f-number equivalent between m43 and FF

Started Mar 25, 2014 | Discussions thread
bobn2
bobn2 Forum Pro • Posts: 52,276
Re: f-number equivalent between m43 and FF

dwalby wrote:

mosswings wrote:

dwalby wrote:

bobn2 wrote:

dwalby wrote:

bobn2 wrote:

dwalby wrote:

bobn2 wrote:

dwalby wrote:

Allan Olesen wrote:

D Cox wrote:

The total area of the sensor is irrelevant. If you put the same lens on various sizes of sensor, the noise level in the part of the image that they all record will be identical. (Assuming they are the same generation of technology.)

Let us compare the Nikon D800 (FF, 36MP) and the Nikon D7000 (APS-C, 16MP).

Same pixel size. Same Sony sensor technology. Same noise per pixel if you expose equally (same f-stop, shutter speed and scene light).

But not same magnification of the final output.

If you take the same photo with these two cameras and enlarge to the same output size, each pixel from the D7000 will be magnified 1.5x more in each direction. Consequently, the final output will appear more noisy.

And you can't really say I am wrong about that because I am actually just repeating your own claims from another post in this thread: The same pixel will look more noisy the more you magnify it.

So yes, it is all about sensor area.

I agree its all about sensor area, but I'm not sure I agree with your example.

In the case of two sensors of different size, but with the same size pixel, you have the same SNR for each pixel, we all agree on that.

Where I don't necessarily agree with you is that the simple act of printing the pixel larger or smaller will have any effect on its SNR.

A pixel has no SNR (in the context of a single photo)

a pixel has a signal level, as measured by the A/D converter, and an uncertainty due to noise.

Wrong. A pixel has a signal level as measured by the A/D converter. That is all, it gives a single value.

-- hide signature --

Bob

OK, let me reword my statement a bit. A pixel contains an analog signal value prior to A/D conversion. Every analog signal has some sort of noise associated with it, and the A/D conversion itself introduces read noise. So each pixel, when sampled, has a signal level, and some uncertainty due to the noise.

Even when your statement is reworded it remains wrong. There is no noise in a single observation. You just read the pixel and get a value. There only becomes noise over a number of observations, and in the context of a single photo, each pixel is read only once. So, the noise becomes apparent when you view a group of pixels, which you believe should have the same value, and they don't. There is no 'uncertainty' of separable noise value in a single pixel, no noise component in a single reading of a pixel. The point about this, in the end, is talking about 'per pixel' noise is a nonsense - what matters is the variation over an area, and in the context of photography, the variation of the same area of the same size output image. More particularly, noise is bandwidth dependent, and if you're comparing noise you need to normalise bandwidth (as is typically done in electronics where noise is given 'per Hz').

-- hide signature --

Bob

OK, now I see your reasoning and I guess its a matter of how you want to define noise. I was defining it in terms of what is the probability that the sampled pixel accurately represents the expected signal value. As you know, in a Gaussian distribution there is an expected probability that can be computed for each value along the distribution. If we know the illuminance, sensor efficiency and pixel area we can compute that distribution and have a good idea what our chances are of sampling a pixel with a given level of confidence that it measured the expected signal level accurately.

So wouldn't you agree that the effect of the noise is still there, even though it may or may not be obvious to the observer of a single pixel. If you crank the ISO up to the highest setting possible, take a picture with some deep shadows, and look at one pixel in the shadow area, the effect of the noise is there in the pixel value itself. If you knew the pixel should be black, and it recorded as blue, you know the noise corrupted the value even without any adjacent pixels to prove that was the case. If you didn't know what color it should be, then the noise might not be obvious to you on observing a single pixel, but that doesn't mean there was no noise contribution to the process.

When you look at the surrounding pixels and see they're all widely varying in values that should be relatively close, then you are more aware of the degree of the noise, but I'd still claim that the noise existed even in the single pixel case. I've also seen numerous articles on sensor performance that all use the term 'per-pixel noise', so I don't think the concept is foreign and I'm just making it up.

A curious exchange. Indeed bob is right in that "noise" is an emergent property of a signal stream, not a single measurement, so a time series for a single sensor or a spatial matrix for pixel array is required to observe it. However dwalby is correct that per-pixel noise is a useful concept, as long as one recognizes how it is being used. In the case of per-pixel noise, this can be considered a useful normalization basis for a time or space series of measurements, just as we normalize the frequency behavior of noise to per-Hz or per root-Hz.

The per-pixel noise idea gets a bit of traction when we consider the internal structure of the pixel - a bucket that converts photons into electrons, some of which are related to fundamental optical processes - photon noise - some of which are generated internally, and some of which are generated in the amplification and conversion processes. This can help the sensor designer achieve design goals for the product, as a sensor is made of these fundamental elements. Again though, those "noise levels" are emergent properties, and you can't tell from a single measurement whether the recorded signal is noisy or not. You have to perform the measurement many times and analyze it.

Another way of saying this is that noise can only be discerned in relation to something else - the next sample in a time series, or the adjacent pixel in an array (actually, both).

Bob is approaching the discussion in terms of what are the limitations of an undersampled system. In DSP terms, if you take one sample of any signal then by observation you could declare it to be sampling a waveform that is DC with no noise. Without a properly sampled system you can't observe the true characteristics of the sampled signal, but at the same time that doesn't mean you can claim those characteristics do not exist. Bob is not claiming noise doesn't exist, just that its not visually detectable in an undersampled system, so why concern yourself with it.

BPSK modulated systems typically make their symbol decisions with a single sample, after being bandlimited to reduce the noise power to the absolute minimum. Just because its a single sample decision doesn't mean there's no noise interacting with that sampling process, and based on Eb/no calculations the probability of an improperly detected bit can be accurately modeled. But what Bob is claiming, and I agree with him in that claim, is that looking at any bit decision value alone doesn't tell you anything about whether it was a correct or incorrect detection of the bit. Where I was coming from is similar to the Eb/no computation that says when I look at any given bit decision I know a priori the probability of that bit being correct, and SNR plays an important part in that calculation, thus my claim that a single pixel sample does have an inherent SNR.

In the case of the sensor its more like 256-QAM than BPSK, but the concepts are the same.

What I'm saying also is that because the signal that the pixel is measuring is itself random, there is no clear 'expectation' of what it 'should' be. The best the sensor can do is accurately record the pattern of photons presented to it.

-- hide signature --

Bob

Post (hide subjects) Posted by
(unknown member)
(unknown member)
Keyboard shortcuts:
FForum PPrevious NNext WNext unread UUpvote SSubscribe RReply QQuote BBookmark MMy threads
Color scheme? Blue / Yellow