fredmiranda.com Login
This topic is in the 2015 archive. You can read it and like its posts here. It is read only.

Archive 2015High DR for 5d4

  
 
alundeb
Offline
Details
p.5 #1High DR for 5d4
snapsy wrote:
alundeb wrote
snapsy wrote
alundeb wrote
snapsy wrote
alundeb wroteWhy does the signal have to be 0?

One common definition of DR uses SNR = 1.
The quantization of the LSB on imaging sensor ADCs is at a mean signal level of zero, with a +/- distribution around zero. I'm still not seeing how you arrive at 15.8EV of DR from the 14-bit quantization. Can you supply the fully formula you're using so that it's clearer? Also, what is the use of > 2^bit_depth DR if the result is fully posterized, irrespective of what the precise DR measurement is between 14.0EV and 15.8EV?
The definition of DR I use is maximum signal divided by (signal where SNR = 1). That is the same definition used by ISO and also in the DxOmark measurements.

Assuming that quantization noise is the only noise, the noise level is 1/sqrt(12).

The formula I use to get 15.8 stops is: log(16383/(1/sqrt(12)))/log2
Why are you dividing the bit-depth scale by 1/sqrt(12)? You seem to be mixing units from the quantization error proof with the bit-depth scale. The proper calculation (source, page 6) yields a 100% SNR (and thus DR) result of 6.02(nbits) + 1.76db, which equates to 14.33EV of DR for a 2^14 ADC.
The maximum signal in your source is a sine wave with peak-peak value equalling the ADC full scale, and the amplitude used in the max SNR calculation is the rms value of the sine. For an image sensor, the maximum signal is a unipolar DC value and not a sine wave. This gives a signal gain of 2*sqrt(2) or 9 dB over the formula in your source.
Quantizied DR is the signal peak to stddev irrespective of what the unbounded scale of the input peak into the ADC is. Why would the relative peak have any effect on the ADC's DR calculation? Both get hashed down to quanitization units, converting an infinite set of input measurements (photons, electrons, etc..) into a discrete set of digital output values. The only effect the relative peak-to-stddev differences have is on the resolution of each discrete step.


I don't mean that the relative peak means anything. What I mean is that the maximum signal in the case of a digital image is defined as the ADC full scale value, but in your source the maximum signal is defined as the rms value. 6.02*(nbits) + 1.76 dB is based on the rms value of the signal, not the peak value.

Your source: Vrms = Vfs/(2*sqrt(2)).
Your source: The expectation value of the quantization noise is Vq^2/12.

If we say that Vq = 1 and Vfs = 16383, we get my 15.8 stops.

Apr 13, 2015 at 01:58 PM
alundeb
Offline
Details
p.5 #2High DR for 5d4
snapsy wrote:
alundeb wrote
curious80 wroteHmmm but unless the ADC is assigning the values in somewhat non-linear way, you just don't have enough bits to encode that many levels.
One cannot measure noise or DR for one pixel only. The signal level we use is the mean value, and the noise level is the standard deviation. As long as there is noise, the mean value or the noise level is not limited by the quantization step.
But DR is limited by one or the other (quantiziation bit-depth or noise) so the effect is the same. In other words, increasing the bit-depth but keeping noise the same just means wasted ADC resolution (Ie, oversampling noise). Conversely, keeping bit-depth the same but decreasing noise means undersampling.


But the digital SNR increases as the analog noise decreases even if the analog noise is undersampled.

The total noise is a combination of analog noise and quantization noise, and it is not a hard limit where the effect of one of them suddenly stops.
Apr 13, 2015 at 02:00 PM
alundeb
Offline
Details
p.5 #3High DR for 5d4
snapsy wrote:
Why are you dividing the bit-depth scale by 1/sqrt(12)?
.


Because the quantization noise is 1/sqrt(12) of 1 LSB.
Apr 13, 2015 at 02:14 PM
snapsy
Offline
Details
p.5 #4High DR for 5d4
alundeb wrote:
snapsy wrote
alundeb wrote
curious80 wroteHmmm but unless the ADC is assigning the values in somewhat non-linear way, you just don't have enough bits to encode that many levels.
One cannot measure noise or DR for one pixel only. The signal level we use is the mean value, and the noise level is the standard deviation. As long as there is noise, the mean value or the noise level is not limited by the quantization step.
But DR is limited by one or the other (quantiziation bit-depth or noise) so the effect is the same. In other words, increasing the bit-depth but keeping noise the same just means wasted ADC resolution (Ie, oversampling noise). Conversely, keeping bit-depth the same but decreasing noise means undersampling.
But the digital SNR increases as the analog noise decreases even if the analog noise is undersampled.

The total noise is a combination of analog noise and quantization noise, and it is not a hard limit where the effect of one of them suddenly stops.

