Will Michael revisit ETTR?

[Guillermo Luijk]

errors in SNR calculation when doing certain operations, say downsampling, as pointed out in this thread and various links provided here.

SNR changes when changing resolution, that is not an error, it's an statistical fact.

Once I measured SNR curves for several cameras:



But since they have different resolutions, I normalised the results for the resolution of one of them (in this case Canon 5D's 12,7Mpx):

[joofa]

SNR changes when changing resolution, that is not an error, it's an statistical fact.

Once I measured SNR curves for several cameras:

But since they have different resolutions, I normalised the results for the resolution of one of them (in this case Canon 5D's 12,7Mpx):

Yes, SNR changes, that is agreed. But by how much, that is the issue. That normalization is the issue. Recall, SNR means signal to noise ratio. How can you normalize SNR without incorporating any notions about signal. Right?

Joofa

[Guillermo Luijk]

Yes, SNR changes, that is agreed. But by how much, that is the issue. That normalization is the issue. Recall, SNR means signal to noise ratio. How can you normalize SNR without incorporating any notions about signal. Right?

The whole process (measuring SNR curves, normalizing,...) is about incorportating notions about signal, I don't understand your question.
In my case I used a simple criteria, but totally consistent with real world downsizing by assuming noise from different pixels adds in quadrature. See Emil's BIG PIXELS vs. small pixels.

Regards

[joofa]

The whole process (measuring SNR curves, normalizing,...) is about incorportating notions about signal, I don't understand your question.
In my case I used a simple criteria, but totally consistent with real world downsizing by assuming noise from different pixels adds in quadrature. See Emil's BIG PIXELS vs. small pixels.

Regards

Please don't pull an Emil on me  . The issue here is the signal changes from the first pixel to the last pixel in the image. So while noise might add in quadrature, but the magnitude of noise being added is dependent on the signal and that is changing in a real world image. Emil has historically shown poor recognition of this problem with his analysis being geared towards a naive assumption of flat patches. But real images are not flat patches. And, few people would appreciate, 10 20, 30, ..., or N number of SNRs figures for a single image coming from each flat patch.

Sincerely,

Joofa

[Guillermo Luijk]

But real images are not flat patches. And, few people would appreciate, 10 20, 30, ..., or N number of SNRs figures for a single image coming from each flat patch.

Once the sensor response is obtained, patches are not needed any more. On post 198 I showed a way to relate each area of an image to a SNR in that area.

[hjulenissen]

So you are the moral police here now.

In no way, where are you getting that idea from?

My original query was not even addressed to you.

Are you suggesting that any thread should have only two participants?

You jumped in and are now hoping to get me banned.

I think that is a misinterpretation of the post that you are quoting. You seem to forget the part concerning: "if you try to force other people to stop discussing a topic..." What do you think is a fair response if any user tries such a thing?

My advise to you is that if problem is too hard for you, sit back and enjoy and ride.

As far as I have seen, you have just repeated the same question over and over: how to estimate SNR from a single real-world photography. It seems to you have some absolute requirements that all real photographers or participants should adhere to.

...Since this is a photography website I would expect you to...
...Remember, we must deal with...

 I have tried to explain why I think it is impossible and fundamentally irrelevant in several different ways.

-h

[joofa]

Once the sensor response is obtained, patches are not needed any more. On post 198 I showed a way to relate each area of an image to a SNR in that area.

Yes, you are correct, and that is a way. And, I agree with that. However, as I said I don't know how many people with appreciate 10, 20, 30, ..., SNRs figures attached to a single image. IMHO, this is not a very conducive situation. We should study SNR more closely and figure out models that yield a single number that we can quote for an image with words to the effect of "this image of yours has this SNR", as opposed to "this image of yours has SNRs of 10db, 0db, 3.2 db, 4db, ....., and wait, finally 6db".

Sincerely,

Joofa

[joofa]

I have tried to explain why I think it is impossible and fundamentally irrelevant in several different ways.

Then you need to study this problem more.

Sincerely,

Joofa

[hjulenissen]

However, as I said I don't know how many people with appreciate 10, 20, 30, ..., SNRs figures attached to a single image.

Not many people are really into photography. Not many photographers choose to spend time on lula. Not many lula-ers contribute to this thread about ETTR... The fact that some topic interest very few people does not make it fundamentally uninteresting, I think.

We should study SNR more closely and figure out models that yield a single number that we can quote for an image with words to the effect of "this image of yours has this SNR", as opposed to "this image of yours has SNRs of 10db, 0db, 3.2 db, 4db, ....., and wait, finally 6db".

