Getting down to the nitty-gritty about noise and it's effect on IQ Locked

Started Apr 15, 2010 | Discussions
This thread is locked.
Crocodile Gena Senior Member • Posts: 1,017
Getting down to the nitty-gritty about noise and it's effect on IQ

It's not a secret that if we compare 100% crops from two equally efficient sensors with the same format (sensor size) but different pixel counts (pixel sizes), that the crop from the sensor with more pixels will appear more noisy.

However, we are comparing different levels of magnification when we make such a comparison. Much more useful, in terms of delivered IQ, is if we compare at the same scene as opposed to the same number of pixels, as this will relate to the appearance of the final image.

To that end, let's consider the following samples:

Does it really make sense to call the 50D pics "more noisy"? Or, instead, does it make more sense to say the noise is basically the same, but the 40D pics are simply "more blurry"?

In my estimation, the latter description seems to more accurately represent the images.

Noise is a vector. That is, it has two components: amplitude and frequency. The greater the standard deviation of the recorded signal from the true (mean) signal, the greater the amplitude. The greater the number of samples (pixels), the greater the frequency.

I hope we can all agree that a greater amplitude of noise is undesireable, whereas a greater frequency of noise is desireable. The problem comes from comparing the amplitudes of noise at different frequencies. If one image has a greater amplitude of noise at a higher frequency than another image which has a lower amplitude of noise at a lower frequency (the case for the images above), does that make it "more noisy"?

In my opinion, the image that most accurately represents the scene is the less noisy of the two, and is consistent with one defintion of noise that means "unwanted signal".

Along these lines, I would argue that if the image with the higher pixel count can be downsampled (and/or NR can be applied) so that maximum frequency of noise matches that of the lower pixel count image, and the amplitude of the downsampled/NR image matches, or is less than, the amplitude of noise of the image with the lower pixel count, then the image with the greater pixel count is not "more noisy".

In other words, using scalar descriptors ("more" / "less") to describe noise in scalar (single-valued) terms makes sense only if the maximum frequencies are the same. But if we are comparing vector quantities (amplitude / frequency), such a simplistic comparison does not accurately relate to the IQ of the final image.

I'd be pleased to hear the opinions of those that have an interest in this topic.

tko Forum Pro • Posts: 12,796
as we all know

More MP make your lenses less sharp. More MP increases diffraction. And more MP increases noise. More MP are bad

Ultimately, a one pixel image is the only option. None of the other flaws in the system are revealed.

pocketfulladoubles Senior Member • Posts: 1,986
Re: as we all know

I'm an engineer not a mathemitician, so I'm treading on the limits of my knowledge, but I believe if you are going to show this is true, you would need to first develop a functional between your hypothesized vector space and the scalars of human perception. This would require a lot of testing. Then you would need to find the function that minimizes your functional. And, with my one year of calculus of variations and functional analysis, I have no idea how you'd accomplish that.

Joofa Senior Member • Posts: 2,655
Re: Getting down to the nitty-gritty about noise and it's effect on IQ

Crocodile Gena wrote:

I hope we can all agree that a greater amplitude of noise is undesireable, whereas a greater frequency of noise is desireable. The problem comes from comparing the amplitudes of noise at different frequencies. If one image has a greater amplitude of noise at a higher frequency than another image which has a lower amplitude of noise at a lower frequency (the case for the images above), does that make it "more noisy"?

Comparing noise amplitude at different frequencies is not well-defined. What you need to compare here is the noise power at different frequencies. Two different things. And, then you have to have a methodology/assumption of going from the non-stationarity of noise in natural images to stationary noise if you are going to talk more meaningful things about power spectrum.

In other words, using scalar descriptors ("more" / "less") to describe noise in scalar (single-valued) terms makes sense only if the maximum frequencies are the same. But if we are comparing vector quantities (amplitude / frequency), such a simplistic comparison does not accurately relate to the IQ of the final image.

There is no problem with the scalar/vector issue here. As the scalar quantity in this case can be obtained from the integral/summation of the vector quantity along the frequency variable.

Joofa

-- hide signature --

Dj Joofa

Steen Bay Veteran Member • Posts: 7,418
Total noise power

Crocodile Gena wrote:

Noise is a vector. That is, it has two components: amplitude and frequency. The greater the standard deviation of the recorded signal from the true (mean) signal, the greater the amplitude. The greater the number of samples (pixels), the greater the frequency.

