A new generation of "Perfect" Sensors?

[Tronhard]

The following article appeared in DP Review with regards to a whole new generation of sensors that may be able to count every photon. Will it result in greatly enhanced camera sensors one day..? Let's wait and see.

https://www.dpreview.com/interviews/...ome-to-cameras

[pnodrog]

These quantum sensors were mentioned last year in an article I was reading about a digital hybrid sensor. It was a sensor that would automatically reset the well when it was full and keep count of how many times it happened (from memory up to 3 times). These counts 0-3 were added to the normal ADC results as the most significant 2 bits. Effectively boosting dynamic range by 2 stops.

I am sure in the next few years we will see a leap in dynamic range. It seems to be a very popular area of research at Universities as well as by the manufacturers.

[TonyW]

These would be an interesting development, especially being able to increase the dynamic range. However, as it says in the article, it does not eliminate noise because 'shot' noise is a major part of the overall noise and this would not be eliminated. Shot noise is caused by the discrete nature of the light gathered which is a random variation.

This raises another question in my mind. I would have thought that the number of photons gathered is so large that the discreteness is not important, but apparently that is not so. Can anybody tell me the order of magnitude of the number of photons gathered by a sensel when a photograph is taken, at least to within a factor of 10 or so?

[pnodrog]

Tony, I understand it is currently between 10,000 and 100,000 per pixel depending on the sensor. Due to the photon absorption characteristics (and conversion to electrons) in silicon Clarkvision claim the capacity is more a function of the pixel's well area rather than volume. If you are interested or have nothing much to do you could read this. It is reasonably easy reading.

[dje]

Yes Bill Claff also has some figures for FWC in electrons here.http://http://www.photonstophotos.net/Charts/Sensor_Characteristics.htm. Select the camera models of interest or look at the list down the page a bit. Not sure how they compare with Roger Clarke's figures.

These are Full well figures, the actual no of electrons for any one pixel depends on the light and could range from a few right up to the fwc.

I'm still struggling to grasp the concept mentioned in the DPReview article but I guess Eric Fossum knows his stuff! As always, it's interesting to read the large number of member comments in the DPReview article, the tone of them being vintage DPReview clientele stuff!
Dave

[TonyW]

Paul, that's an interesting reference. For my Camera, Canon 5d3 for example, the "full well capacity" is quoted at 70,000 electrons/photons. It is supposed to have a dynamic range of 12 stops at lowest ISO so that at the bottom of the range there would be 70000/4096 electrons, which comes to 17 electrons. I can believe that shot noise would start to be important at that number.

Given what I learnt from the recent thread on photosites' capacitors, I would deduce that the "full well capacity" is the number of electrons needed to fully discharge the capacitor. Does that sound right? My knowledge of semi-conductors is hazy.

It also makes sense that recharging the capacitor during an exposure increases the effective capacity of the sensel and thereby increases the dynamic range.

[xpatUSA]

Paul, that's an interesting reference. For my Camera, Canon 5d3 for example, the "full well capacity" is quoted at 70,000 electrons/photons.

You can't really include photons with electrons that way, sorry Tony. What goes into the capacitance is incident photons times the quantum efficiency (QE) of the sensor pixel at some specified wavelength, normally 555nm. In other words, only the captured electrons count. Of course we should realize that the QE has a spectral value with respect to wavelength, so it ain't that simple.

For example and by coincidence, one of my cameras has the same full-well capacity of 70,000e-. Here's the QE spectra of it's sensor:



Please ignore the obfuscatory black line ... that's Foveon's Marketing Dept. at work. If the individual color QE's got big enough, the efficiency would go over 1.0, obviously impossible unless our sensors suddenly turned into photo-multipliers, LOL.

For those puzzling over the response at 700nm+ the curves are for a bare sensor with no UV/IR blocking filter.

[pnodrog]

Paul, that's an interesting reference. For my Camera, Canon 5d3 for example, the "full well capacity" is quoted at 70,000 electrons/photons. It is supposed to have a dynamic range of 12 stops at lowest ISO so that at the bottom of the range there would be 70000/4096 electrons, which comes to 17 electrons. I can believe that shot noise would start to be important at that number.

Given what I learnt from the recent thread on photosites' capacitors, I would deduce that the "full well capacity" is the number of electrons needed to fully discharge the capacitor. Does that sound right? My knowledge of semi-conductors is hazy.

It also makes sense that recharging the capacitor during an exposure increases the effective capacity of the sensel and thereby increases the dynamic range.

The typical efficiency with the current sensors seems to be in the region of 40% to 60% but some of the back-illuminated sensors are approaching 80%. Some of the noise is due to the rather random rate at which the photons arrive. They are very disorderly and unfortunately don't space themselves out evenly. If they were orderly collecting them as a true representative sample at low light levels would be much simpler. The highlights should have a closer statistical relationship to the actual brightness than the shadows where too few or too many turn up for the given exposure. For long exposures, the sensor noise to a limited extent can be estimated by taking a second exposure and then corrected by subtracting it from the actual exposure. For short exposures, the sensor noise is too random and too low to use this method to make a meaningful correction.

http://www.cse.wustl.edu/~jain/cse567-11/ftp/imgsens/index.html

A bit more reading. Actually, the amount of information available online is amazing. The hard bit is sorting out how relevant it is.

From a photography point of view understanding the intricacies of how sensors work is probably less important than knowing how they perform.

[dje]

For my Camera, Canon 5d3 for example, the "full well capacity" is quoted at 70,000 electrons/photons. It is supposed to have a dynamic range of 12 stops at lowest ISO so that at the bottom of the range there would be 70000/4096 electrons, which comes to 17 electrons. I can believe that shot noise would start to be important at that number.

