Joofa wrote:

Great Bustard wrote:

Joofa wrote:

Under ordinary setups we don't see photons.

This is a complete non-sequitur.

Yeah right.

We only have access to some measure of photoelectrons at best.

Photons fall on the sensor, they release electrons, and the resulting signal is measured.

So what does that say about the measure of photons? How do you get that? I thought your so called "density of light" was a quantitative measure? Perhaps not and it is just a rhetorical device.

With the usual digital cameras around there is no reliable way of measuring the variation in photon flux, and hence, effectively the so-called "density of light", unless you want to take an image of a uniform brick wall.

Again, a complete non-sequitur. Who's talking about "variations in photon flux"?

I thought for a usual picture the photon flux varies spatially. No? May be you only take pictures of uniform brick walls as I suspected

You're trying to divert to your old tired argument that the photon noise is not accurately described by the square root of the signal.

I think you are out of your mind. I have never said that.

You've been asked over and over to support that claim, and you never deliver.

Pass along what is that you are smoking. I have never disagreed with what you are pretending that I disagreed with.

Now that I've swallowed your bait, please go start a new thread where you demonstrate that the photon noise is not accurately described by the square root of the signal, where the signal is measured in electrons. I'm sure Dr. Martinec would love to read all about how you have proved him wrong:

http://theory.uchicago.edu/~ejm/pix/20d/tests/noise/#shotnoise

Don't try to play to authority. When have I said that I'm disagreeing with what Prof. Martinec is saying in that article. Can you please point me to a link where I have said that I disagree with that article of his?

Being nearly Groundhog Day, I was just reminiscing about the famous "woolly concept" disclaimer:

Re: Measured Signal vs True Signal in a Poisson Distribution

ejmartin wrote:

Joofa wrote:

If you take that image on which Emil has done his noise-frequency analysis you should be able to find many sort of uniform patches where like him you can do a noise-spectrum analysis locally. However, the problem is that as you just said above the measured noise is described by sqrt being a Poissonion process, then a darker patch will you give one number and a lighter patch will give you a different number. So far so good. Now if I ask you that what is the "image-level-noise" (whatever that means) for this image, then which number are you going to quote to me?

I won't give you a number but rather a function. The point is that in uniform patches the noise is spatially white to a good approximation, so the 2d frequency spectrum is constant over a wide range of frequencies from Nyquist to the inverse patch size. There is an overall normalization of the spectrum which depends on the average tonality of the patch, which is where the Poisson property comes in. So one has a function of spatial frequency which to a good approximation is constant wrt frequency and depending on the local average tonality as the sqrt of the average intensity.

I never use the phrase "image noise" because it is indeed such a woolly concept.

http://forums.dpreview.com/forums/post/36073105

And the ensuing "image-noise" debate:

Re: Defining 'image level noise'

Joofa wrote:

Great Bustard wrote:

Great Bustard wrote:

I think that we need to define "image level noise" before trying to quantify and compute it.

Joofa wrote:

They way you have defined it you have reduced the whole image to a single random variable. That is fine. There are other ways also.

Well, the way I see it, noise has an amplitude at a particular spatial frequency. My treatment has been to compute "total image noise" at a given frequency.

I think there is no concept of spatial frequency here as what you have now is a big fat single random variable - each image will give you just a single number.

But what's interesting about my definition and computation is that if we repeated the calculations for a different number of pixels, we'd get the same answer.

Won't you get a different answer. I thought it is affected by the number of pixels as sqrt (n). Let me recheck.

But you took the stationary case (like Emil). This is not the real image case.

In other words, "total image noise" is scale invariant for the way I defined it.

See above comment.

I think we agreed above that if the true signal were unknown, then using the measured signal in place of the mean signal would not affect this by any significant amount.

Only in relatively high illuminant areas. For areas that are in low light this relation will break down as sqrt becomes increasingly an appreciable portion of the signal content.

Let's stop at this point. I don't understand what you mean -- it's your turn to work an example for me.

Take your number 10,000 (a case of high illumination), which sqrt is 100. I.e., 100/10000 = 1/100 = 1% variation. Now take 25 (a case of low light), which sqrt is 5, so 5/25 = 1/5 = 20% variation. See the sqrt is becoming an increasingly appreciable amount of the signal so it is becoming increasingly difficult for us to take the sqrt of the measured signal as equal to its true unknown value.

http://forums.dpreview.com/forums/post/36080460

.

DM...