I hope we can all agree that a greater amplitude of noise is undesireable, whereas a greater frequency of noise is desireable. The problem comes from comparing the amplitudes of noise at different frequencies. If one image has a greater amplitude of noise at a higher frequency than another image which has a lower amplitude of noise at a lower frequency (the case for the images above), does that make it "more noisy"?

I'd say yes, a higher resolution image has more noise. A 20mp FF image has 41% higher pixel-level (shot/photon) noise than a 10mp FF image, and the 20mp image has also 41% higher 'total noise power', like explained by ejmartin in the link below. The 20mp image has the same noise at the shared frequencies as the 10mp image, plus the additional noise from its higher frequencies, and the fact that the 'total noise power' always is directly proportional to the pixel-level noise means that the pixel-level noise actually can be seen as a direct measure for the sum of noises at all frequencies in the image, and not just the amplitude of noise at Nyquist.

http://forums.dpreview.com/forums/read.asp?forum=1019&message=33180640

Joofa Senior Member • Posts: 2,655
Re: Total noise power

Steen Bay wrote:

I'd say yes, a higher resolution image has more noise.

IMHO, you are conflating a number of issues here. If the noise power spectrum reaches out to higher frequencies in a higher resolution image that does not necessarily mean that a given pixel in a higher MP Image has more noise. Consider this: If all other variables are the same then lets assume that a pixel in the lower MP image has a mean signal of N. Then the "usual" notion of Poisson noise power would be N (SNR=sqrt(N)). However, if we assume a set of 2x2 pixels in place here and assume that the mean signal in the smaller pixel is now N/4, then the noise power is N/4 (SNR = sqrt (N/4). So the actual average noise value on the smaller pixel (higher MP image) is smaller than a bigger pixel in the lower MP image, but its SNR is poorer, of course.

So what do you mean then that the "higher resolution image has more noise?"

The 20mp image has the same noise at the shared frequencies as the 10mp image, plus the additional noise from its higher frequencies,

Using the numbers from above under the assumption of stationary white noise the power spectrum for higher MP has a flat value of N/4, where as the power spectrum of lower MP image has a flat value of N, then how do you claim that in the shared range of frequencies the noise is the "same", whatever that means, which is not entirely clear to me as I mentioned above.

http://forums.dpreview.com/forums/read.asp?forum=1019&message=33180640

In the link above, I think that Emil Martinec has made some simplifying assumptions such as a flat field and the stationarity of noise, so he has to justify how to go from there to the images of natural scenes.

Joofa

-- hide signature --

Dj Joofa

pocketfulladoubles Senior Member • Posts: 1,986
Re: Total noise power

PSD isn't going to change if the total noise power doesn't change. Twice the resolution seems like -3dB per bin for a random signal to me, no? Are we even sure that we have a pure random signal? I know that's the theory, but what's the reality? I'd like to see some FFT spectra showing a flat line with no tones.

In the end, you're just defining methods to quantify noise, but this has nothing to do with determining a minimum function of noise perception.

You're also assuming that noise perception is a single variable from good to bad. What about human perception of noise as a function of hue, saturation and brightness? In other words, are we more forgiving of noise in bright, high contrast, blue areas, and less so in flat shadowed brown areas?

This whole concept of noise magnitude, save for the technical thought experiment, seems ill-posed.

Joofa Senior Member • Posts: 2,655
Re: Total noise power

pocketfulladoubles wrote:

PSD isn't going to change if the total noise power doesn't change.

Using the numbers from my example above the noise power for a pixel of higher MP count is N/4, and for lower MP count it is N. So PSD is going to change.

I'd like to see some FFT spectra showing a flat line with no tones.

That depends upon what PSD determination method is chosen.

In the end, you're just defining methods to quantify noise, but this has nothing to do with determining a minimum function of noise perception.

You're also assuming that noise perception is a single variable from good to bad. What about human perception of noise as a function of hue, saturation and brightness? In other words, are we more forgiving of noise in bright, high contrast, blue areas, and less so in flat shadowed brown areas?

I have never said anything about noise perception yet. Did I? So far my argument has all been along noise measurement. Perception of that measured quantity is a whole new field and it is not a small discussion like for e.g., PSDs.

