Doug Kerr
Well-known member
We often hear of "quantizing noise" as one of the noise components in an overall digital imaging chain.
That term is borrowed (inappropriately and unprofitably) from another field, the digital representation of audio waveforms.
Here, for reference, is some background on that.
*********
When we represent an audio waveform in digital form, quantizing error gives errors in the individual waveform samples as reconstructed, and thus to "distortion" of the reconstructed waveform (quantizing distortion).
When the digital representation of audio waveforms was first introduced into telecommunication practice, there was need for a metric that would describe how "severe" quantizing distortion was (in terms of its perceptual impact to the listener) - one with a clear theoretical definition, and hopefully one that could be determined in an actual case with practical test instruments. (At the time, there was no computer-based or even digital measurement - most instruments involved only filters, amplifiers, maybe modulators, rectifiers, and analog meters!)
This need was met by the following train of thought:
• Expand our definition of "noise" to include anything "in" the delivered waveform that was not "in" the original waveform.
• Then "noise" is manifest by the difference between the delivered waveform and the original waveform - even if that is a result of quantizing distortion.
• If we subtract the original waveform from the delivered waveform (or remove it in some other way), we have a waveform that is the "distortion", isolated - which we treat as a special form of "noise".
• We then analyze this "noise" waveform using the same technique used to assess noise proper, which includes recognition of the differing sensitivity of the human ear to waveform components of different frequency.
We do this using "noise measuring instruments" that we already have, with the addition of a simple accessory to "remove the original waveform". (It was actually just a notch filter; we always made the test with an "original waveform" that was a sine wave at a certain frequency, which this filter just removed from the reconstructed waveform before it was fed to the "noise meter")
We found that the result of this (our "metric" for quantizing distortion) tracked pretty well with the human perception of the "badness" of quantizing distortion. And even if was not ideal in that regard, we could measure it with instruments we had widely deployed in the field. So it was adopted. And the term "quantizing noise" came into use in recognition of the back story - perhaps life imitating art.
*********
Now lets go to digital photography. What does quantizing error do to the image constructed from data from our sensor?
If we have areas of slowly varying color (color of course including luminance) we see the impact as banding. We easily recognize that this is not "natural" ("the sky never looks like that"). It's not handsome, and its an "impairment", but certainly from the viewer's standpoint, it is hard to think of it as "noise".
If we have a scene area with a lot of small detail, not forming an expected geometric pattern (think a gravel road), or even a shag rug (remember those?), the quantizing error in fact changes the image from what it "should be", but not in a way that we probably recognize by examining the image in any usual way (if at all).
So the characterization of the impact of quantizing error as noise isn't apt from the viewer's perspective.
But what about from a technical perspective? When we speak of such things as "signal to noise ratio", we mean by noise a randomly varying departure of the reported pixel color from what the sensor would "report" if there were no noise. The way we measure it is to apply a consistent photometric exposure (perhaps to a grid of sensels) and look at the variation in the digtial values (DN) the ADC reports across those pixels. (We typically use the measure "standard deviation" as the metric of that noise.) Or we might do it by taling several shots with the camera locked in position and looking at the variation in the DN reported by a single sensel.
How does quantization get in on this?
Lets think of the latter test model, with a constant photometric exposure on a certain sensel. Then in general, quantization causes no "randomly varying" discrepancy. It causes an error, but it is fixed (because the actual sensel votage is fixed - we're ignoring for now sources of "real" noise). Hardly noise-like.
Now, suppose we do this with a luminance that we vary slowly between tests. At any given stage of this, the effect of the error is fixed (and does not constitute a "random variation", nor any variation at all), but as the test luminance crosses certain boundaries, the error jumps (never being greater than half a DN number step). Perhaps we should consider the RMS error over a range of such slow change in luminance. If we feel a real need to talk about "quantizing noise", then we could consider that determination to be a measure of it. But it's not clear what that tells us. And of course the magntiude of the RMS assessment will be in the area of a fraction of a DN unit - very small.
But there is an effect of quantizing on the impact of "real" random noise. If the random noise in the usual sense (shot noise, read noise, etc), for a certain range of pixel luminance, is small compared to the size of the quantizing "step" (one unit of the data number, or DN), then for certain luminance values, the noise as observed at the digital output of the ADC will be zero; for other luminance values, the noise as observed at the digital output of the ADC will be greater than the actual random "voltage". Overall, for a random range of photometric exposures, the noise as seen in the DN output of the ADC may be on average greater than the actual noise on an analog basis going into the ADC, or it may be smaller.
But it's hard to think of this as noise caused by quantizing. (If quantizing causes the randomness of the digital output of the ADC to be less than the actual noise voltage, would that be a negative quantizing noise?)
Rather, it is a (slight) corruption of the transfer of the noise from analog form to digital form through the ADC, just as quantizing corrupts the transfer of the signal voltage from analog form to digital form through the ADC.
And it is almost negligible as soon as the RMS value of the actual noise gets as large as one quantizing step (one DN unit). At that point, the average noise (as seen out of the ADC) is about 1.125 times the actual noise voltage into the ADC.
If the RMS noise voltage into the ADC is 5 times the quantizing step (one DN unit), then the noise as seen out of the ADC (in digital form) is a little less than the noise voltage into the ADC.
*********
Overall, the bottom line is this: the concept of "quantizing noise" is a useful conceit in the matter of the digital representation of audio waveforms, since it allows us to make an assessment of the impact of quantizing distortion using concepts, and actual test instruments, developed for the measurement of actual noise.
