« on: February 24, 2014, 04:52:55 AM »
I'm really not convinced that pixel size has as much to do with image noise as some say it does. Over the years as sensor resolution has continued to rise, high ISO noise has also gone down.
I actually suspect that the high ISO capabilities of current full frame cameras have as much to do with the processor as any aspect of the sensor. There are also many things that can be improved irrespective of pixel size, maybe Canon saves their best designs, materials and components for high end models. They cripple other things commonly enough, why not specifically produce worse low light on low end models?
Lower resolution means you can do more processing per image. If it is mostly processing power making the difference then a 40MP camera with the same processor as the 1Dx should have the same ISO quality at half the burst speed. I'm looking forward to seeing the results from Canon's first big MP full frame.
The readout rate is only going to affect read noise levels. Slower frequency components will introduce less noise of their own into the signal. If you look into astrophotography CCD cameras, they have very low readout rates, often reading fewer megapixels per second than there are in-total in the sensor. This, along with cooling by TEC, aims to reduce read noise (both dark current by reducing temperature, and high frequency noise by reducing readout rate (i.e. using slower frequency ADC units and the like.))
Read noise only affects the shadows, however. Overall noise is the result of the randomness with which photons strike photodiodes and free an electron. Photon shot noise, the primary source of high ISO noise, follows a Poisson distribution. The only link this has with the electronics of the sensor is via quantum efficiency. Increase Q.E., and you increase the ratio of incident photons to an increase in charge by one electron. Beyond that, photon shot noise is not an electrical phenomena, it is a physical and natural phenomena. Q.E. is already relatively high in sensors...most have around 50% give or take 5% or so. To reduce high ISO noise (at any ISO) by a factor of two, one must double Q.E. That means we can only half high ISO noise once (at best), but that would require 100% Q.E. (and not even astro cams with low frequency readout (i.e. 0.3 frames per second) and three-stage water-cooled TEC with a 70°C delta-T can achieve that!)
That leaves only one other option for reducing noise at high ISO: pixel size. Photon shot noise is the result of the randomness with which photons strike pixels. An increase in pixel size is effectively the same thing as an analog averaging algorithm...larger area, more incident photons per pixel, less total variation across pixels. There is absolutely no question that larger pixels result in less noise (especially at high ISO, where read noise is already a minimal contribution of total noise, far, far less than it is at lower ISO...for example, the 1D X has 2.2e- read noise at ISO 3200, but 38.2e- at ISO 100! The 5D III has 3.1e- at ISO 3200, and 33.1e- at ISO 100. Read noise is a negligible amount at high ISO.
However, contrast the charge saturation point of larger pixels of FF with the smaller pixels of APS-C, and there is a very clear benefit to FF sensors. The 1D X has 3069e- and 10.5 stops DR at ISO 3200. The 5D III has 2179e- and 9.5 stops DR at ISO 3200, where as the 7D has 1067e- and 8.5 stops DR at ISO 3200. The 7D has 2.9e- read noise at ISO 3200, approximately the same negligible amount as the 1D X and 5D III. The 1D X has a three-fold noise advantage AND a two-stops dynamic range advantage at the same ISO, and that is all thanks to the larger physical photodiode area. The 5D III has a clear two-fold noise advantage over the 7D.
The logical next step is to say the 70D has better pixels. The 70D has better pixels at LOW ISO (a good thing, certainly for Canon) thanks to a higher FWC (26726e- vs. 20187e-), however it's smaller pixel area than the 7D still results in a lower saturation point at higher ISO. The 70D gets a mere 999e- charge saturation at ISO 3200. It has a slight 0.2 stop DR advantage thanks to ever so slightly less read noise, but that difference is within the margin of error of Canon's metering sensor.
When it comes to high ISO image quality, there is absolutely no substitute for total photodiode area. One either increases pixel size, or one could, as I mentioned before, possibly layer photodiodes in each pixel. Three times the photodiodes in depth, and one could potentially double the sensitivity of APS-C pixels. Combined with BSI, and one could make APS-C pixels pretty small and still have the IQ be acceptable, even quite good. But, one could also always employ the same technology in FF sensor pixels...so APS-C will never have the IQ advantage. Except in the case of very reach-limited scenarios...no amount of pixel processing can really overcome the benefit of 4x more pixels.