Originally posted by jeffkrol Shot noise, I believe, would rarely ever overwhelm read noise.
[...]
The major difference is that read noise sits like a blanket or a cloud in the image; obscuring the hills and valleys, making them them all flat through a lack of contrast.
[...]
Read noises generated by cameras after the photonic capture also have macro and micro patterns in them; one-dimenional components, etc.
Jeff, thanks for the quote.
First, wording ...
I meant the term "read-out noise" exactly in the way as your citation above did. However, one should note that it is more about dark currents and artifacts than noise, actually. I called it read-out noise though because this is a more commonly used term.
Second, shot vs. read noise ...
You didn't see the images before the 16x super imposition ... Totally useless. Actually, shot noise can be computed:
1 W (watt) = 93 lm (lumen = 1 cd·sr = 1 lx·m^2) (using the luminous efficiency of sunlight of 14% only);
1 green photon = 3.6E-19 J (joule).
18% graycard at 1000 lx (lux = 1 lm/m^2) = 9 EV (light value)
[source: f8 1/30 ISO400
KODAK: Estimating Luminance and Illuminance: Pub A-105KIC]
Luminous flux onto sensor = 18% * Luminous flux onto greycard / (2 * pi * f-stop^2)
[source: own]
Example for EV 2 at 1s exposure at ISO 1600 (=> f/8):
- 18% graycard at 7.8 lx
- sensor at 0.0035 lm or 0.038 mW (milli watt)
- 1.0E14 photons hitting sensor within 1s.
- Imperfect bayer filter (15%), quantum efficiency (25%), 14.6 Mio. Pixels =>
- 260000 photons detected by a sensor pixel
- Shot noise: 500 (Poisson law) (affecting 9.0th bit) for mid gray (rgb 117)
Now for RGB value 10 (1/224 of 18% gray only; sRGB gamma correction!!):
- 1160 photons detected by a sensor pixel
- Shot noise: 34 (Poisson law) (affecting 5.0th bit) for dark gray (rgb 10)
And if you push 14 Bit RAW darks (RGB value 1, 1/30000 of 18% gray only):
- 9 photons detected by a sensor pixel
- Shot noise: 3 (Poisson law) (affecting 1.5th bit) for black (rgb 1)
And you loose 1/2 additional bit per one stop decrease of EV in scene or increase in ISO. So, at ISO25600, you have lost 2 more bits, leaving 3 bits for dark gray.
Just because there are a finite number of photons around only ...
DISCLAIMER:
I do not claim that my above calculation is correct. It is of the back-of-an-envelope kind
P.S.
You can invert all that fuzz to compute quantum efficiencies for the 3 colors from noise measurements. That would be useful numbers to classify sensors, as opposed to the DxO wizardry. I think, Gordon B Good tries this with some success and computes the full well capacities (a closely related measure) from the noise. Also, while noise measurements at high rgb values give you the quantum efficiency or full well capacity, any excess noise seen in low rgb values can be used to compute the read-out noise from.
Last edited by falconeye; 08-17-2009 at 04:17 AM.