Sensor DR vs Camera DR

[douglasf13]

Apical, who makes the chip in the A700 for DRO+, has this technology that some of you may be interested int.

http://www.apical-imaging.com/adiso

[Cartman]

Apical, who makes the chip in the A700 for DRO+, has this technology that some of you may be interested int.

http://www.apical-imaging.com/adiso
[a href=\"index.php?act=findpost&pid=211853\"][{POST_SNAPBACK}][/a]

Thanks for the link -- new to me.

[Guillermo Luijk]

Apical, who makes the chip in the A700 for DRO+, has this technology that some of you may be interested int.

http://www.apical-imaging.com/adiso
[a href=\"index.php?act=findpost&pid=211853\"][{POST_SNAPBACK}][/a]

That idea would be super-optimized, but I wonder how complex is to implement an analogue ISO gain varying pixel to pixel. The PDF just explains the idea, but gives no details at all about circuit implementation.

[ejmartin]

Apical, who makes the chip in the A700 for DRO+, has this technology that some of you may be interested int.

http://www.apical-imaging.com/adiso
[a href=\"index.php?act=findpost&pid=211853\"][{POST_SNAPBACK}][/a]

I looked at their white paper, which makes it clear that what is being discussed is tone-mapping post-capture, in the digital domain.  In short, it will do nothing to increase dynamic range or improve shadow S/N, which is the goal of the proposal.

[douglasf13]

I looked at their white paper, which makes it clear that what is being discussed is tone-mapping post-capture, in the digital domain.  In short, it will do nothing to increase dynamic range or improve shadow S/N, which is the goal of the proposal.
[a href=\"index.php?act=findpost&pid=211917\"][{POST_SNAPBACK}][/a]

  Yeah, I didn't mean to infer that Apical uses the same technology that you guys are referring to.  I just wanted to show an alternative idea that was out there.

[BJL]

Emil, I am still puzzled. The 1DMkIII and D3  both use 14 bit ADC's, and yet apparently neither gives as much as 12 stops of DR at any ISO, so the observed DR could be handled by 12 bit output with well chosen level. I see only these possibilities:

1) Canon and Nikon's nominally 14-bit ADC's introduce noise levels which reduce the S/N ratio of the digital output to under 12 stops; less than what the 12 most significant bits of the ADC output contain, so the two least significant bits are garbage.

2) The full headroom of the ADC is not being used at low ISO speeds: low ISO amplification is converting even full well signals to voltages giving to an output level well below maximum, so the two most significant bits are zeros.

3) The ADC's have S/N ratio better than any 12-bit ADC could give, but the signal entering the ADC's has DR less than this (due to noise earlier in the process, like in ISO gain pre-amplification)*

4) Some combination of the above.


The only way I can see that Canon and Nikon are not being highly misleading in their touting the advantages of the 14-bit ADC's in these cameras is that the 14-bit range is used so that pre-amplification does not have to achieve a very precise match of maximum input voltage to the voltage that gives maximum output in ADU. It may be easier to drive a signal with DR almost 12 stops wide through a tunnel 14 bits wide than through one that is only 12 bits wide, if you wish to avoid hitting one side and so losing some DR.

So I vote for some combination of (2) and (3).


* Olympus and Panasonic have indicated that the 10MP NMOS sensor of the E-3 achieves reduced noise levels and thus greater DR compared to earlier versions of that 10MP sensor due to reducing the noise produced in the on-sensor amplification. This suggests to me that amplifiers are a plausible source of significant noise in a CMOS sensor too.

[ejmartin]

Emil, I am still puzzled. The 1DMkIII and D3  both use 14 bit ADC's, and yet apparently neither gives as much as 12 stops of DR at any ISO, so the observed DR could be handled by 12 bit output with well chosen level. I see only these possibilities:

1) Canon and Nikon's nominally 14-bit ADC's introduce noise levels which reduce the S/N ratio of the digital output to under 12 stops; less than what the 12 most significant bits of the ADC output contain, so the two least significant bits are garbage.

2) The full headroom of the ADC is not being used at low ISO speeds: low ISO amplification is converting even full well signals to voltages giving to an output level well below maximum, so the two most significant bits are zeros.