Joofa

-- hide signature --

Dj Joofa

Paul Beagle Regular Member • Posts: 192
Re: Total noise power

Steen Bay wrote:

Crocodile Gena wrote:

Noise is a vector. That is, it has two components: amplitude and frequency. The greater the standard deviation of the recorded signal from the true (mean) signal, the greater the amplitude. The greater the number of samples (pixels), the greater the frequency.

I hope we can all agree that a greater amplitude of noise is undesireable, whereas a greater frequency of noise is desireable. The problem comes from comparing the amplitudes of noise at different frequencies. If one image has a greater amplitude of noise at a higher frequency than another image which has a lower amplitude of noise at a lower frequency (the case for the images above), does that make it "more noisy"?

I'd say yes, a higher resolution image has more noise.

Is this what you see in the samples posted by Gena?

A 20mp FF image has 41% higher pixel-level (shot/photon) noise than a 10mp FF image,

Is not the shot noise in a sample given by the square root of the number of photons in the sample, and therefore is it not smaller the smaller the pixel?

and the 20mp image has also 41% higher 'total noise power', like explained by ejmartin in the link below.

The 20mp image has the same noise at the shared frequencies as the 10mp image, plus the additional noise from its higher frequencies, and the fact that the 'total noise power' always is directly proportional to the pixel-level noise

how can that be when the pixel level noise is smaller for a smaller pixel, while the 'total noise power' is higher over the extended bandwidth made possible by smaller pixels?

means that the pixel-level noise actually can be seen as a direct measure for the sum of noises at all frequencies in the image, and not just the amplitude of noise at Nyquist.

why then are there separate measures, why not settle on just one? If we have a 10MP image and a 20MP image with half the noise power per pixel they have the same 'total noise power' but different noise power per pixel, so how can you say that pixel level noise is a direct measure of total noise power?

Does your theory accord with what you see in the images posted by CG?

What do you think is the relationship between 'total noise power' and visible noise?

Is not the comparison of noise power over different bandwidths the same as comparing noise at different scales?

Paul Beagle Regular Member • Posts: 192
Re: Total noise power

pocketfulladoubles wrote:

This whole concept of noise magnitude, save for the technical thought experiment, seems ill-posed.

Yet people want to talk about 'more' or 'less' noise. And then, what is 'magnitude'? Is it 'amplitude' or 'power' or 'spectral density' or what? How does it relate to what we perceive?

pocketfulladoubles Senior Member • Posts: 1,986
Re: Total noise power

Paul Beagle wrote:

pocketfulladoubles wrote:

This whole concept of noise magnitude, save for the technical thought experiment, seems ill-posed.

Yet people want to talk about 'more' or 'less' noise. And then, what is 'magnitude'? Is it 'amplitude' or 'power' or 'spectral density' or what? How does it relate to what we perceive?

You agree on a model with f(x1, x2, xn) variables, and hold all but one fixed at a time, and doing double-blind experiments with the question of "Is this noisy?" Find what is changing. Find what is dependant. Find a function to minimize the functional. Sounds like a whole PhD thesis to me. But this really is the point. Who cares what you can measure if your measurement has no meaning?

Paul Beagle Regular Member • Posts: 192
Re: Total noise power

pocketfulladoubles wrote:

Paul Beagle wrote:

pocketfulladoubles wrote:

This whole concept of noise magnitude, save for the technical thought experiment, seems ill-posed.

Yet people want to talk about 'more' or 'less' noise. And then, what is 'magnitude'? Is it 'amplitude' or 'power' or 'spectral density' or what? How does it relate to what we perceive?

You agree on a model with f(x1, x2, xn) variables, and hold all but one fixed at a time, and doing double-blind experiments with the question of "Is this noisy?" Find what is changing. Find what is dependant. Find a function to minimize the functional. Sounds like a whole PhD thesis to me. But this really is the point. Who cares what you can measure if your measurement has no meaning?

That's about the measure of it :). And you're right about the 'point', just so long as people say camera with 'x' have more noise, take your choice as to what 'x' is.

pocketfulladoubles Senior Member • Posts: 1,986
Re: Total noise power

Joofa wrote:

pocketfulladoubles wrote:

PSD isn't going to change if the total noise power doesn't change.

Using the numbers from my example above the noise power for a pixel of higher MP count is N/4, and for lower MP count it is N. So PSD is going to change.

