Andrew Roos
Post count: 964

@Henk Coetzee 249134 wrote:

This is a debate I get involved in from time to time. My personal take is that at the scale of a specific size print, more resolution does not equal more noise and that sensor size (not pixel size) determines noise at this scale. Here is an interesting comparison of the Sony NEX-5 and NEX-7 which seems to confirm this.

This is correct, but only in the region where photon shot noise is the dominant noise source. Photon shot noise is the result of light travelling in discrete bundles of energy, called photons. You can receive 10 photons or 11 photons but not 10.5 photons. So if each pixel-sized area of a uniformly illuminated subject is emitting on average 10.5 photons in your exposure time, then some photosites on the sensor will receive 10 photons and others will receive 11 (and also, with lower probability, 9 or 12, etc). This variation in the number of photons received by different photosites all imaging the same uniformly illuminated subject is photon shot noise. Photon shot noise is a result of the number of photons actually received from the subject with a given illumination and exposure time, so it does not depend on the size of the individual photosites.

As photosites get smaller, eventually read noise dominates. Read noise is the additional noise, per photosite, added as part of the process of resetting, exposing and then reading out and digitising the amount of light received by a photosite. When read noise dominates, then image scale noise does degrade as you make photosites smaller (because read noise occurs per photosite, so the more photosites you have, the greater the noise).

Read noise is fairly amenable to technological advances (as opposed to photon shot noise, which is determined by physics) so as newer generations of sensor are engineered, the line where read noise becomes dominant moves, allowing sensors with smaller photosites to be competitive noise-wise (on image scale of course) with sensors with larger photosites. At the current state of technology, I agree that Sony’s new 24 MP APS-C sensor (used in the NEX-7) with 4 um pixel pitch seems to be photon shot noise limited, not read noise limited.

Although read noise is currently amenable to technological advance, ultimately it will run up against the thermal noise barrier – the amount of noise that is present in all electrical components simply due to their temperature. When this happens we will probably end up with a photosite size that you cannot make smaller without increasing image scale noise, unless you are prepare to cool the sensor in order to reduce thermal noise (this is almost universal in astronomy imagers).

(At some point the optical diffraction limit will also make it pointless to make photosites any smaller, as no lens with a reasonable f-number will be able to resolve images to the photosite size. For an f/2 lens the diffraction limited pixel size is around 1 um).

Just to confuse things even more, the currently highest ranked camera on DxOMark – a medium format back scores lower than the APS-C cameras on low-light ISO, beating them only really on colour depth.

MF backs are designed for studio work, with optimum lighting. Resolution and colour depth are their raison d’etre. It’s a bit like comparing the quarter-,ile sprint times of a Rolls with that of a Subaru Imprezza. (Actually, in real life, the Rolls will win. The scooby will be stopped by the barriers put up by the traffic police while the VIPs in their Rolls waft past at a stately 40 km/h).

This really makes me wonder what makes good “image quality”, assuming that DxOMark’s scores are actually representative of what makes a visually appealing print.

I find the DXOMark measurements (the actual graphs) interesting, but the individual scores less so and the total score is imo pretty pointless. I think their suggestion that the overall score somehow correlates to a maximum print size is pretty ludicrous becaue max print size is going to depend on subject, subject and camera movement, lighting and artistic intent and the one final score simply does not convey enough information to be even slightly meaningful across all possible scenarios.