sCMOS sensor technology

[BJL]

A company called SCMS is offering sensors using an ideas similar to one discussed in these forums: reading each photosite twice, with low and high gain, to increase dynamic range: see
http://www.scmos.com/
http://www.scmos.com/downloads/
http://www.scmos.com/files/high/scmos_white_paper_8mb.pdf

P.S. As keeps happening with new high DR sensor technologies, the company seemed aimed at special purpose markets, not mentioning "normal" photography at all.

P.P.S. One partners in this company is Fairchild Imaging, a major player in areas like large sensors for satellite imaging. So they look serious. The white paper also has some good background on sensor technologies, like the pros and cons of CCD vs CMOS.

[ejmartin]

An interesting read.  Seems like the well capacity is not all that great?  The read noise is about 1.5 e-, but the DR is 16000:1, so it would seem the well capacity is only about 10-11K electrons for a 6.5µ pixel.  The 6.4µ pixels in the 20D have a well capacity of over 50K electrons.  I'm wondering how to square this with their claim of 60% QE.  

Anyway, very good DR but it would seem geared toward low light/high sensitivity applications.

[BJL]

Seems like the well capacity is not all that great?  The read noise is about 1.5 e-, but the DR is 16000:1, so it would seem the well capacity is only about 10-11K electrons for a 6.5µ pixel.  The 6.4µ pixels in the 20D have a well capacity of over 50K electrons.  I'm wondering how to square this with their claim of 60% QE.  

Anyway, very good DR but it would seem geared toward low light/high sensitivity applications.

Firstly, the spec is >16,000:1 at 30fps, and at that frame rate, the read noise is given as <2e-: if we take these as 16,000 and 2e- respectively, the maximum signal would be 32,000e- wouldn't it? Not great, but not terrible. A few years ago, Kodak FF CCD's with 6.8 micron cell size has saturation signal 40,000e-, but more recently that is up to 60,000e- due in part to using deeper wells. It does not surprise me that the SCMOS makers are a bit behind the MOS sensor juggernauts (Sony, Canon, Panasonic) in some details.

Secondly, the sensor has microlenses and no CFA (it is monochrome) so 60% QE is good but not stunning: for example the Kodak KAF-16803 is a rather old monochrome CCD with microlenses, and is also 60% QE at 550nm.

P.S. It seems likely that the extreme DR here is far more relevant to scientific detection situations that "normal" or "artistic" photography. In technical photography, to which this sensor is targeted, it can be very useful to detecting even a few photons producing only a a couple of electrons and thus a local S/N ratio near 1:1. For normal photography, such deep shadow details with such low S/N will be either too dark to matter or horribly noisy to look at. My rule of thumb, suggested by a Kodak document, is that you usually want at least a 10:1 local S/N ratio, requiring at least 100 photons counted and at least 10e- of shot noise. If so, getting dark noise below about 5e- is not of much relevance to normal photography.

The same reasoning is probably why the Electron Multiplying CCD's mentioned in that document are relevant to some technical photography, but have generated no interest for mainstream digital photography: they reduce dark noise at the cost of making shot noise worse; probably a net loss for normal photography.

[ejmartin]

I was looking at figures 5, 7, and 8 of the white paper, which quote a 1.5e- read noise level, but perhaps this refers to another chip.

As far as getting dark noise below 5e- not being of much relevance, I think the biggest advance is getting pattern noise down, something not conveniently measured in terms of electrons.  This was a major advance of the D3 and D3x, and now perhaps the 7D.  I suspect getting the overall read noise down in electrons means getting the pattern noise down along with it.  If my simulation of dual gains has any validity, the low read noise makes for superb shadows at low ISO.  Pattern noise seems largely an artifact of amplification, and so should go down in shadows due to the use of dual amplification.  Low ISO read noise in current DSLR's is mostly amplifier noise, and about 10 electrons (D3x) to 20 electrons (1D3); dual amplification will reduce that to about 2 electrons (7D).  

BTW, for a full well of 32000 electrons at ISO 100, that's 1000 electrons at RAW saturation at ISO 3200; midtones are then at 100 electrons/pixel.  Getting the read noise down below two electrons is going to make for much cleaner shadows at high ISO.  The sample images in the white paper bear this out.

[joofa]

The same reasoning is probably why the Electron Multiplying CCD's mentioned in that document are relevant to some technical photography, but have generated no interest for mainstream digital photography: they reduce dark noise at the cost of making shot noise worse; probably a net loss for normal photography.