I'm losing you, but I could be wrong. Just as a thought experiment, say you have some PSD function with 1 unit of some arbitrary resolution. Now take your 2x2 pixel per original pixel increase in resolution, basically you now have four times the pixels, and each pixel has 1/4 the noise power. Each bin of your FFT is now 1/4 the bandwidth, but you also have four times the probability per bin, so you're right back where you started. Hencse, the PSD has not changed. It is a density, not an absolute. It also isn't useful for non-random signals, which I'm still not sure is completely the case. It seems intuitive to me that you may have something along the way of interference patterns, moire, or repeated results being a function from the uniform layout of a digital sensor, but I'm really guessing.

Joofa Senior Member • Posts: 2,655
Re: Total noise power

pocketfulladoubles wrote:

I'm losing you, but I could be wrong. Just as a thought experiment, say you have some PSD function with 1 unit of some arbitrary resolution. Now take your 2x2 pixel per original pixel increase in resolution, basically you now have four times the pixels, and each pixel has 1/4 the noise power. Each bin of your FFT is now 1/4 the bandwidth, but you also have four times the probability per bin, so you're right back where you started. Hencse, the PSD has not changed. It is a density, not an absolute.

PSD is the Fourier transform of the autocorrelation function. Under the white noise assumption, the autocorrelation for lower MP is N*delta(t) and for higher MP it is N/4*delta(t). Now you can take the Fourier transform and see that the PSDs are different.

It also isn't useful for non-random signals, which I'm still not sure is completely the case.

The concept of PSD is perfectly valid for non-random or deterministic signals.

Joofa

-- hide signature --

Dj Joofa

OP Crocodile Gena Senior Member • Posts: 1,017
The utility of noise as a measure of IQ.

Steen Bay wrote:

I'd say yes, a higher resolution image has more noise.

So, just so I have it clear: you are saying that the 50D images that I posted in the OP (and reproduced again below) are more noisy than the 40D images, right? Next question: which images are more pleasing, the 50D images or the 40D images?

If you agree with me that the 50D images are more pleasing, then you are in the difficult position of saying that "more noise" is "better than" less noise, which goes against "common wisdom".

Thus, if we have to qualify the conditions when "more noise" is more pleasing than "less noise", is noise a useful measure for decribing IQ? Or, is it more that we cannot talk about noise in a vacuum? That is, whenever we discuss noise, we must discuss rendered detail. But, if so, then that brings us full circle to noise amplitude and frequency, so why not simply always mention both when taking about noise?

Steen Bay Veteran Member • Posts: 7,418
Noise and detail

Crocodile Gena wrote:

Steen Bay wrote:

I'd say yes, a higher resolution image has more noise.

So, just so I have it clear: you are saying that the 50D images that I posted in the OP (and reproduced again below) are more noisy than the 40D images, right? Next question: which images are more pleasing, the 50D images or the 40D images?

I said that a higher resolution image has more shot/photon noise (assuming same format, technology etc.), but I can't/won't comment on this particular comparison using ISO 1600 (?) images from IR, because there are all sorts of potential problems with it. Paul used Gaussian blur in the resampling (which affects the noise), and we don't know whether the focus is optimal in both images (which affects the detail).

Also, If we're looking at deep shadows at ISO 1600, then we're in the read noise affected/dominated area where the 50D has a relative advantage, since its read noise is much lower than the 40D's (something like 2.5 vs 4.5e- I think), and things like that (and others) becomes rather important when we're comparing images with such relatively small differences in theoretical noise levels (about 22%, or so).

If you agree with me that the 50D images are more pleasing, then you are in the difficult position of saying that "more noise" is "better than" less noise, which goes against "common wisdom".

I'm not agains more MPs, quite the contrary actually. I'm just saying that noise and detail are like two sides of the same coin. We can't get the (good) detail without also getting the (bad) noise, and whether more noise/detail is to be preferred, that'll always depend on the particular image (e.g. a blue sky won't have more detail, just more noise) and personal preferences.

Thus, if we have to qualify the conditions when "more noise" is more pleasing than "less noise", is noise a useful measure for decribing IQ? Or, is it more that we cannot talk about noise in a vacuum? That is, whenever we discuss noise, we must discuss rendered detail. But, if so, then that brings us full circle to noise amplitude and frequency, so why not simply always mention both when taking about noise?