Ok, I'm still having trouble conceptualizing how the formula translates into achieving > 2^bit_depth DR. To help suss out where my understanding is falling short, here's a quick spreadsheet of a sample ADC encoding, using a D750 as an example (ISO 100 FWC 81608 with read noise of 5.5e-, with a per-pixel DR of 13.9EV):

http://horshack.smugmug.com/photos/i-RNZFTLR/0/O/i-RNZFTLR.jpg

For argument's sake lets say the D750 has exactly 1 e- of read noise instead of its actual 5.5e-, meaning it has a pixel-DR approaching 16EV. Can you walk me through how the formula would relate to the 14-bit ADC encoding for EVs 15 and 16? As far as I can comprehend right now, the minimum e- sample from a 14-bit ADC would be 5 e-, with the 1/2 LSB swing below and above that, bringing the total DR to the approx 14.33EV. Can you walk me through how the formula's encoding of the quantum bit (LSB) brings the representable DR up to 15.8EV?
Apr 13, 2015 at 07:06 PM
Paul Mo
Offline
Details
p.5 #5High DR for 5d4
I bet Canon are greatful to you guys for figuring out how to get wider DR where they couldn't. No doubt a cheque is the mail.
Apr 13, 2015 at 07:33 PM
alundeb
Offline
Details
p.5 #6High DR for 5d4
snapsy wrote:
alundeb wrote
snapsy wrote
alundeb wrote
curious80 wroteHmmm but unless the ADC is assigning the values in somewhat non-linear way, you just don't have enough bits to encode that many levels.
One cannot measure noise or DR for one pixel only. The signal level we use is the mean value, and the noise level is the standard deviation. As long as there is noise, the mean value or the noise level is not limited by the quantization step.
But DR is limited by one or the other (quantiziation bit-depth or noise) so the effect is the same. In other words, increasing the bit-depth but keeping noise the same just means wasted ADC resolution (Ie, oversampling noise). Conversely, keeping bit-depth the same but decreasing noise means undersampling.
But the digital SNR increases as the analog noise decreases even if the analog noise is undersampled.

The total noise is a combination of analog noise and quantization noise, and it is not a hard limit where the effect of one of them suddenly stops.
Ok, I'm still having trouble conceptualizing how the formula translates into achieving > 2^bit_depth DR. To help suss out where my understanding is falling short, here's a quick spreadsheet of a sample ADC encoding, using a D750 as an example (ISO 100 FWC 81608 with read noise of 5.5e-, with a per-pixel DR of 13.9EV):

http://horshack.smugmug.com/photos/i-RNZFTLR/0/O/i-RNZFTLR.jpg

For argument's sake lets say the D750 has exactly 1 e- of read noise instead of its actual 5.5e-, meaning it has a pixel-DR approaching 16EV. Can you walk me through how the formula would relate to the 14-bit ADC encoding for EVs 15 and 16?
...Show more →

There are two things that can help us conceptualize how DR > 2^N.

One is that the threshold between ADU 0 and 1 in the conversion corresponds to 1/2 ADU, not 1 ADU. Analog values corresponding to between -1/2 and 1/2 ADU are represented as 0, values between 1/2 and 3/2 ADU are represented as 1, and so on.

Using your example, we have a table showing what the ADU values represent in terms of minimum, average and maximum values of e-:

http://i306.photobucket.com/albums/nn242/Overflate/ADU14bits_zpsoleut7j4.png

The other thing is that noise has a probability density function that extends beyond the standard deviation. If we use gaussian noise with standard deviation 1 e- in this example, the approximate values for the probability are:

http://i306.photobucket.com/albums/nn242/Overflate/Gauss_1ev_zpsxrdanxhy.png

Finally if we add this 1 e- of gaussian read noise to the signal with fwc 81608 e-, and run it through an ideal A/D converter, the probabilities of getting certain ADU values for a given signal level are given in the following table. What I should have done is to calculate the exact probability for each ADC value, so the table here is not accurate , but it shows the concept.

http://i306.photobucket.com/albums/nn242/Overflate/ADC_14bits_prob_1ev_zpsjikqnwld.png

Edit: I forgot to normalize the noise to ADU. The last table is a bit different, corrected now.

EV 1 is encoded as 'always' 16383
....
EV 14 is encoded as 'always' 2.
EV 15 is encoded as 'always' 1.
EV 16 is encoded as about 50% 0's and 50% 1's
EV 17 is encoded as about 90 % 0's and 10% 1's
EV 18 is not shown, but encoded as 'always" 0


Edited on Apr 14, 2015 at 03:22 AM · View previous versions
Apr 14, 2015 at 01:42 AM
therealthings
Offline
Details
p.5 #7High DR for 5d4
Is there someone to write this all down in Layman's terms?
Apr 14, 2015 at 02:23 AM
alundeb
Offline
Details
p.5 #8High DR for 5d4
therealthings wrote:
Is there someone to write this all down in Layman's terms?


I am sorry, I am not able to do that. What is important to know, and relevant to the thread, is that the extended DR of current Canon sensors with the proposed method can still be encoded into the usual 14 bits RAW files, so the file size won't have to be any bigger
Apr 14, 2015 at 02:41 AM
ben egbert
Offline
Details
p.5 #9High DR for 5d4
The question about layman's terms is a good one. I get that this will improve DR, the lead title says so. Maybe high ISO as well? Maybe video?