Trying to transport that concept to the area of digital imaging does not seem to lead to anything profitable.
That term is borrowed (inappropriately and unprofitably) from another field, the digital representation of audio waveforms.
Here, for reference, is some background on that.
*********
When we represent an audio waveform in digital form, quantizing error gives errors in the individual waveform samples as reconstructed, and thus to "distortion" of the reconstructed waveform (quantizing distortion).
When the digital representation of audio waveforms was first introduced into telecommunication practice, there was need for a metric that would describe how "severe" quantizing distortion was (in terms of its perceptual impact to the listener) - one with a clear theoretical definition, and hopefully one that could be determined in an actual case with practical test instruments. (At the time, there was no computer-based or even digital measurement - most instruments involved only filters, amplifiers, maybe modulators, rectifiers, and analog meters!)
This need was met by the following train of thought:
• Expand our definition of "noise" to include anything "in" the delivered waveform that was not "in" the original waveform.
• Then "noise" is manifest by the difference between the delivered waveform and the original waveform - even if that is a result of quantizing distortion.
• If we subtract the original waveform from the delivered waveform (or remove it in some other way), we have a waveform that is the "distortion", isolated - which we treat as a special form of "noise".
• We then analyze this "noise" waveform using the same technique used to assess noise proper, which includes recognition of the differing sensitivity of the human ear to waveform components of different frequency.
We do this using "noise measuring instruments" that we already have, with the addition of a simple accessory to "remove the original waveform". (It was actually just a notch filter; we always made the test with an "original waveform" that was a sine wave at a certain frequency, which this filter just removed from the reconstructed waveform before it was fed to the "noise meter")
We found that the result of this (our "metric" for quantizing distortion) tracked pretty well with the human perception of the "badness" of quantizing distortion. And even if was not ideal in that regard, we could measure it with instruments we had widely deployed in the field. So it was adopted. And the term "quantizing noise" came into use in recognition of the back story - perhaps life imitating art.
*********
Now lets go to digital photography. What does quantizing error do to the image constructed from data from our sensor?
If we have areas of slowly varying color (color of course including luminance) we see the impact as banding. We easily recognize that this is not "natural" ("the sky never looks like that"). It's not handsome, and its an "impairment", but certainly from the viewer's standpoint, it is hard to think of it as "noise".
If we have a scene area with a lot of small detail, not forming an expected geometric pattern (think a gravel road), or even a shag rug (remember those?), the quantizing error in fact changes the image from what it "should be", but not in a way that we probably recognize by examining the image in any usual way (if at all).
So the characterization of the impact of quantizing error as noise isn't apt from the viewer's perspective.
But what about from a technical perspective? When we speak of such things as "signal to noise ratio", we mean by noise a randomly varying departure of the reported pixel color from what the sensor would "report" if there were no noise. The way we measure it is to apply a consistent photometric exposure (perhaps to a grid of sensels) and look at the variation in the digtial values (DN) the ADC reports across those pixels. (We typically use the measure "standard deviation" as the metric of that noise.) Or we might do it by taling several shots with the camera locked in position and looking at the variation in the DN reported by a single sensel.
How does quantization get in on this?
Lets think of the latter test model, with a constant photometric exposure on a certain sensel. Then in general, quantization causes no "randomly varying" discrepancy. It causes an error, but it is fixed (because the actual sensel votage is fixed - we're ignoring for now sources of "real" noise). Hardly noise-like.
Now, suppose we do this with a luminance that we vary slowly between tests. At any given stage of this, the effect of the error is fixed (and does not constitute a "random variation", nor any variation at all), but as the test luminance crosses certain boundaries, the error jumps (never being greater than half a DN number step). Perhaps we should consider the RMS error over a range of such slow change in luminance. If we feel a real need to talk about "quantizing noise", then we could consider that determination to be a measure of it. But it's not clear what that tells us. And of course the magntiude of the RMS assessment will be in the area of a fraction of a DN unit - very small.
But there is an effect of quantizing on the impact of "real" random noise. If the random noise in the usual sense (shot noise, read noise, etc), for a certain range of pixel luminance, is small compared to the size of the quantizing "step" (one unit of the data number, or DN), then for certain luminance values, the noise as observed at the digital output of the ADC will be zero; for other luminance values, the noise as observed at the digital output of the ADC will be greater than the actual random "voltage". Overall, for a random range of photometric exposures, the noise as seen in the DN output of the ADC may be on average greater than the actual noise on an analog basis going into the ADC, or it may be smaller.
But it's hard to think of this as noise caused by quantizing. (If quantizing causes the randomness of the digital output of the ADC to be less than the actual noise voltage, would that be a negative quantizing noise?)
Rather, it is a (slight) corruption of the transfer of the noise from analog form to digital form through the ADC, just as quantizing corrupts the transfer of the signal voltage from analog form to digital form through the ADC.
And it is almost negligible as soon as the RMS value of the actual noise gets as large as one quantizing step (one DN unit). At that point, the average noise (as seen out of the ADC) is about 1.125 times the actual noise voltage into the ADC.
If the RMS noise voltage into the ADC is 5 times the quantizing step (one DN unit), then the noise as seen out of the ADC (in digital form) is a little less than the noise voltage into the ADC.
*********
Overall, the bottom line is this: the concept of "quantizing noise" is a useful conceit in the matter of the digital representation of audio waveforms, since it allows us to make an assessment of the impact of quantizing distortion using concepts, and actual test instruments, developed for the measurement of actual noise.
Trying to transport that concept to the area of digital imaging does not seem to lead to anything profitable.