So, we can all agree that noise and detail goes hand in hand?

pocketfulladoubles Senior Member • Posts: 1,986
Re: Total noise power

Joofa wrote:

pocketfulladoubles wrote:

I'm losing you, but I could be wrong. Just as a thought experiment, say you have some PSD function with 1 unit of some arbitrary resolution. Now take your 2x2 pixel per original pixel increase in resolution, basically you now have four times the pixels, and each pixel has 1/4 the noise power. Each bin of your FFT is now 1/4 the bandwidth, but you also have four times the probability per bin, so you're right back where you started. Hencse, the PSD has not changed. It is a density, not an absolute.

PSD is the Fourier transform of the autocorrelation function. Under the white noise assumption, the autocorrelation for lower MP is N*delta(t) and for higher MP it is N/4*delta(t). Now you can take the Fourier transform and see that the PSDs are different.

I'll try it on my signal analyzer here at work, your assumption still doesn't seem right to me. You are increasing resolution by 4, so you could expect a -6dB change per bin. However, you now have 4 times as many sites to sample and as a whole you would get +6dB in distribution. The net effect is zero. From another view, the probability distribution as a system has not changed just because you slice it up more. This is density.

In fact, do you have any SDF files from something like a 35670A that would back up your argument? Send it to me. I and another here would love to see it.

It also isn't useful for non-random signals, which I'm still not sure is completely the case.

The concept of PSD is perfectly valid for non-random or deterministic signals.

The concept is valid. However, brace yourself for misleading answers. Imagine a pure tone with a signal well above the noise floor. Say your bin bandwidth is 1 unit and the amplitude is 20dB. Now, double the resolution. Since this is a density function, your denominator is units or (units^.5) depending on how you like to view your data. Essentially, increasing resolution makes the amplitude of the tone increase. Let the resolution go to something small, and the tone get huge, but the original signal hasn't changed. So what's the right answer? They're all correct. However, and this happens to me all the time, project specs may be given in terms of PSD and engineers measure signals and report them. Depending on the resolution selected, you can get results that pass or fail. In other words, in the real world, PSD should be limited to random signals.

Joofa

-- hide signature --

Dj Joofa

Joofa Senior Member • Posts: 2,655
Re: Total noise power

pocketfulladoubles wrote:

I'll try it on my signal analyzer here at work, your assumption still doesn't seem right to me.

Hi,

You don't need a signal analyzer for this. It is simple math that can be checked by hand. May be I am wrong, but still you should be able to verify it by hand.

You are increasing resolution by 4, so you could expect a -6dB change per bin. However, you now have 4 times as many sites to sample and as a whole you would get +6dB in distribution. The net effect is zero. From another view, the probability distribution as a system has not changed just because you slice it up more. This is density.

I was taking along theoretical ensemble statistics of white noise at a given pixel and its associated PSD so regardless of frequency it should not dip/rise and have a flat line.

Imagine a pure tone with a signal well above the noise floor. Say your bin bandwidth is 1 unit and the amplitude is 20dB. Now, double the resolution. Since this is a density function, your denominator is units or (units^.5) depending on how you like to view your data. Essentially, increasing resolution makes the amplitude of the tone increase. Let the resolution go to something small, and the tone get huge, but the original signal hasn't changed. So what's the right answer? They're all correct.

The area under a PSD should give the variance of noise on a given pixel. Right? However, in our example the smaller pixel had N/4 noise power and the bigger pixel had N, based upon Poisson statistics. If the PSDs are identical everywhere in the normalized [-1/2,1/2] frequency range then they will produce the same number and will be contrary to the fact that we have different variances on the smaller and bigger pixels as mentioned before.

pocketfulladoubles wrote:

Joofa wrote

pocketfulladoubles wrote:

It also isn't useful for non-random signals, which I'm still not sure is completely the case.

The concept of PSD is perfectly valid for non-random or deterministic signals.

The concept is valid. However, brace yourself for misleading answers. ... In other words, in the real world, PSD should be limited to random signals.

Again, PSD is perfectly applicable for a non-random/deterministic signal. Please see the following link where it says that it applies to deterministic signals:
http://en.wikipedia.org/wiki/Spectral_density

The concept of PSD for a random signal is an extension of the notion of PSDs for non-random/deterministic signals. There are more complications in the convergence of PSDs for random signals, though.

Joofa

-- hide signature --

Dj Joofa

OP Crocodile Gena Senior Member • Posts: 1,017
Re: Noise and detail

Steen Bay wrote:

Crocodile Gena wrote:

So, just so I have it clear: you are saying that the 50D images that I posted in the OP (and reproduced again below) are more noisy than the 40D images, right?