What sort of photographer will choose this over the 5DS?

If I am a landscaper and want lots of pixels, and of course I want DR, but I have work arounds for DR, but not for more pixels, how would this fit in?
Apr 14, 2015 at 09:13 AM
snapsy
Offline
Details
p.5 #10High DR for 5d4
alundeb wrote:
snapsy wrote
alundeb wrote
snapsy wrote
alundeb wrote
curious80 wroteHmmm but unless the ADC is assigning the values in somewhat non-linear way, you just don't have enough bits to encode that many levels.
One cannot measure noise or DR for one pixel only. The signal level we use is the mean value, and the noise level is the standard deviation. As long as there is noise, the mean value or the noise level is not limited by the quantization step.
But DR is limited by one or the other (quantiziation bit-depth or noise) so the effect is the same. In other words, increasing the bit-depth but keeping noise the same just means wasted ADC resolution (Ie, oversampling noise). Conversely, keeping bit-depth the same but decreasing noise means undersampling.
But the digital SNR increases as the analog noise decreases even if the analog noise is undersampled.

The total noise is a combination of analog noise and quantization noise, and it is not a hard limit where the effect of one of them suddenly stops.
Ok, I'm still having trouble conceptualizing how the formula translates into achieving > 2^bit_depth DR. To help suss out where my understanding is falling short, here's a quick spreadsheet of a sample ADC encoding, using a D750 as an example (ISO 100 FWC 81608 with read noise of 5.5e-, with a per-pixel DR of 13.9EV):

http://horshack.smugmug.com/photos/i-RNZFTLR/0/O/i-RNZFTLR.jpg

For argument's sake lets say the D750 has exactly 1 e- of read noise instead of its actual 5.5e-, meaning it has a pixel-DR approaching 16EV. Can you walk me through how the formula would relate to the 14-bit ADC encoding for EVs 15 and 16? As far as I can comprehend right now, the minimum e- sample from a 14-bit ADC would be 5 e-, with the 1/2 LSB swing below and above that, bringing the total DR to the approx 14.33EV. Can you walk me through how the formula's encoding of the quantum bit (LSB) brings the representable DR up to 15.8EV?
There are two things that can help us conceptualize how DR > 2^N.

One is that the threshold between ADU 0 and 1 in the conversion corresponds to 1/2 ADU, not 1 ADU. Analog values corresponding to between -1/2 and 1/2 ADU are represented as 0, values between 1/2 and 3/2 ADU are represented as 1, and so on.

Using your example, we have a table showing what the ADU values represent in terms of minimum, average and maximum values of e-:

The other thing is that noise has a probability density function that extends beyond the standard deviation. If we use gaussian
...Show more →

In your final chart your "average" e- value is actually the max e- value for each tier rather than the average. Why is that? That allowed you to half the average toward the quantum end of the e- scale to include EV15 into the detectable e- range but that doesn't seem accurate. Based on my original scale for a 1 e- D750 on a 14-bit ADC, the lowest measurable signature at EV14 would be 5 e- +/- stddev.
Apr 14, 2015 at 09:18 AM
alundeb
Offline
Details
p.5 #11High DR for 5d4
snapsy wrote:
alundeb wrote
snapsy wrote
alundeb wrote
snapsy wrote
alundeb wroteOne cannot measure noise or DR for one pixel only. The signal level we use is the mean value, and the noise level is the standard deviation. As long as there is noise, the mean value or the noise level is not limited by the quantization step.
But DR is limited by one or the other (quantiziation bit-depth or noise) so the effect is the same. In other words, increasing the bit-depth but keeping noise the same just means wasted ADC resolution (Ie, oversampling noise). Conversely, keeping bit-depth the same but decreasing noise means undersampling.
But the digital SNR increases as the analog noise decreases even if the analog noise is undersampled.

The total noise is a combination of analog noise and quantization noise, and it is not a hard limit where the effect of one of them suddenly stops.
Ok, I'm still having trouble conceptualizing how the formula translates into achieving > 2^bit_depth DR. To help suss out where my understanding is falling short, here's a quick spreadsheet of a sample ADC encoding, using a D750 as an example (ISO 100 FWC 81608 with read noise of 5.5e-, with a per-pixel DR of 13.9EV):

http://horshack.smugmug.com/photos/i-RNZFTLR/0/O/i-RNZFTLR.jpg

For argument's sake lets say the D750 has exactly 1 e- of read noise instead of its actual 5.5e-, meaning it has a pixel-DR approaching 16EV. Can you walk me through how the formula would relate to the 14-bit ADC encoding for EVs 15 and 16? As far as I can comprehend right now, the minimum e- sample from a 14-bit ADC would be 5 e-, with the 1/2 LSB swing below and above that, bringing the total DR to the approx 14.33EV. Can you walk me through how the formula's encoding of the quantum bit (LSB) brings the representable DR up to 15.8EV?
There are two things that can help us conceptualize how DR > 2^N.

