (for handling the total charge if you do charge-domain binning) to preserve the whole dynamic range: implies noise is also going up by 2X in electrons. Therefore I think Adi's comment above. But if you use it situationally (i.e., limit intra-scene dynamic range, which usually is the case with many applications in industry)you can benefit from binning/and re-use the same camera also for other situations. This is the second comment by Anon at 5.22 pm
In my opinion, the ~2 X Cfd is not needed.
DeleteThe binning is either useful for low light (as Arnaud mentioned above) or high SNR in shot noise limited case just because of the larger full well (as Eric and Yang mentioned above).
In low light, the 2 X Cfd is not needed, one full Cfd will bring you in to shot noise limited anyway.
In the shot noise limited case, like explained in Albert's blog, the digital (voltage) binning is the same as charge domain. No reason to incresase Cfd here.
Peng
Peng, in 2-pixel shared FD configuration CG reduces due to the second tranfer gate contributing to CFD. Say 30% contribution from second TG. The max signal CFD can handle increases by 30%, but your noise increases by 30% as well.
DeleteIn a non shared digital binned configuration your noise increases as sqrt(2), while your max signal increases by 2x. With similar noise performance, the DR of the digitally binned 2 pixels is 3dB higher than that of charge-binned pixels. Also the max SNR is higher, while the noise is very similar.
Unless you find a way to decouple CFD from TG cap contribution, the digital binning almost always wins in CMOS.