Given what I learnt from the recent thread on photosites' capacitors, I would deduce that the "full well capacity" is the number of electrons needed to fully discharge the capacitor. Does that sound right? My knowledge of semi-conductors is hazy.

Tony the lower end of the Dynamic Range corresponds to the point where the signal is just useable. It is probably most common to use the so called Engineering Dynamic Range definition in which this point is considered to be equal to the read noise level. Another definition called Photographic Dynamic range considers the lowest useable signal to be somewhat higher than the read noise by a few db. Either way, at this signal level, read noise is the dominant source of noise. Shot noise increases with the square root of the signal level and becomes the dominant source of noise at higher signal levels than the bottom end of the DR. S/N measurements are often quoted at 18% grey level and here the shot noise is dominating.

In a CMOS sensor, the capacitance is not fully discharged at FWC because the voltage-charge response curve is becoming non-linear at low charge levels. So typically the FWC corresponds to the largest negative output voltage swing possible in the linear region of the curve.

Dave

[pnodrog]

Tony the lower end of the Dynamic Range corresponds to the point where the signal is just useable. It is probably most common to use the so called Engineering Dynamic Range definition in which this point is considered to be equal to the read noise level. Another definition called Photographic Dynamic range considers the lowest useable signal to be somewhat higher than the read noise by a few db. Either way, at this signal level, read noise is the dominant source of noise. Shot noise increases with the square root of the signal level and becomes the dominant source of noise at higher signal levels than the bottom end of the DR. S/N measurements are often quoted at 18% grey level and here the shot noise is dominating.

In a CMOS sensor, the capacitance is not fully discharged at FWC because the voltage-charge response curve is becoming non-linear at low charge levels. So typically the FWC corresponds to the largest negative output voltage swing possible in the linear region of the curve.

Dave

I totally agree with Dave but feel I need to clarify in case it is misinterpreted. Because shot noise increases with the square root of the signal it does overtake and then dominates the read noise as the signal increases. However, the total noise compared to the signal quickly diminishes so it is less apparent in the lighter parts of the image.

[dje]

I totally agree with Dave but feel I need to clarify in case it is misinterpreted. Because shot noise increases with the square root of the signal it does overtake and then dominates the read noise as the signal increases. However, the total noise compared to the signal quickly diminishes so it is less apparent in the lighter parts of the image.

This diagram from Emil Martinec's well known article on noise in digital cameras shows the situation quite well. S/N is shown in stops but can be converted to db by multiplying by 6. DXOMark publish figures like this but that particular download doesn't seem to be working at present. The area where the slope of the lines change is where shot noise is taking over from read noise as the main noise contributor.


Fig. 12 - Signal-to-noise ratio vs. exposure of the 1D3, for various ISO. The horizontal axis is the raw value (signal) in EV
horizontal coordinate x corresponds to raw value S=2^x-1); (that is, the vertical axis is S/N ratio in stops (for S/N in dB, multiply by six).
Only read and photon noise have been taken into account in generating the above plot; PRNU and other noise sources are not incorporated.

[TonyW]

This discussion is reaching the usual state where the number of terms used exceeds the number understandable in finite time.

I made a mistake in conflating the number of photons with the number of electrons, i.e. ignoring the "quantum efficiency".

Dave, I don't understand your statement about shot noise. With a Poisson distribution, the standard deviation increases in proportion to the share root of the mean, so that the signal to noise ratio INCREASES in proportion to the square root of the mean. Hence the shot noise should become less important as the signal increases.

I still stand by my original statement that the number of electrons collected at the bottom of the range is quite small so that it is not surprising that shot noise is important there. I don't properly understand read noise or how it behaves. That is beyond me at this stage.

[dje]

Dave, I don't understand your statement about shot noise. With a Poisson distribution, the standard deviation increases in proportion to the share root of the mean, so that the signal to noise ratio INCREASES in proportion to the square root of the mean. Hence the shot noise should become less important as the signal increases.

I still stand by my original statement that the number of electrons collected at the bottom of the range is quite small so that it is not surprising that shot noise is important there. I don't properly understand read noise or how it behaves. That is beyond me at this stage.

Tony I'll refer to and explain further what I said in my post 9

"Either way, at this signal level, read noise is the dominant source of noise. Shot noise increases with the square root of the signal level and becomes the dominant source of noise at higher signal levels than the bottom end of the DR."

I was referring to the noise levels, not S/N. Read noise is the noise that is read with no signal (ie black) and is independant of signal level. Shot noise on the other hand increases with signal level. Paul in his post 10 went on to clarify the effect on S/N and I also posted a graph in post 11 that re-enforces what he said. I'm not saying that shot noise dominates the signal more and more as the signal level rises, only that it dominates the other noise contributors.

If you look at the graphs in post 11, on the left hand side, the S/N is rising stop for stop as the read noise is constant- each stop increase in signal increases the S/N by 1 stop. On the right hand side, the S/N is rising at a slower rate because shot noise is dominating and the noise is only increasing as the square root of the signal. This is all explained in the article I referenced in post 9.

Yes there will be some shot noise for a signal level equal to the read noise but it will be significantly lower than the read noise.

Dave

[xpatUSA]

Yes, there will be some shot noise for a signal level equal to the read noise but it will be significantly lower than the read noise.

Dave

Just to put some numbers to that, one of my camera sensor data-sheets quotes an RMS read noise level of 70-e (electrons).

If now we capture a mean signal of 70-e the shot noise for the signal would be about 8e-, much less than 70.

If then we capture the recommended maximum (linear) amount of 45,000e- the shot noise for that would be about 850e-, much, much more than 70.

Not saying anything new to this thread - just confirming some statements made to date.