One is that the threshold between ADU 0 and 1 in the conversion corresponds to 1/2 ADU, not 1 ADU. Analog values corresponding to between -1/2 and 1/2 ADU are represented as 0, values between 1/2 and 3/2 ADU are represented as 1, and so on.

Using your example, we have a table showing what the ADU values represent in terms of minimum, average and maximum values of e-:

The other thing is that noise has a probability density function that extends beyond the standard deviation. If we use gaussian noise with standard deviation 1 e- in this example, the approximate values for the probability are:

http://i306.photobucket.com/albums/nn242/Overflate/Gauss_1ev_zpsxrdanxhy.png

Finally if we add this 1 e- of gaussian read noise to the signal with fwc 81608 e-, and run it through an ideal A/D converter, the probabilities of getting certain ADU values for a given signal level are given in the following table. What I should have done is to calculate the exact probability for each ADC value, so the table here is not accurate , but it shows the concept.

Edit: I forgot to normalize the noise to ADU. The last table is a bit different, corrected now.

EV 1 is encoded as 'always' 16383
....
EV 14 is encoded as 'always' 2.
EV 15 is encoded as 'always' 1.
EV 16 is encoded as about 50% 0's and 50% 1's
EV 17 is encoded as about 90 % 0's and 10% 1's
EV 18 is not shown, but encoded as 'always" 0
In your final chart your "average" e- value is actually the max e- value for each tier rather than the average. Why is that? That allowed you to half the average toward the quantum end of the e- scale to include EV15 into the detectable e- range but that doesn't seem accurate. Based on my original scale for a 1 e- D750 on a 14-bit ADC, the lowest measurable signature at EV14 would be 5 e- +/- stddev.


It is not correct to use tiers for the top end to measure DR as you suggest. The top end of the DR is the maximum value and not the range between 1/2 FS and FS. When you measure DR as saturation divided by read noise, what value do you use for saturation?
Apr 14, 2015 at 09:26 AM
snapsy
Offline
Details
p.5 #12High DR for 5d4
alundeb wrote:
snapsy wrote
alundeb wrote
snapsy wrote
alundeb wrote
snapsy wroteBut DR is limited by one or the other (quantiziation bit-depth or noise) so the effect is the same. In other words, increasing the bit-depth but keeping noise the same just means wasted ADC resolution (Ie, oversampling noise). Conversely, keeping bit-depth the same but decreasing noise means undersampling.
But the digital SNR increases as the analog noise decreases even if the analog noise is undersampled.

The total noise is a combination of analog noise and quantization noise, and it is not a hard limit where the effect of one of them suddenly stops.
Ok, I'm still having trouble conceptualizing how the formula translates into achieving > 2^bit_depth DR. To help suss out where my understanding is falling short, here's a quick spreadsheet of a sample ADC encoding, using a D750 as an example (ISO 100 FWC 81608 with read noise of 5.5e-, with a per-pixel DR of 13.9EV):

http://horshack.smugmug.com/photos/i-RNZFTLR/0/O/i-RNZFTLR.jpg

For argument's sake lets say the D750 has exactly 1 e- of read noise instead of its actual 5.5e-, meaning it has a pixel-DR approaching 16EV. Can you walk me through how the formula would relate to the 14-bit ADC encoding for EVs 15 and 16? As far as I can comprehend right now, the minimum e- sample from a 14-bit ADC would be 5 e-, with the 1/2 LSB swing below and above that, bringing the total DR to the approx 14.33EV. Can you walk me through how the formula's encoding of the quantum bit (LSB) brings the representable DR up to 15.8EV?
There are two things that can help us conceptualize how DR > 2^N.

One is that the threshold between ADU 0 and 1 in the conversion corresponds to 1/2 ADU, not 1 ADU. Analog values corresponding to between -1/2 and 1/2 ADU are represented as 0, values between 1/2 and 3/2 ADU are represented as 1, and so on.

Using your example, we have a table showing what the ADU values represent in terms of minimum, average and maximum values of e-:

The other thing is that noise has a probability density function that extends beyond the standard deviation. If we use gaussian noise with standard deviation 1 e- in this example, the approximate values for the probability are:

http://i306.photobucket.com/albums/nn242/Overflate/Gauss_1ev_zpsxrdanxhy.png

Finally if we add this 1 e- of gaussian read noise to the signal with fwc 81608 e-, and run it through an ideal A/D converter, the probabilities of getting certain ADU values for a given signal level are given in the following table. What I should have done is to calculate the exact probability for each ADC value, so the table here is not accurate , but it shows the concept.

Edit: I forgot to normalize the noise to ADU. The last table is a bit different, corrected now.