Steen Bay wrote:

I said that a higher resolution image has more shot/photon noise (assuming same format, technology etc.)...

Again, the issue is the utility of the word "more" when comparing difference frequencies. Does it make sense to say that a higher amplitude of noise at a higher frequency represents "more noise" than a lower amplitude of noise at a lower frequency, when the image with the higher frequency noise can be downsampled (or NR applied) so that when the maximum frequencies are the same, the amplitude is the same (or even less)?

I would argue that the word "more" in this context is misleading as it does not accurately represent the appearance of the image.

...but I can't/won't comment on this particular comparison using ISO 1600 (?) images from IR, because there are all sorts of potential problems with it. Paul used Gaussian blur in the resampling (which affects the noise), and we don't know whether the focus is optimal in both images (which affects the detail).

Well, we have a problem then, don't we? Because all comparison images have "potential" problems. If we go that route, divorcing application from theory, then what's the point?

Now, in terms of applying a Gaussian blur before resampling, well, that's just proper resampling technique, not "cheating". In fact, a good resampling routine would do it automatically, but I don't know how the resampling routines work in various editing programs.

Also, If we're looking at deep shadows at ISO 1600, then we're in the read noise affected/dominated area where the 50D has a relative advantage, since its read noise is much lower than the 40D's (something like 2.5 vs 4.5e- I think), and things like that (and others) becomes rather important when we're comparing images with such relatively small differences in theoretical noise levels (about 22%, or so).

In terms of read noise, that's just about right. That is, we would expect that for the same technology that the read noise is proportional to the pixel area. Since the pixel area is proportional to the pixel count (for the same size sensor), then 4.5 e- x 15 MP / 10 MP = 3 e-, which is pretty much on the mark.

So, if you would, could you please answer the question about the posted images, with all the proper assumptions in place (perfectly focused, same lens, etc., etc.)? The images once again (for convenience):

The question: Does it really make sense to call the 50D pics "more noisy"? Or, instead, does it make more sense to say the noise is basically the same, but the 40D pics are simply "more blurry"?

Next question: which images are more pleasing, the 50D images or the 40D images? If you agree with me that the 50D images are more pleasing, then you are in the difficult position of saying that "more noise" is "better than" less noise, which goes against "common wisdom".

I'm not agains more MPs, quite the contrary actually. I'm just saying that noise and detail are like two sides of the same coin. We can't get the (good) detail without also getting the (bad) noise, and whether more noise/detail is to be preferred, that'll always depend on the particular image (e.g. a blue sky won't have more detail, just more noise) and personal preferences.

For sure, for a given sensor size and efficiency, we're not going to be able to pack any arbitrary amount of pixels in the image and have the same noise at the pixel level -- I said exactly that in the first paragraph of the OP.

But, as I said, when I look at the pics above, I do not see the 50D pics as "more noisy", instead, I see the 40D pics as "more blurry". Do you disagree?

Thus, if we have to qualify the conditions when "more noise" is more pleasing than "less noise", is noise a useful measure for decribing IQ? Or, is it more that we cannot talk about noise in a vacuum? That is, whenever we discuss noise, we must discuss rendered detail. But, if so, then that brings us full circle to noise amplitude and frequency, so why not simply always mention both when taking about noise?

So, we can all agree that noise and detail goes hand in hand?

We can agree that noise at the pixel level goes hand-in-hand with captured detail. But, as I keep saying, I do not agree that noise at the image level goes hand-in-hand with captured detail, as the above pics demonstrate.

However, if you were to say that, yes, the 50D pics above appear "more noisy" than the corresponding 40D pics, then I could accept that interpretation and say that we have a different interpretation of what constitutes "noise". Then, we could discuss that point.

But, as you know, I prefer to discuss photographic parameters as it applies to the visual properties of the displayed image, rather than in an academic vacuum. That's not to say that I dismiss the semantics of it all (indeed, I'm probably one of the more tenacious people in that regard), but that I also try to "keep it real".

So, if you would, with all the "proper assumptions" in place, please tell me which images appear "more noisy".

Steen Bay Veteran Member • Posts: 7,418
Re: Noise and detail

Crocodile Gena wrote:

Steen Bay wrote:

Steen Bay wrote:

I said that a higher resolution image has more shot/photon noise (assuming same format, technology etc.)...