3) The ADC's have S/N ratio better than any 12-bit ADC could give, but the signal entering the ADC's has DR less than this (due to noise earlier in the process, like in ISO gain pre-amplification)*

4) Some combination of the above.
The only way I can see that Canon and Nikon are not being highly misleading in their touting the advantages of the 14-bit ADC's in these cameras is that the 14-bit range is used so that pre-amplification does not have to achieve a very precise match of maximum input voltage to the voltage that gives maximum output in ADU. It may be easier to drive a signal with DR almost 12 stops wide through a tunnel 14 bits wide than through one that is only 12 bits wide, if you wish to avoid hitting one side and so losing some DR.

So I vote for some combination of (2) and (3).
* Olympus and Panasonic have indicated that the 10MP NMOS sensor of the E-3 achieves reduced noise levels and thus greater DR compared to earlier versions of that 10MP sensor due to reducing the noise produced in the on-sensor amplification. This suggests to me that amplifiers are a plausible source of significant noise in a CMOS sensor too.
[{POST_SNAPBACK}][/a]

My understanding is a combination of (1) and (3), at least for the 1D3.  Measurements of read noise vs ISO are well modeled in the 1D3 by a two-stage amplification:

[a href=\"http://theory.uchicago.edu/~ejm/pix/20d/tests/noise/noise-p2.html#read_vs_iso]http://theory.uchicago.edu/~ejm/pix/20d/te...tml#read_vs_iso[/url]

See the table below Figure 15.  There it is observed that the dominant noise source occurs between the two stages of amplification, with a smaller but substantial contribution from sources downstream of the second amplification.

[BJL]

... at least for the 1D3 ... it is observed that the dominant noise source occurs between the two stages of amplification, with a smaller but substantial contribution from sources downstream of the second amplification.
[a href=\"index.php?act=findpost&pid=213413\"][{POST_SNAPBACK}][/a]

Thanks. So the main noise source at low ISO settings is before input to the ADC, since it is in particular before the secondary ISO adjustment analog amplification stage. That make sense to me, and refutes the idea that Canon is discarding substantial potential DR improvement in these high end cameras simply by using inadequate 14-bit ADC's. (Cheap tires on an expensive sports car!)

The idea that noise and DR are significantly affected by analog stage noise sources such as in amplification for ISO adjustment also fits with the comment from Olympus and Panasonic that improved amplifier designs lead to less noise (and thus higher DR) with the 10MP NMOS sensor for the E-3, compared to their other 10MP NMOS sensors from earlier in the same year.

The analog stages between photosite and ADC might be the last main frontier of noise reduction.


P. S. Do you know if one or both of these amplification stages are done on the sensor chip or off it?
I am fairly sure that A/D conversion is done off-chip in Canon DSLR's, the D3 and probably in most DSLR's. Maybe the only SLR sensors with on-board A/D conversion are the Sony EXMOR ones, as in the D300 and A700.

[EricV]

Another way to increase dynamic range is to make two exposures, one for the highlights and another for the shadows, then combine the digital images offline.  GLuijk has discussed this technique exhaustively in the "zero noise" thread.  

This technique could be implemented in the camera with no extra electronics.  For example, the user presses the shutter button, set say to 1/125 sec, the camera acquires a 12-bit image at 1/125 sec for the highlights, then aquires another 12-bit image at 1/8 sec for the shadows, then merges the two images into a 16-bit final output.  

As this example shows, the technique would work best for fairly static subjects with a tripod mounted camera, since the exposure for the shadows would tend to be long, but the technology for this is here today, and it would extend the dynamic range of any sensor as much as desired.

[BJL]

Another way to increase dynamic range is to make two exposures, one for the highlights and another for the shadows, then combine the digital images offline.
[a href=\"index.php?act=findpost&pid=213428\"][{POST_SNAPBACK}][/a]

And Panasonic has already announced sensors that so something like this, aimed for now at surveillance video cameras.

Another surveillance video oriented sensor does multiple partial reads during a single exposure, so moving subjects could be handled. Main idea: detect nearly full wells early in the exposure and do A/D conversion right then at each pixel, storing the digital result at the pixel, and then reading off the sensor after the whole exposure is finished.


The surveillance video market seems to make greater demands for DR than DLSRs!