EV 1 is encoded as 'always' 16383
....
EV 14 is encoded as 'always' 2.
EV 15 is encoded as 'always' 1.
EV 16 is encoded as about 50% 0's and 50% 1's
EV 17 is encoded as about 90 % 0's and 10% 1's
EV 18 is not shown, but encoded as 'always" 0
In your final chart your "average" e- value is actually the max e- value for each tier rather than the average. Why is that? That allowed you to half the average toward the quantum end of the e- scale to include EV15 into the detectable e- range but that doesn't seem accurate. Based on my original scale for a 1 e- D750 on a 14-bit ADC, the lowest measurable signature at EV14 would be 5 e- +/- stddev.
It is not correct to use tiers for the top end to measure DR as you suggest. The top end of the DR is the maximum value and not the range between 1/2 FS and FS. When you measure DR as saturation divided by read noise, what value do you use for saturation?


I'm not following you. Are you saying the top tier bit of ADU coding is dedicated only to full e- saturation values?
Apr 14, 2015 at 09:42 AM
alundeb
Offline
Details
p.5 #13High DR for 5d4
snapsy wrote:
alundeb wrote
snapsy wrote
alundeb wrote
snapsy wrote
alundeb wroteBut the digital SNR increases as the analog noise decreases even if the analog noise is undersampled.

The total noise is a combination of analog noise and quantization noise, and it is not a hard limit where the effect of one of them suddenly stops.
Ok, I'm still having trouble conceptualizing how the formula translates into achieving > 2^bit_depth DR. To help suss out where my understanding is falling short, here's a quick spreadsheet of a sample ADC encoding, using a D750 as an example (ISO 100 FWC 81608 with read noise of 5.5e-, with a per-pixel DR of 13.9EV):

http://horshack.smugmug.com/photos/i-RNZFTLR/0/O/i-RNZFTLR.jpg

For argument's sake lets say the D750 has exactly 1 e- of read noise instead of its actual 5.5e-, meaning it has a pixel-DR approaching 16EV. Can you walk me through how the formula would relate to the 14-bit ADC encoding for EVs 15 and 16? As far as I can comprehend right now, the minimum e- sample from a 14-bit ADC would be 5 e-, with the 1/2 LSB swing below and above that, bringing the total DR to the approx 14.33EV. Can you walk me through how the formula's encoding of the quantum bit (LSB) brings the representable DR up to 15.8EV?
There are two things that can help us conceptualize how DR > 2^N.

One is that the threshold between ADU 0 and 1 in the conversion corresponds to 1/2 ADU, not 1 ADU. Analog values corresponding to between -1/2 and 1/2 ADU are represented as 0, values between 1/2 and 3/2 ADU are represented as 1, and so on.

Using your example, we have a table showing what the ADU values represent in terms of minimum, average and maximum values of e-:

The other thing is that noise has a probability density function that extends beyond the standard deviation. If we use gaussian noise with standard deviation 1 e- in this example, the approximate values for the probability are:

http://i306.photobucket.com/albums/nn242/Overflate/Gauss_1ev_zpsxrdanxhy.png

Finally if we add this 1 e- of gaussian read noise to the signal with fwc 81608 e-, and run it through an ideal A/D converter, the probabilities of getting certain ADU values for a given signal level are given in the following table. What I should have done is to calculate the exact probability for each ADC value, so the table here is not accurate , but it shows the concept.

Edit: I forgot to normalize the noise to ADU. The last table is a bit different, corrected now.

EV 1 is encoded as 'always' 16383
....
EV 14 is encoded as 'always' 2.
EV 15 is encoded as 'always' 1.
EV 16 is encoded as about 50% 0's and 50% 1's
EV 17 is encoded as about 90 % 0's and 10% 1's
EV 18 is not shown, but encoded as 'always" 0
In your final chart your "average" e- value is actually the max e- value for each tier rather than the average. Why is that? That allowed you to half the average toward the quantum end of the e- scale to include EV15 into the detectable e- range but that doesn't seem accurate. Based on my original scale for a 1 e- D750 on a 14-bit ADC, the lowest measurable signature at EV14 would be 5 e- +/- stddev.
It is not correct to use tiers for the top end to measure DR as you suggest. The top end of the DR is the maximum value and not the range between 1/2 FS and FS. When you measure DR as saturation divided by read noise, what value do you use for saturation?
I'm not following you. Are you saying the top tier bit of ADU coding is dedicated only to full e- saturation values?