Again, the issue is the utility of the word "more" when comparing difference frequencies. Does it make sense to say that a higher amplitude of noise at a higher frequency represents "more noise" than a lower amplitude of noise at a lower frequency, when the image with the higher frequency noise can be downsampled (or NR applied) so that when the maximum frequencies are the same, the amplitude is the same (or even less)?

I would argue that the word "more" in this context is misleading as it does not accurately represent the appearance of the image.

...but I can't/won't comment on this particular comparison using ISO 1600 (?) images from IR, because there are all sorts of potential problems with it. Paul used Gaussian blur in the resampling (which affects the noise), and we don't know whether the focus is optimal in both images (which affects the detail).

Well, we have a problem then, don't we? Because all comparison images have "potential" problems. If we go that route, divorcing application from theory, then what's the point?

Now, in terms of applying a Gaussian blur before resampling, well, that's just proper resampling technique, not "cheating". In fact, a good resampling routine would do it automatically, but I don't know how the resampling routines work in various editing programs.

Also, If we're looking at deep shadows at ISO 1600, then we're in the read noise affected/dominated area where the 50D has a relative advantage, since its read noise is much lower than the 40D's (something like 2.5 vs 4.5e- I think), and things like that (and others) becomes rather important when we're comparing images with such relatively small differences in theoretical noise levels (about 22%, or so).

In terms of read noise, that's just about right. That is, we would expect that for the same technology that the read noise is proportional to the pixel area. Since the pixel area is proportional to the pixel count (for the same size sensor), then 4.5 e- x 15 MP / 10 MP = 3 e-, which is pretty much on the mark.

So, if you would, could you please answer the question about the posted images, with all the proper assumptions in place (perfectly focused, same lens, etc., etc.)? The images once again (for convenience):

The question: Does it really make sense to call the 50D pics "more noisy"? Or, instead, does it make more sense to say the noise is basically the same, but the 40D pics are simply "more blurry"?

Next question: which images are more pleasing, the 50D images or the 40D images? If you agree with me that the 50D images are more pleasing, then you are in the difficult position of saying that "more noise" is "better than" less noise, which goes against "common wisdom".

I'm not agains more MPs, quite the contrary actually. I'm just saying that noise and detail are like two sides of the same coin. We can't get the (good) detail without also getting the (bad) noise, and whether more noise/detail is to be preferred, that'll always depend on the particular image (e.g. a blue sky won't have more detail, just more noise) and personal preferences.

For sure, for a given sensor size and efficiency, we're not going to be able to pack any arbitrary amount of pixels in the image and have the same noise at the pixel level -- I said exactly that in the first paragraph of the OP.

But, as I said, when I look at the pics above, I do not see the 50D pics as "more noisy", instead, I see the 40D pics as "more blurry". Do you disagree?

Thus, if we have to qualify the conditions when "more noise" is more pleasing than "less noise", is noise a useful measure for decribing IQ? Or, is it more that we cannot talk about noise in a vacuum? That is, whenever we discuss noise, we must discuss rendered detail. But, if so, then that brings us full circle to noise amplitude and frequency, so why not simply always mention both when taking about noise?

So, we can all agree that noise and detail goes hand in hand?

We can agree that noise at the pixel level goes hand-in-hand with captured detail. But, as I keep saying, I do not agree that noise at the image level goes hand-in-hand with captured detail, as the above pics demonstrate.

However, if you were to say that, yes, the 50D pics above appear "more noisy" than the corresponding 40D pics, then I could accept that interpretation and say that we have a different interpretation of what constitutes "noise". Then, we could discuss that point.

But, as you know, I prefer to discuss photographic parameters as it applies to the visual properties of the displayed image, rather than in an academic vacuum. That's not to say that I dismiss the semantics of it all (indeed, I'm probably one of the more tenacious people in that regard), but that I also try to "keep it real".

So, if you would, with all the "proper assumptions" in place, please tell me which images appear "more noisy".

OK, since you insist I prefer the images with most noise/detail, which are the 50D images at 15, 22 and 30mp, and the 40D image at 10mp (not sure about the 6.7mp images), but the differences are so small that a bit of sharpening probably would make it very difficult to choose. And btw, I remember that Paul once posted the same images, but without Gaussian blur, and that I preferred the more noisy/detailed version of the images in that comparison.

Keyboard shortcuts:
FForum MMy threads