f-number equivalent between m43 and FF

Started Mar 25, 2014 | Discussions thread
Jack Hogan Veteran Member • Posts: 6,507
Re: SNAP! Where it is?

crames wrote:

Jack Hogan wrote:

crames wrote:

dwalby wrote: When you bin, the processing advantage only applies if the noise samples are uncorrelated, which they are at native resolution. If you upsample the smaller image to the larger image pixel count, your noise samples after upsampling are no longer uncorrelated. Therefore if you later bin both images, one will enjoy the sqrt(N) SNR improvement, and the other will not, because the larger one has uncorrelated noise samples and the (upsampled) smaller one does not.

Another way to look at it is in the frequency domain .

I wonder if that's not a misleading graph in this context, as I think it applies just to read noise: it shows the amplitude of the signal decreasing with increasing frequency, but the noise being constant with it. I understand this discussion to be mainly about rms shot noise, which I think would also decrease with a decreasing signal.

Poisson or shot noise is white noise, that is, more or less flat spectrum. A little beside the point because the first graph on the page was to show how why SNR might change when downsampling.


I should have just posted one of the graphs, this one, which can illustrate the upsampling case:

Upsampling (or low pass filter)

In the ideal case, when upsampling, you add pixels without adding any new frequency content. Say you have an image with a certain number of pixels with an MTF as shown in blue to the left of the dashed-green line (the Nyquist frequency). Upsampling would result in the Nyquiest frequency moving up to the green line labelled n1. Upsampling by itself would add no new content, neither signal nor noise, in the added frequency range between the 2 green lines. That's why the blue line is shown flat zero in that range. After upsampling, the image is bigger but still has only its original frequency content. In signal processing terms, the image has been zero-padded in the frequency space, which corresponds to ideal sinc interpolation in the spatial domain.

This illustrates how upsampling does not increase noise, (except if done badly), and certainly no new photon noise since this is all happening well-past the point where we are dealing with photons.

I would tend to agree.  The question is how do we work the fact that one image is pixelated and the other one is not in our equivalence discussion?  Perhaps we cannot.  Perhaps our IQ parameters allow us to only compare 'equivalent' images captured at the same resolution or at some downsampled common denominator - but not after upsampling.

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