No, I'm saying that DR is measured as a single value (Full Scale) vs another single value (the value where SNR = 1). Using bits, and tiers between bits, only confuses things and have no direct place in a SNR / DR analysis.
Apr 14, 2015 at 09:50 AM
snapsy
Offline
Details
p.5 #14High DR for 5d4
alundeb wrote:
snapsy wrote
alundeb wrote
snapsy wrote
alundeb wrote
snapsy wroteOk, I'm still having trouble conceptualizing how the formula translates into achieving > 2^bit_depth DR. To help suss out where my understanding is falling short, here's a quick spreadsheet of a sample ADC encoding, using a D750 as an example (ISO 100 FWC 81608 with read noise of 5.5e-, with a per-pixel DR of 13.9EV):

http://horshack.smugmug.com/photos/i-RNZFTLR/0/O/i-RNZFTLR.jpg

For argument's sake lets say the D750 has exactly 1 e- of read noise instead of its actual 5.5e-, meaning it has a pixel-DR approaching 16EV. Can you walk me through how the formula would relate to the 14-bit ADC encoding for EVs 15 and 16? As far as I can comprehend right now, the minimum e- sample from a 14-bit ADC would be 5 e-, with the 1/2 LSB swing below and above that, bringing the total DR to the approx 14.33EV. Can you walk me through how the formula's encoding of the quantum bit (LSB) brings the representable DR up to 15.8EV?
There are two things that can help us conceptualize how DR > 2^N.

One is that the threshold between ADU 0 and 1 in the conversion corresponds to 1/2 ADU, not 1 ADU. Analog values corresponding to between -1/2 and 1/2 ADU are represented as 0, values between 1/2 and 3/2 ADU are represented as 1, and so on.

Using your example, we have a table showing what the ADU values represent in terms of minimum, average and maximum values of e-:

The other thing is that noise has a probability density function that extends beyond the standard deviation. If we use gaussian noise with standard deviation 1 e- in this example, the approximate values for the probability are:

http://i306.photobucket.com/albums/nn242/Overflate/Gauss_1ev_zpsxrdanxhy.png

Finally if we add this 1 e- of gaussian read noise to the signal with fwc 81608 e-, and run it through an ideal A/D converter, the probabilities of getting certain ADU values for a given signal level are given in the following table. What I should have done is to calculate the exact probability for each ADC value, so the table here is not accurate , but it shows the concept.

Edit: I forgot to normalize the noise to ADU. The last table is a bit different, corrected now.

EV 1 is encoded as 'always' 16383
....
EV 14 is encoded as 'always' 2.
EV 15 is encoded as 'always' 1.
EV 16 is encoded as about 50% 0's and 50% 1's
EV 17 is encoded as about 90 % 0's and 10% 1's
EV 18 is not shown, but encoded as 'always" 0
In your final chart your "average" e- value is actually the max e- value for each tier rather than the average. Why is that? That allowed you to half the average toward the quantum end of the e- scale to include EV15 into the detectable e- range but that doesn't seem accurate. Based on my original scale for a 1 e- D750 on a 14-bit ADC, the lowest measurable signature at EV14 would be 5 e- +/- stddev.
It is not correct to use tiers for the top end to measure DR as you suggest. The top end of the DR is the maximum value and not the range between 1/2 FS and FS. When you measure DR as saturation divided by read noise, what value do you use for saturation?
I'm not following you. Are you saying the top tier bit of ADU coding is dedicated only to full e- saturation values?
No, I'm saying that DR is measured as a single value (Full Scale) vs another single value (the value where SNR = 1). Using bits, and tiers between bits, only confuses things and have no direct place in a SNR / DR analysis.

The ADC is a linear device that hashes ranges of analog input measurements into discrete output measurements. So it operates on ranges of measurements for input and tiers for output. The only difference is that the input values are subject to noise deviations. DR is saturation to stddev...why wouldn't an ADC attached to a particular sensor be tuned specifically for that sensor's FWC of e- for its input range?
Apr 14, 2015 at 09:54 AM
alundeb
Offline
Details
p.5 #15High DR for 5d4
snapsy wrote:
alundeb wrote
snapsy wrote
alundeb wrote
snapsy wrote
alundeb wroteThere are two things that can help us conceptualize how DR > 2^N.

One is that the threshold between ADU 0 and 1 in the conversion corresponds to 1/2 ADU, not 1 ADU. Analog values corresponding to between -1/2 and 1/2 ADU are represented as 0, values between 1/2 and 3/2 ADU are represented as 1, and so on.

Using your example, we have a table showing what the ADU values represent in terms of minimum, average and maximum values of e-:

The other thing is that noise has a probability density function that extends beyond the standard deviation. If we use gaussian noise with standard deviation 1 e- in this example, the approximate values for the probability are:

http://i306.photobucket.com/albums/nn242/Overflate/Gauss_1ev_zpsxrdanxhy.png

Finally if we add this 1 e- of gaussian read noise to the signal with fwc 81608 e-, and run it through an ideal A/D converter, the probabilities of getting certain ADU values for a given signal level are given in the following table. What I should have done is to calculate the exact probability for each ADC value, so the table here is not accurate , but it shows the concept.

Edit: I forgot to normalize the noise to ADU. The last table is a bit different, corrected now.

