What causes differences in "noise grain" among cameras?

I often see statements about one or another DSLR's exhibiting good (or bad) quality of the "noise grain" in their images, or that model A has better "noise grain" than model B. Imatest can spit out a measure of noise as a function of spatial frequency to try to quantify this.

My question is, why should there be such a difference, unless it is introduced by differences in the raw conversion being used? If we are talking about noise in midtones and highlights, that is dominated by photon shot noise which is purely white as a function of spatial frequency; it cannot differ from one camera to another as long as the raw data are raw and losslessly encoded. There can be a difference in shadows, since read noise can vary from camera to camera and contain more or less banding etc; though if we are talking about the grain aspect of noise rather than banding, that too is spatially white and my question still stands. So why is the "noise grain" a property of the camera and not the raw converter?

Of course there might be differences when viewing the full frame image due to differences in megapixel count, but I presume people are making these judgments by viewing the image at 100%, where the playing field should be level.

Any ideas?

why should there be such a difference, unless it is introduced by differences in the raw conversion being used? If we are talking about noise in midtones and highlights, that is dominated by photon shot noise which is purely white as a function of spatial frequency; it cannot differ from one camera to another........

Click to expand...

Hi Emil,

I assume that you're asking why one camera model exhibits more noise than another?

Well, one very good reason is the number of pixels photo-sites packed into the sensor. The greater the number of pixels, the smaller they have to be to make them fit. The smaller they are, the fewer photons they can collect. However, the random noise in the circuitry remains fairly constant, so the signal to noise ration is worse.

Am I being too simplistic here?

Regards,

Stuart

StuartRae said:

Hi Emil,

I assume that you're asking why one camera model exhibits more noise than another?

Well, one very good reason is the number of pixels photo-sites packed into the sensor. The greater the number of pixels, the smaller they have to be to make them fit. The smaller they are, the fewer photons they can collect. However, the random noise in the circuitry remains fairly constant, so the signal to noise ration is worse.

Am I being too simplistic here?

Regards,

Stuart

Click to expand...

No, I'm not referring to the total amount of noise, but rather its spatial distribution. For instance, this recent statement by Rob Galbraith:

"...examples of how the D3 stacks up against Canon's EOS-1D Mark III at higher ISOs, and those examples would reveal that, while there are differences in the appearance of image graininess - Canon's grain pattern is tighter - there's no doubt that the D3 produces a less noisy..."

I see statements of this sort in many places. I just can't figure out what they're talking about unless they mean that DPP's grain pattern is tighter than Capture NX's.

And yes, the D3 is less noisy, by about a 50% per pixel in terms of number of photons captured, or 25% in S/N ratio, in the raw files.

No, I'm not referring to the total amount of noise, but rather its spatial distribution.

Click to expand...

I see. Sorry for the misunderstanding. I shall have to engage the spare brain sell and have a think........

Regards,

Stuart

Emil Martinec said:

My question is, why should there be such a difference, unless it is introduced by differences in the raw conversion being used?

Click to expand...

Yes, Raw conversion is a major contributor to (the visibility of) lower spatial frequency noise, mainly because of the different sampling densities of Red and Blue versus Green. Whatever noise there is in the R/B filtered sensels, after quantization it will be interpolated more to fill-in-the-blancs (resulting in lower spatial frequencies for their noise) than the Green filtered sensels. Green is different in the sense that it is mostly interpolated in horizontal/vertical directions, and much less in diagonal directions (exception being the Fuji 45 degree 'rotated' sensel layout). So different emphasis on detail extraction versus noise per channel by Raw converters will amplify that (by now) spectrally non-uniform colored noise.

This is also amplifying the different noise levels from the R, G, or B filtered sensels, even before quantization and demosaicing. When you derive the 'analog gain' levels for the different channels, you will likely find that they are different (to compensate for different transparency per filter color). Different gains indicate that some colors will 'score' fewer photons for a given spectrally uniform level of exposure, so gain will compensate (and amplify channel noise). The noise differences can be small for large samples, but as always small samples of only a few pixels fluctuate more about the mean level, which is what is perceived as 'grain'.

Of course there might be differences when viewing the full frame image due to differences in megapixel count, but I presume people are making these judgments by viewing the image at 100%, where the playing field should be level.

Click to expand...

I try to not presume anything objective based on the majority of posts on various forums, unless I have a good understanding of the skill level of the person making the claims/observations. The S/N ratio is pretty poor on most forums when it comes to scientific analysis. There is also a good level of interpretation involved in human vision, we are not very good in objective quantification, and only a bit better at narrow field of vision comparisons.

And then there is noise reduction ...

Bart