I was taking line profiles of blue sky noise in two images--an 8080 image posted recently and 3020 image from my camera. If I ratioed the peak to peak variation --11levels for the 8080 and 5 for the 3020, I can say the 8080 is 2.2 times more noisy than my camera. If I ratio the standard deviations-- the rms noise- I get 1.6 or roughly the square root of 2.2. In electrical engineer parlance I believe the first ratio is the equivent of the variations in the voltage while the second is the variations in the power, and power noise is normally used as the standard. In visual terms I'd think it should be the other way around. We see differences in pixel intensity not pixel power intensity. By that standard the 8080 is 2.2 times more noisy than my camera. Anyone have a different opinion? jpc
I try to answer the original question. The gray levels on the photo (digital or film) are produced by photo energy (which is power x time). When you read out the voltage (or current) from the photo, the voltages represent the original photo or noise power. In my opinion, you don't need to do square root again.
You are interpreting your results incorrectly. The relationship between P-t-P and RMS noise is not one of voltage vs. power. The keys is that the "R" in RMS stands for root. Thus, both measurements are in the same units. P-t-P measures maximum excursions, including the rare "tails" of the random distribution describing the noise. RMS describes a type of average. The result is a number which, *if* squared, would give you a power equal to the average power. Using your voltage analogy, and with a sine wave that varies between -1V and 1V, the peak-to-peak measurement would be 2V and the RMS would be 0.707 or sqrt(2)/2 V. David
The 2.2 to 1 ratio of p-p to rms only applies to sinusoidal waveforms. Noise in a camera is not sinusoidal, so do not expect that ratio to work. The noise is likely to be a form of Poisson distribution.
Unlike some, I shall refrain from lecturing you on the math of rms vs pp. I think you know all that. I think the answer to your musing is that "it depends". Probably no-one has done any research on the *subjective* perception of image noise. That would require a large number of tests with many people observing many pictures and scoring them wrt noise. I suspect if such work was done the outcome would be a pink noise/white noise equivalent weighting that takes into account such matters as spatial bandwidth, the colouration of the noise, the amount of image detail etc. In other words, the rms vs pp vs whatever other measures question can't be resolved. It's an exact equivalent to the grain debate. There's nice grain and ugly grain. People used to argue interminably Tri-X grain versus HP3 grain or whatever, or the grain produced by one developer or the other, or at 20'C versus 25'C. As to the two cameras you compared. The ratio is 1.6:1 or 2.2:1. Either way I'd say "about twice as much". Anything else is measurebation.
I often wonder why digital still camera noise levels aren't quoted in decibels (db) as in other type of electronic signal devices? For example a good broadcast television camera's signal to noise ratio (the rms noise level relative to a signal voltage at peak white) would be (say) 50 db and would be better by a voltage factor of two than a camera with a signal to noise ratio of 44db (under the same conditions). Using this system we would have precise and easily measurable noise figures with which to objectively compare cameras. The electronics industries have used this method of noise measurement for the last 80 odd years in all kinds of signal transducers (TV cameras, microphones, record players, telephones etc.) and processing circuitry (amplifiers, transmitters, telephone lines etc.). There is no reason (that I can see) why this method shouldn't be used for digital still cameras. Why has this time proven method been ignored by the digital photography industry? Regards Rescho
[POSTED TO rec.photo.digital - REPLY ON USENET PLEASE] In <406e3512$0$12740$> on Sat, 3 Apr 2004 Perhaps because of the common use of noise reduction, which makes any such spec problematic.
Perhaps because of the common use of noise reduction, which makes any such If you are referring to "in camera" noise reduction then that would be included in the s/n figure for the camera and used when evaluating the camera. If you are referring to post camera noise reduction the I can't see that it's relevant. For example, with my previous camera (Minolta 7i) I had to use noise reduction to improve images taken at (say) 400 ISO, now with my Canon 300D, the noise levels are usually acceptable without post camera noise processing at the same ISO setting. This is because the Canon has a better s/n ratio in the first place. Regards Rescho.
[POSTED TO rec.photo.digital - REPLY ON USENET PLEASE] In <406e3f5a$0$439$> on Sat, 3 Apr 2004 14:35:52 It's not simple to either define or to measure what's going on.