EV 1 is encoded as 'always' 16383
....
EV 14 is encoded as 'always' 2.
EV 15 is encoded as 'always' 1.
EV 16 is encoded as about 50% 0's and 50% 1's
EV 17 is encoded as about 90 % 0's and 10% 1's
EV 18 is not shown, but encoded as 'always" 0
In your final chart your "average" e- value is actually the max e- value for each tier rather than the average. Why is that? That allowed you to half the average toward the quantum end of the e- scale to include EV15 into the detectable e- range but that doesn't seem accurate. Based on my original scale for a 1 e- D750 on a 14-bit ADC, the lowest measurable signature at EV14 would be 5 e- +/- stddev.
It is not correct to use tiers for the top end to measure DR as you suggest. The top end of the DR is the maximum value and not the range between 1/2 FS and FS. When you measure DR as saturation divided by read noise, what value do you use for saturation?
I'm not following you. Are you saying the top tier bit of ADU coding is dedicated only to full e- saturation values?
No, I'm saying that DR is measured as a single value (Full Scale) vs another single value (the value where SNR = 1). Using bits, and tiers between bits, only confuses things and have no direct place in a SNR / DR analysis.
The ADC is a linear device that hashes ranges of analog input measurements into discrete output measurements. So it operates on ranges of measurements for input and tiers for output. The only difference is that the input values are subject to noise deviations. DR is saturation to stddev...why wouldn't an ADC attached to a particular sensor be tuned specifically for that sensor's FWC of e- for its input range?


Now I'm not following you.

But yes, the ADC is a linear device, and operates with tiers of fixed size 1 ADU if you want to put it that way. The ADC is tuned to the sensor's FWC, represented by the tier between 16382 and 16383.
Apr 14, 2015 at 09:59 AM
snapsy
Offline
Details
p.5 #16High DR for 5d4
alundeb wrote:
snapsy wrote
alundeb wrote
snapsy wrote
alundeb wrote
snapsy wroteIn your final chart your "average" e- value is actually the max e- value for each tier rather than the average. Why is that? That allowed you to half the average toward the quantum end of the e- scale to include EV15 into the detectable e- range but that doesn't seem accurate. Based on my original scale for a 1 e- D750 on a 14-bit ADC, the lowest measurable signature at EV14 would be 5 e- +/- stddev.
It is not correct to use tiers for the top end to measure DR as you suggest. The top end of the DR is the maximum value and not the range between 1/2 FS and FS. When you measure DR as saturation divided by read noise, what value do you use for saturation?
I'm not following you. Are you saying the top tier bit of ADU coding is dedicated only to full e- saturation values?
No, I'm saying that DR is measured as a single value (Full Scale) vs another single value (the value where SNR = 1). Using bits, and tiers between bits, only confuses things and have no direct place in a SNR / DR analysis.
The ADC is a linear device that hashes ranges of analog input measurements into discrete output measurements. So it operates on ranges of measurements for input and tiers for output. The only difference is that the input values are subject to noise deviations. DR is saturation to stddev...why wouldn't an ADC attached to a particular sensor be tuned specifically for that sensor's FWC of e- for its input range?
Now I'm not following you.

But yes, the ADC is a linear device, and operates with tiers of fixed size 1 ADU if you want to put it that way. The ADC is tuned to the sensor's FWC, represented by the tier between 16382 and 16383.


Maybe we're looking at tiers from opposite end. The top tier of this ADC near FWC would capture e- values from 40,804 through 81,607. In your chart you list the average e- value of the top tier as 81608, leaving the entire MSB of the ADC's encoding dedicated to it.
Apr 14, 2015 at 10:04 AM
alundeb
Offline
Details
p.5 #17High DR for 5d4
snapsy wrote:
alundeb wrote
snapsy wrote
alundeb wrote
snapsy wrote
alundeb wroteIt is not correct to use tiers for the top end to measure DR as you suggest. The top end of the DR is the maximum value and not the range between 1/2 FS and FS. When you measure DR as saturation divided by read noise, what value do you use for saturation?
I'm not following you. Are you saying the top tier bit of ADU coding is dedicated only to full e- saturation values?
No, I'm saying that DR is measured as a single value (Full Scale) vs another single value (the value where SNR = 1). Using bits, and tiers between bits, only confuses things and have no direct place in a SNR / DR analysis.
The ADC is a linear device that hashes ranges of analog input measurements into discrete output measurements. So it operates on ranges of measurements for input and tiers for output. The only difference is that the input values are subject to noise deviations. DR is saturation to stddev...why wouldn't an ADC attached to a particular sensor be tuned specifically for that sensor's FWC of e- for its input range?
Now I'm not following you.

But yes, the ADC is a linear device, and operates with tiers of fixed size 1 ADU if you want to put it that way. The ADC is tuned to the sensor's FWC, represented by the tier between 16382 and 16383.
Maybe we're looking at tiers from opposite end. The top tier of this ADC near FWC would capture e- values from 40,804 through 81,607. In your chart you list the average e- value of the top tier as 81608, leaving the entire MSB of the ADC's encoding dedicated to it.