If the phenomenon we are studying is complex, it can probably only be described with some complexity. Of course, there may be some simpler phenomenon at its core that is obsured by our observations and understanding, but we wont know until we.. know.

Is the practical application to brag about how much better "SNR" your image has than that of your pal? Or is it to better understand the perceptual gains of doing an optimal exposure? Or something entirely different?

-h

[hjulenissen]

Then you need to study this problem more.

I think that this line of communication is not really helping either of us.

-h

[Guillermo Luijk]

I said I don't know how many people with appreciate 10, 20, 30, ..., SNRs figures attached to a single image. IMHO, this is not a very conducive situation. We should study SNR more closely and figure out models that yield a single number that we can quote for an image with words to the effect of "this image of yours has this SNR", as opposed to "this image of yours has SNRs of 10db, 0db, 3.2 db, 4db, ....., and wait, finally 6db".

That is possible too: if we can calculate the SNR of each area of the image, it's trivial to calculate the mean, the median of even the histogram of the different SNR found in the image.

But IMO this summarized 'SNR single figure' would only be useful from an engineering point of view. In practice, every photographer processes his pictures in several ways that make that single SNR figure quite theoretical: denoising plain areas will improve perceived SNR, contrast (slopes >1) applied to some area will make SNR worst in that area, while de-contrast (slopes <1) curves will improve it. Darkened areas (no matter if they keep or even improve their SNR), will be less visible to the observer and hence less relevant to the final 'noisy' perception while SNR in the medium and highlights will contribute more to the final perception of noise, etc...

Regards

[joofa]

That is possible too: if we can calculate the SNR of each area of the image, it's trivial to calculate the mean, the median of even the histogram of the different SNR found in the image.

True, I agree with you. A weighted average would make sense. But there are fundamental problems regarding patch extraction. Segmentation of images to obtain patches is possible, but perhaps is not always a well defined process, and certain assumptions have to be imposed regarding when pixels are sufficiently different to stop classifying them to the same patch. In any case, if we go this patch-based-segmentation route, do you think it can answer other fundamental questions such as SNR variation in image resolution changes, etc., that you mentioned above?

Sincerely,

Joofa

[Guillermo Luijk]

Segmentation of images to obtain patches is possible, but perhaps is not always a well defined process, and certain assumptions have to be imposed regarding when pixels are sufficiently different to stop classifying them to the same patch. In any case, if we go this patch-based-segmentation route, do you think it can answer other fundamental questions such as SNR variation in image resolution changes, etc., that you mentioned above?

To apply the 'patch concept' in measuring SNR we actually don't need real patches, I'd apply a continuos solution instead:

• Calculate exposure on each pixel (from RAW data) with respect to sensor saturation
• Apply gaussian blur (small radius) so that each pixel becomes represented by the average exposure in its surrounding area
• Match those values obtained for each pixel against the SNR response, providing a continuous map of SNR of the RAW file

SNR correction because of resizing is also easy, just needs to use a corrected SNR response. For example to normalise SNR to be comparable to Canon 5D's 12Mpx, I applied this simple equation derived from the quadrature principle:

SNR_norm_dB = SNR_perpixel_dB + 20 * log10 [(Mpx / 12,7)^0.5]

being Mpx the pixelcount in Mpx of the camera used.

[joofa]

• Apply gaussian blur (small radius) so that each pixel becomes represented by the average exposure in its surrounding area

So you are applying a low pass filtering even before SNR calculation. To me that doesn't seem the right thing to do as it will disturb the image signal and noise relationship in the actual data.

• Match those values obtained for each pixel against the SNR response, providing a continuous map of SNR of the RAW file

Which SNR respone? I don't understand.

SNR correction because of resizing is also easy, just needs to use a corrected SNR response. For example to normalise SNR to be comparable to Canon 5D's 12Mpx, I applied this simple equation derived from the quadrature principle:

SNR_norm_dB = SNR_perpixel_dB + 20 * log10 [(Mpx / 12,7)^0.5]

being Mpx the pixelcount in Mpx of the camera used.

I don't think that is technically the correct formula to be used. Remember, I asked where is the "signal" in the normalization process. Can you please comment where or what is the signal component in this equation?

Sincerely,

Joofa

[PierreVandevenne]

Yes, you are correct, and that is a way. And, I agree with that. However, as I said I don't know how many people with appreciate 10, 20, 30, ..., SNRs figures attached to a single image.

