hjkarten Loc: San Diego, California

Dear Bill (bclaff),
I skimmed the Martinec article, and realized that I had read it some years ago. It is definitely worth another reading.
Many thanks,
Harvey

bclaff Loc: Sherborn, MA (18mi SW of Boston)

You are getting closer.
There is no conflating of dynamic range (EV) and bit-depth.
Bit-depth sets an upper limit on dynamic range. Dynamic range is *never* expressed as bit-depth.
FWIW, there are many sensor specification sheets PDFs available and you'll never see dynamic range listed as bit-depth, usually as dB.
You may be making an assumption that the chip manufacturer sets bit-depth to match required dynamic range.
That is not true. There are many 16-bit sensors that are well under 16 stops of dynamic range.
And I know of several 12-bit sensors that would have higher dynamic range if they had 14-bit ADCs.

Regards

TriX Loc: Raleigh, NC

Harvey, thanks for your response. Don’t worry in the least about how grating the minute details of digitizing may be to those that just want to take images - we do it all the time, and they’re not forced to read it. We’re only beginning page 7 - we have at least another 5-6 pages to go. 🙀

Cheers,
Chris

Canisdirus

This thread is an example of why Photographers are the third-worst ppl to converse with at social gatherings...
Right behind Dentists and Accountants.

TriX Loc: Raleigh, NC

And you are now part of that group

hjkarten Loc: San Diego, California

In actual practice, I have the impression that EV values are a messy combination of light transmittance of the properties of the lens at different wavelengths and diffractive state, the aperture, the efficiency of the chip, the spectral response of the sensor at different wavelengths and different intensities - especially when either at saturation with photon overflow or at very low photon number with increased shot noise , the nominal ISO chosen, properties of the ADC and those of the processor dealing with the output of the ADC. (Probably also include the software).
Dynamic Range of the system is equally messy, and the blessing of modern digital cameras is that they work as well as they do! (WHEW! That was a long sentence.)
I would have no problem using dB, but habit might prompt me to think of dB in terms of shot noise at the low end and overflow at the high end, particularly if overloading the electron well. What would be the best parameter to use to deal with dynamic range once you consider the whole process from the perspective of image processing?
How many discrete steps of brightness and at each of the different wavelengths do I have to deal with to cover my original analog input if I want an accurate representation of the original field? I can't do that calculation using EV values, and I would have to undergo all sorts of contortions to do that using dB. (I am not smart enough to even know how it can be done?) That is where digital image processing saves our hides. And that is done at the bit level. It is not simple, and is made again more difficult when you know that the final output has to be re-represented as an analogue product.

bclaff Loc: Sherborn, MA (18mi SW of Boston)

You may not know how to do it but dB and EV are entirely interchangeable; just like feet and meters.
For a rough estimate just divide dB by 6 to get EV.

Dynamic range is simply saturation for the high value and, in the case of a pixel, read noise for the low value.
It's not any more complicated than that (and has nothing to do with photon (shot) noise).

You can take this ratio of high / low and convert it to dB or EV.
Personally I think EV makes more sense in the context of photography.

This is dynamic range at the sensor data was represented by raw (linear) data.
It doesn't map easily to dynamic range in the (perceptual not linear) final output.

TriX Loc: Raleigh, NC

Yep, 1 EV = 1 bit = 6.02 dB. Question: why, for very low illuminance of the sensor, do you feel that shot noise doesn't contribute to the noise floor?

hjkarten Loc: San Diego, California

Hi TriX,
I certainly did not mean to suggest that. Perhaps I made a typo?
I even have a paper on the role of shot noise contributing to the noise in the retina. We argued that one of the benefits of binocular vision in nocturnal animals is that the S/N improves as function of the ^^2, and helps differentiate noise from valid signal. This was first proposed by Dan Steinman (Univ of Maryland) more than 30 years ago. The S/N enhancement probably happens at the level of the visual cortex. A similar benefit can be achieved with on chip integration or using longer dwell time between pixel readouts. But that is difficult to accomplish in the retina. Instead the two conjugate points of the two retinae do it by spatial integration and/or averaging. The first location where a single neuron receives topographical input from two corresponding points of the two retinae is at the cortex. A similar pathway is used in the analysis of stereopsis.

Aha! 1 EV = 1 bit!? = 6.02 dB
My search for that conversion drew a blank this morning.
If so, would that also mean that the Dynamic Range of the system can be directly expressed in bits. End of argument.
Many thanks. Also just got the good news that I am expected at the clinic on Wednesday morning to get my COVID vaccine.
Harvey

TriX Loc: Raleigh, NC

Actually, my question was directed to bclaff. My thought was that maybe the contention is that in most photographic applications, there are enough photons striking the sensor that the Poisson distribution is close to a normal distribution, but I’m waiting to hear his thoughts.

Congratulations on the vaccine - my wife’s is tomorrow, and mine is Wednesday also.

Chris

bclaff Loc: Sherborn, MA (18mi SW of Boston)

The noise floor is generally taken as read noise.
Read noise is noise that is present when there is no light.
If there is no light then there is no photon (shot) noise.
So the determination of EDR, pixel level dynamic range, has absolutely nothing to do with shot noise.

DxOMark print dynamic range is simply EDR that is normalized so it also has nothing to do with shot noise (which is one reason it's not that good a measure).

PhotonsToPhotos Photographic Dynamic Range (PDR) does take all forms of noise including shot noise into consideration. The low value in PDR is determined using a Signal to Noise Ratio (SNR) that takes (standard) final image size and viewing distance into consideration.

TriX Loc: Raleigh, NC

Thanks, I take your point, and I share your preference for the methodology of PhotonstoPhotos.

Cheers,
Chris

bclaff Loc: Sherborn, MA (18mi SW of Boston)

So long as you are willing to use fractional bit values rather than whole numbers in which case you just saying EV another way (13.5 EV = 13.5 bits). It still is *not* the bit-depth of the ADC that establishes dynamic range only the upper limit.

bclaff Loc: Sherborn, MA (18mi SW of Boston)

Thanks for that.
You realize I'm a bit biased in that regard; right?

TriX Loc: Raleigh, NC

I didn’t when I mentioned my preference earlier in the thread (and I have many times in the past), but I do now 😎. Lots of interesting links on your page, Bill.

Cheers,
Chris