It's not simple to either define or to measure what's going on. I don't agree. A low noise digital to analogue converter to change the camera's digital signal to a signal that would look like the analogue output of a computer video card channel, photograph a white card against a black background. Calibrate an oscilloscope (CRO) to the peak to peak voltage of the signal (white to black) then increase the CRO gain to measure the residual noise voltage at the black level. Quick calculation and you've got the s/n in decibels. That's the basic method that was used for measuring noise in broadcast TV cameras and the technology in still cameras is essentially the same except for the initial digital to analogue conversion part. There is always an existing analogue output signal in TV cameras. It wouldn't be the sort of thing you could or would do at home but manufacturers and professional camera evaluators could use something like this to give is more objective camera noise figures than we currently get. Manufacturers probably already do. We probably don't these figures because we don't ask for them. Regards Rescho
Primarily because the MTF (frequency response) isn't flat. A camera with the same sensor and electronics but more built-in sharpening would appear to have a poorer S/N than one with less sharpening. For a fair comparison, the MTFs would all need to be the same, and they are not. What I think you need to do is: to put optical sine waves into the camera, measure the narrow-band S/N at that spatial frequency, and integrate the results with some known eye-noise-sensitivity rating like the RIAA filters for audio noise measurements. This would make an averaged perceived S/N for the camera. To measure the whole system you need to include the display MTF and the eye's MTF as well, but not just for comparing cameras. Most other electronic devices aim for a flat frequency response.... Cheers, David
Why on earth would you want to do this in the analog domain? You've already got the numbers in the digital file, no need to measure again. Nobody's arguing that it would be hard to do. It would be easy to do. What they're saying is that the result would be grossly misleading, because in-camera noise reduction doesn't work equally well for all images. So, you'd get a measurement for the noise level on a white card that would be accurate if and only if all you ever did was take photographs of white cards. Andrew.
There is no reason not to do it in the digital domain - I used that process as an example to explain one method that could be used. Unless digital cameras use some sort of dynamic processing of images that doesn't show up in Exif data and that I'm not aware of I can't see why the camera noise levels should vary with image content. Could elaborate on that? I used the white card as an example and one that I know is used for broadcast television cameras - there may be a more appropriate reference image that could be used for still digital cameras. Whatever reference is decided upon then all camera s/n measurements would use that standard so the results would give a consistent and usable indication of all cameras s/n ratio. regards Rescho
The point could be made that it would be fair not to compensate for any in camera enhancement when measuring noise in a camera as it's the final camera output is what counts. If the final image has measured noise because of resolution enhancement and it's unacceptable then so be it - the resultant image would have to be post processed to remove that noise, a process that in an ideal world should not be necessary. As I visualise it, the problem of putting optical sine waves into the camera as in your suggestion above, would involve the added variable of lens resolution as that is the only way that it could be done (that I can see). i.e. a test chart with sine bars on it. I agree to be acceptable some sort of frequency response weighting would be necessary to standardise the procedure and make it match the visual perception of noise to the human eye. The problem then is of course - who follows standards in anything digital? :>) Regards Rescho
[] Yes, it would have be an optical input, and therefore involve the lens. But for fixed lens cameras (the bulk of the market) you would have to do something like that in any case. For a given light level, by varying the zoom you would end up with an integrated SNR versus focal length plot, providing a direct camera-to-camera comparison. Standards - an issue, yes, but the audio guys got to it together (for weighted SNR or sound level measurements....). Cheers, David
Indeed. That's the point. The new digital cameras use a bunch of prorietary image processing techniques to reduce noise. These can make a huge differene to image quality, as seen particularly with the Kodak DCS/14 when its firmware was upgraded. But as I said, these techniques are image dependent. It's no as easy as that. You have far too simple a model of what goes on in a digital camera. Andrew.
The only tricky parts are probably getting things down to the n'th decimal place, but the testing procedure is essentially how Rescho describes (there is really not much else that can be done). See: http://www.dolabs.com/Photography/DxO_Analyzer.html A long write-up for this product: http://luminous-landscape.com/reviews/dxo-explained.shtml
For a basic SNR test, you want to illuminate each pixel as equally as you can, so the field is flat. This factors things like lens MTF, and even critical focus, from the equation. More imporant are geometric distortion and vignetting -- sources of non-flatness in the illumination.
When I reread what I said in the part about electrical engireer parlance I doubt it. What I intended to say--and this may not be totally correct--is that EE terms like db are define by rms/standard deviation numbers. I don't think you can talk about a peak to peak db without making several assumptions about wave forms and other thing that don't normally apply to digital imagery. And if I'm wrong, correct me. Just include a website or two that explains the details. Not that I doubt you but this is rec.photo.digital You maybe right but I have another idea to toss in the pot. When someone pull out a magifing glass to compare two prints or blows up a couple digititized images to 800 percent on a monitor, he is looking at peak to peak noise. Then he is making a worse case evaluation of the differences. But when he pulls back and studies the photos as an average person would he is looking at rms noise. Under these conditions the S/N variations aren't so extreme and the photos don't look that different. When the differences are around a factor of 2, maybe it is measurebation. But if you like to push the camera to the limits while taking low light non flash images, you need all the S/N you can dig out of the camera. For those who think it is possible to use as single number to define a camera's S/N performance don't forget that CCD's are background noise limited. And as the technology improves and the noise floor dropps they will become even more background noise limited. jpc