As far as photography is concerned, it is probably of no practical interest. In scientific or medical imaging, it is very important. When you image multiple point sources, such as stars, they will all have different SNR. The SNR gives you a good idea of the precision of your measure (not necessarily of its accuracy). There's a treshold under which your data can't reasonably be significant. In other fields, such as medical imaging, you'll want to know how little you can spray the patient with potentially damaging x-ray photons and still get a significant image of the intended target.
But again, the core issue is simple. There is an optimal way to expose a sensor: maximize the signal without clipping it (and ideally keep it in the sensor's linear zone). The two problems ETTR addresses are 1) some cameras have sensors whose dynamic range is too small to cover a scene as we'd like to see it. 2) those cameras typically implement ETTL - they expose to the left of the ideal exposure to protect most users in most practical scenarios. ETTR simply slides the dynamic range window where it should be. With a camera with a wide enough dynamic range window, you don't care. Whether it increases the number of levels, whether we perceive those levels, whether Michael's explanation was 100% accurate, etc... is essentially irrelevant because there is only one exposure that is ideal for the sensor, the one that puts the maximum of photons in the best zone.

[hjulenissen]

• Apply gaussian blur (small radius) so that each pixel becomes represented by the average exposure in its surrounding area

Why is it that spatial averaging is needed again?

If you want to define a spatial neighborhood based on brightness similarity (and spatial proximity), you might want to do something like bilateral filtering.

-h  

[Guillermo Luijk]

So you are applying a low pass filtering even before SNR calculation. To me that doesn't seem the right thing to do as it will disturb the image signal and noise relationship in the actual data.

Which SNR respone? I don't understand.

The SNR is not obtained from the image, but from the SNR response (i.e. the SNR curves that provide the SNR in that sensor for each exposure value).

The only reason for the blurring is to obtain the average exposure in the pixel's surrounding area. A single pixel's SNR cannot be measured unless you have a before/after noise addition pair of images, and even in that case is irrelevant since the observer doesn't perceive individual pixel's SNR, but local SNR. The local averaging (blurring) provides this average perceived exposure that we just need to translate to the SNR curve of the sensor to obtain the average local SNR.

I don't think that is technically the correct formula to be used. Remember, I asked where is the "signal" in the normalization process. Can you please comment where or what is the signal component in this equation?

That formula is the extrapolation of the 4 pixels binning formula, very well explained in Emil's article.

Joining 4 pixels in one improves SNR by 2, the scaling factor:

(4/1)^0,5=2


The scaling factor from a X Mpx camera to a 12,7Mpx camera is:

(X/12,7)^0,5

The formula I showed is just this scaling in dB.

Perhaps the tutorial about measuring the SNR curves can help to understand the process.

[joofa]

The SNR is not obtained from the image, but from the SNR response (i.e. the SNR curves that provide the SNR in that sensor for each exposure value).

Were those curves derived from real images or flat patches?

That formula is the extrapolation of the 4 pixels binning formula, very well explained in Emil's article

Joining 4 pixels in one improves SNR by 2, the scaling factor:

(4/1)^0,5=2


The scaling factor from a X Mpx camera to a 12,7Mpx camera is:

(X/12,7)^0,5

The formula I showed is just this scaling in dB.

Technically, Emil is incorrect here: This formula will work (1) only for a flat patch, and not for a real image, and (2) for downsampling by binning, which is a poor method. How will the formula change if I use Lanczos that is what many (most?) people will prefer for downsampling as opposed to binning?

Sincerely,

Joofa

[Guillermo Luijk]

Were those curves derived from real images or flat patches?

From both, real images of flat patches as explained in the tutorial.

Technically, Emil is incorrect here: This formula will work (1) only for a flat patch, and not for a real image, and (2) for downsampling by binning, which is a poor method. How will the formula change if I use Lanczos that is what many (most?) people will prefer for downsampling as opposed to binning?

No idea how the formula will change for other type of rescaling, I just chose this criteria because the only intention was to compare sensors. With other type of rescaling the change in SNR will be different of course, but for typical interpolation methods I'm pretty sure this simple model works well.

Regards

[hjulenissen]

The only reason for the blurring is to obtain the average exposure in the pixel's surrounding area. A single pixel's SNR cannot be measured unless you have a before/after noise addition pair of images, and even in that case is irrelevant since the observer doesn't perceive individual pixel's SNR, but local SNR. The local averaging (blurring) provides this average perceived exposure that we just need to translate to the SNR curve of the sensor to obtain the average local SNR.

So there are two reasons:
1. To avoid noise affecting the (effectively) estimation of signal in isolation
2. To better emulate the human perception