No, the MSB (binary 10000000000000) represents 40804.
81608 is represented by binary 11111111111111, or the sum of the values represented by all bits.
Apr 14, 2015 at 10:08 AM
snapsy
Offline
Details
p.5 #18High DR for 5d4
alundeb wrote:
snapsy wrote
alundeb wrote
snapsy wrote
alundeb wrote
snapsy wroteI'm not following you. Are you saying the top tier bit of ADU coding is dedicated only to full e- saturation values?
No, I'm saying that DR is measured as a single value (Full Scale) vs another single value (the value where SNR = 1). Using bits, and tiers between bits, only confuses things and have no direct place in a SNR / DR analysis.
The ADC is a linear device that hashes ranges of analog input measurements into discrete output measurements. So it operates on ranges of measurements for input and tiers for output. The only difference is that the input values are subject to noise deviations. DR is saturation to stddev...why wouldn't an ADC attached to a particular sensor be tuned specifically for that sensor's FWC of e- for its input range?
Now I'm not following you.

But yes, the ADC is a linear device, and operates with tiers of fixed size 1 ADU if you want to put it that way. The ADC is tuned to the sensor's FWC, represented by the tier between 16382 and 16383.
Maybe we're looking at tiers from opposite end. The top tier of this ADC near FWC would capture e- values from 40,804 through 81,607. In your chart you list the average e- value of the top tier as 81608, leaving the entire MSB of the ADC's encoding dedicated to it.
No, the MSB (binary 10000000000000) represents 40804.
81608 is represented by binary 11111111111111, or the sum of the values represented by all bits.


There are two ranges - the input analog range of e- (fractions of volts specifically) and the output range of digital ADUs. The MSB relates to the digital output of the ADC (ADUs), not the input (e-).
Apr 14, 2015 at 10:11 AM
alundeb
Offline
Details
p.5 #19High DR for 5d4
snapsy wrote:
alundeb wrote
snapsy wrote
alundeb wrote
snapsy wrote
alundeb wroteNo, I'm saying that DR is measured as a single value (Full Scale) vs another single value (the value where SNR = 1). Using bits, and tiers between bits, only confuses things and have no direct place in a SNR / DR analysis.
The ADC is a linear device that hashes ranges of analog input measurements into discrete output measurements. So it operates on ranges of measurements for input and tiers for output. The only difference is that the input values are subject to noise deviations. DR is saturation to stddev...why wouldn't an ADC attached to a particular sensor be tuned specifically for that sensor's FWC of e- for its input range?
Now I'm not following you.

But yes, the ADC is a linear device, and operates with tiers of fixed size 1 ADU if you want to put it that way. The ADC is tuned to the sensor's FWC, represented by the tier between 16382 and 16383.
Maybe we're looking at tiers from opposite end. The top tier of this ADC near FWC would capture e- values from 40,804 through 81,607. In your chart you list the average e- value of the top tier as 81608, leaving the entire MSB of the ADC's encoding dedicated to it.
No, the MSB (binary 10000000000000) represents 40804.
81608 is represented by binary 11111111111111, or the sum of the values represented by all bits.
There are two ranges - the input analog range of e- and the output range of digital ADUs. The MSB relates to the digital output of the ADC (ADUs), not the input (e-).


But there is a linear relationship between the input range and the output range, as shown in my first table.

I really don't understand why you call the range between MSB and FS a tier.
Apr 14, 2015 at 10:19 AM
snapsy
Offline
Details
p.5 #20High DR for 5d4
alundeb wrote:
snapsy wrote
alundeb wrote
snapsy wrote
alundeb wrote
snapsy wroteThe ADC is a linear device that hashes ranges of analog input measurements into discrete output measurements. So it operates on ranges of measurements for input and tiers for output. The only difference is that the input values are subject to noise deviations. DR is saturation to stddev...why wouldn't an ADC attached to a particular sensor be tuned specifically for that sensor's FWC of e- for its input range?
Now I'm not following you.

But yes, the ADC is a linear device, and operates with tiers of fixed size 1 ADU if you want to put it that way. The ADC is tuned to the sensor's FWC, represented by the tier between 16382 and 16383.
Maybe we're looking at tiers from opposite end. The top tier of this ADC near FWC would capture e- values from 40,804 through 81,607. In your chart you list the average e- value of the top tier as 81608, leaving the entire MSB of the ADC's encoding dedicated to it.
No, the MSB (binary 10000000000000) represents 40804.
81608 is represented by binary 11111111111111, or the sum of the values represented by all bits.
There are two ranges - the input analog range of e- and the output range of digital ADUs. The MSB relates to the digital output of the ADC (ADUs), not the input (e-).
But there is a linear relationship between the input range and the output range, as shown in my first table.

I really don't understand why you call the range between MSB and FS a tier.


Yes, there is a linear relationship between the but the ADC is hashing a larger universe of analog e- values into a smaller universe of digital ADU values. Each doubling of e- values within the ADC's programmed input range correlates to the next MSB used in its output range (ie, a tier)...but the actual encoding of the bits correspond to ADU values (0..2^14-1) and not directly to e- values.
Apr 14, 2015 at 10:28 AM
FM ForumsCanon ForumJoin Upload & Sell
  
 

You are not logged inLog inRegister