I think electron multiplying CCD's can be read with gain turned off (effectively reducing them to CCD type reading), so making shot noise worse may not be an issue here.

[BJL]

BTW, for a full well of 32000 electrons at ISO 100, that's 1000 electrons at RAW saturation at ISO 3200; midtones are then at 100 electrons/pixel.  Getting the read noise down below two electrons is going to make for much cleaner shadows at high ISO.  The sample images in the white paper bear this out.

With 100e- and thus 10e of shot noise, 5e- of red noise gives a total of 11.1e of noise, S/N ratio 9:1, whereas completely eliminating read noise gets you to 10e-, 10:1. Is that really a visibly significant improvement? The interline CCD compared to is very loosely specified as 4 to 10e- of noise: I am, sure that 10e- can look significantly worse!

I am even more sure that further lowering of read noise from <2e- to the <1e- of EMCCD is irrelevant to image quality in "normal photography". (EMCCD also seems to need large wells for all those extra electrons, reducing resolution.)


Still, when you have to push high ISO, read noise can be significant at some level (2e-? 5e-?, 10e-?) But that situation does not benefit from this dual gain, high overall DR solution; it just needs the "high gain, early in the signal chain" approach. THis is achieved by the approach (also used by SCMOS) of applying variable gain for ISO speed adjustment as part of the signal transfer from photosite to column bottom. Sony Exmor CMOS sensors already do this, and Canon is quite likely doing so with recent sensors.

[ejmartin]

With 100e- and thus 10e of shot noise, 5e- of red noise gives a total of 11.1e of noise, S/N ratio 9:1, whereas completely eliminating read noise gets you to 10e-, 10:1. Is that really a visibly significant improvement? The interline CCD compared to is very loosely specified as 4 to 10e- of noise: I am, sure that 10e- can look significantly worse!

What I'm saying is that at low ISO, however one measures the pattern noise, it is 32x worse at ISO 100 than it is at ISO 3200, relative to signal, assuming it is a fixed postamplification noise.  So if they really are using dual amplification of the sort that I described, this will be a big help in cutting down on pattern noise at low ISO.  And the pattern noise does seem to be pretty good in the white paper, though it looks like the monkeyed with the black point of the CCD example to make it look nasty in comparison.

It's also not just the std dev of noise that is important; if the noise is more impulsive, it will be visually more distracting.  Read noise does seem to have heavier tails than a pure Gaussian.

I am even more sure that further lowering of read noise from <2e- to the <1e- of EMCCD is irrelevant to image quality in "normal photography". (EMCCD also seems to need large wells for all those extra electrons, reducing resolution.)

A read noise enough lower than 1e-, so that one is counting photons, would be fantastic.  It would make possible large reductions in pixel size without loss in image quality, assuming saturation density and QE can be kept fixed.

Still, when you have to push high ISO, read noise can be significant at some level (2e-? 5e-?, 10e-?) But that situation does not benefit from this dual gain, high overall DR solution; it just needs the "high gain, early in the signal chain" approach. THis is achieved by the approach (also used by SCMOS) of applying variable gain for ISO speed adjustment as part of the signal transfer from photosite to column bottom. Sony Exmor CMOS sensors already do this, and Canon is quite likely doing so with recent sensors.

I think a lot of low light photographers would love to have 10 stops of DR, without pattern noise, at ISO 3200.

[BJL]

I can see one possible benefit from very low read noise, like <1e-, if achieved without multiplication of shot noise: by making read noise more or less irrelevant, it allows the use of very many smallish photosites to get essentially the same combined noise (shot plus read) over any given region of the image as given by fewer, larger photosites on the same size of sensor. Then one has the advantage of extra resolution when that is useful without any "image level" disadvantage in noise levels, handling of high subject brightness range and so on. This because with only shot noise to worry about, processing options like downsampling to lower pixel counts can give the same per pixel signal and noise as if one used fewer, bigger photosites on the same sensor size to start with. In a sensor with negligible read noise, I would vote for about 16,000e- well capacity, not much more. At base ISO with scenes of normal SBR, midtones would get a healthy 1600e-, and so an excellent local S/N of 40:1. Using bigger photosites with greater well capacity would just throw away resolution for no good reason.

About ISO 3200 performance: unless I am missing something, SCMOS could be described a single ISO speed approach: one never needs to adjust ISO in the analog signal processing: all the "high ISO information" and "low ISO information" is in the pair of 11 bit output values for each photosite.